Multi-agent collaborative advertising optimization method and system

Abstract

This application proposes a multi-agent collaborative advertising optimization method and system, belonging to the field of advertising technology. The method involves an operations manager agent parsing advertising goals and constraints, generating structured advertising task sheets, and writing them into a shared data space. Standardized task cards are then distributed to each agent. Creative creation, audience targeting and strategy, and advertising agent collaborate to complete creative generation, strategy combination, cross-platform campaign creation, multi-dimensional A / B testing, and automatic parameter tuning. All data is synchronized to the shared data space. After multi-level effect attribution by the monitoring and evaluation layer, each agent updates its knowledge base and iterates the baseline campaign combination, forming a closed-loop self-optimization system. This significantly improves campaign efficiency and effect controllability, enabling continuous self-iterative optimization of the campaign execution strategy.

Description

Technical Field

[0001] This application relates to the field of advertising delivery technology, and in particular to a multi-agent collaborative advertising delivery optimization method and system. Background Technology

[0002] In today's advertising industry, with the rise of multiple advertising platforms, cross-platform advertising has become a challenge that advertisers must face. Advertisers hope to run ads on multiple platforms to maximize their audience reach and advertising effectiveness. However, due to differences between platforms and resource limitations, advertisers face problems such as uneven resource allocation and unstable campaign results when running ads across platforms.

[0003] The existing advertising delivery system has three major problems: role fragmentation, localized automation, and lack of cross-platform information sharing. Specifically, these problems are: (1) The cost of manual collaboration between creative, delivery, and operation teams is high, and only partial automation such as bidding / creativity can be achieved, which cannot cover the entire link of "target audience – content – ​​delivery – attribution". (2) The data standards of each platform are different, and the performance indicators are difficult to guide the front-end decision-making. Moreover, it is difficult for humans to efficiently complete multi-dimensional A / B testing, which limits the overall optimization capability. (3) Most of the existing related technologies are single agent / single module optimizations, or only focus on single links such as creative generation and bidding adjustment, and lack a full-link multi-agent collaborative architecture and cross-role information sharing and responsibility attribution mechanism. Summary of the Invention

[0004] To address the aforementioned issues, the main objective of this application is to propose a multi-agent collaborative advertising delivery optimization method and system. This system aims to organically combine a multi-agent collaborative operation mechanism, a fully automated delivery loop, multi-dimensional A / B testing, and a strategy self-iterative optimization system. While achieving intelligent execution and continuous optimization of the entire cross-platform advertising delivery process, it also solves problems such as high reliance on human experience, low cross-platform collaboration efficiency, and weak controllability of delivery results in traditional advertising delivery technologies. This system combines delivery effectiveness, execution efficiency, cost controllability, and compliance.

[0005] To achieve the above objectives, this application proposes a multi-agent collaborative advertising delivery optimization method, the method comprising:

[0006] The operations manager intelligent agent receives the goals and constraints of advertising placement, parses the goals and constraints into machine-readable indicators and constraints, generates a structured placement task sheet, and writes it into the shared data space. Based on the delivery task order, the operations manager intelligent agent determines the benchmark delivery combination and breaks down the delivery task into standardized task cards corresponding to the creative production intelligent agent, the audience targeting and strategy intelligent agent, and the advertising delivery intelligent agent, and issues them out. The standardized task card includes at least the input data source, output data field, result acceptance quantitative conditions, and task execution deadline. The creative production agent generates and filters candidate creative packages based on the corresponding standardized task cards, audience profiles, advertising platform material specifications and compliance constraints, and writes the unique identifier and metadata corresponding to the candidate creative packages into the shared data space. The target audience and strategy agent generates candidate group targeting strategies and candidate delivery execution strategies based on the corresponding standardized task cards. It then combines and trims the candidate group targeting strategies and candidate delivery execution strategies with the candidate creative package to generate a candidate delivery combination list. The candidate delivery combination list is output to the advertising agent and written into the shared data space. Each candidate delivery combination in the candidate delivery combination list is a one-to-one correspondence of candidate creative package, candidate group targeting strategy, and candidate delivery execution strategy. The advertising agent creates a cross-platform advertising plan based on the advertising task order and the candidate advertising combination list, and writes the mapping relationship between the unique identifier of each advertising platform and the unique identifier of the candidate advertising combination into the shared data space; The advertising agent pulls and processes real-time advertising data from various advertising platforms based on the advertising task order. Using the benchmark advertising combination as a reference, it conducts multi-dimensional A / B testing on each candidate advertising combination in the candidate advertising combination list, automatically adjusts the bid and budget within preset constraints, and writes the parameter adjustment records into the shared data space. The monitoring and evaluation layer collects relevant data from various advertising platforms based on the delivery task order, and performs multi-level effect attribution analysis on the relevant data according to the creative dimension, audience targeting strategy dimension, delivery execution strategy dimension and advertising platform dimension, generates attribution results and writes them into the shared data space; Each agent updates its knowledge base based on the attribution results, and updates the candidate campaign combination with the best performance in this campaign cycle as the benchmark campaign combination. Then, the operations manager agent returns to the campaign task order to determine the benchmark campaign combination, break down the campaign task into standardized task cards corresponding to each agent, and distributes the standardized task cards to the corresponding creative production agent, audience targeting and strategy agent, and ad placement agent.

[0007] To achieve the above objectives, this application proposes a multi-agent collaborative advertising delivery optimization system for executing the method described in any embodiment of this application. The system includes a business and configuration layer, a multi-agent collaboration layer, a data and feature layer, a monitoring and evaluation layer, and an external advertising platform layer. The business and configuration layer is used for business personnel to configure deployment targets and constraints, store risk control and compliance rules, and provide input and constraints for the multi-agent collaboration layer. The multi-agent collaboration layer includes an operations manager agent, a creative production agent, a target audience and strategy agent, and an advertising agent. Each agent collaborates to complete task decomposition, creative generation, strategy design, and campaign execution, and achieves information exchange through a shared data space. The data and feature layer includes a user profile and feature library, a historical deployment log library, an attribution result library, and a knowledge base, which provide decision-making basis for the multi-agent collaboration layer and store full-process data. The monitoring and evaluation layer is used to perform data retrieval, preprocessing, A / B testing, multi-level attribution analysis, and optimization suggestion generation to feed back data to the multi-agent collaboration layer. The external advertising platform layer includes multiple cross-channel advertising platforms, which connect with the advertising agent via API to receive placement configurations and transmit real-time placement data.

[0008] In the technical solution provided in this application embodiment, the operation manager intelligent agent parses the manually input advertising placement goals and constraints into machine-readable standardized indicators and rules, generates a structured placement task sheet, and completes the determination of the benchmark placement combination and the fine decomposition of placement tasks based on this. It issues exclusive standardized task cards with clear input and output, acceptance standards and timeliness requirements to the creative production intelligent agent, the target audience and strategy intelligent agent, and the advertising placement intelligent agent, respectively. This can solve the core pain points of traditional advertising placement, such as unclear responsibilities in manual task decomposition, inconsistent execution standards and weak full-process control capabilities, from the top level. It can also establish a unified execution specification and data benchmark for cross-link collaborative operations of multiple intelligent agents. At the same time, relying on the shared data space, it realizes the unified flow and shared storage of full-process placement data, creative resources, strategy configuration and execution records, completely breaks down the data barriers of each link of advertising placement, and avoids the optimization disconnect problem caused by information silos. Building upon this foundation, a creative production agent, based on standardized task cards, automates the generation, compliance verification, and performance screening of advertising creatives that meet platform specifications and compliance requirements, forming a standardized candidate creative package. Then, a target audience and strategy agent simultaneously automates the generation of candidate group targeting strategies and campaign execution strategies, achieving a one-to-one correspondence and compliant tailoring of creatives, audience targeting strategies, and campaign execution strategies. This results in a list of feasible candidate campaign combinations, significantly reducing manual operation costs and reliance on human experience in advertising creative production and strategy design. It addresses the issues of disconnect between creative and strategy optimization and the difficulty of cross-platform strategy adaptation in traditional advertising. Furthermore, it provides a standardized, traceable, and variable-controllable minimum candidate campaign combination for subsequent multi-dimensional A / B testing, ensuring the accuracy and scientific rigor of campaign performance comparison tests. Subsequently, the ad placement agent automates the creation and tagging of cross-platform campaigns based on a list of candidate placement combinations. Simultaneously, it pulls and standardizes real-time placement data from various platforms, conducting multi-dimensional parallel A / B testing against a benchmark placement combination. Within preset constraints, it automatically iterates and adjusts bids and budgets, enabling intelligent and automated control of the entire ad placement execution process. This solves the problems of low efficiency, delayed parameter adjustments, and uncontrollable costs and pace associated with traditional ad placement methods involving multiple accounts across platforms. Furthermore, preset constraints ensure the stability and compliance of placement adjustments.Ultimately, by aggregating all relevant data across the entire campaign chain through monitoring and evaluation layers, the system performs multi-dimensional, multi-level performance attribution analysis across creative content, audience targeting strategies, campaign execution strategies, and advertising platforms. This accurately quantifies the conversion contribution and campaign performance at each stage. Furthermore, each intelligent agent updates its knowledge base and iteratively optimizes benchmark campaign combinations based on the effective attribution results, forming a closed-loop self-optimization system encompassing "campaign execution - performance attribution - strategy iteration - optimized campaign." This system not only enables continuous iteration of campaign execution strategies and steady improvement in campaign performance but also allows for the systematic accumulation and reuse of high-quality campaign experience. It addresses the problems of inaccurate performance attribution, lack of data support for optimization actions, and the inability to systematically reuse high-quality campaign experience in traditional campaigns. The overall solution combines excellent campaign performance, execution efficiency, cost control, reusability, and compliance.

[0009] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] Figure 1 This is a flowchart of a multi-agent collaborative advertising delivery optimization method provided in an embodiment of this application.

[0012] Figure 2 This is a flowchart illustrating the steps of a creative creation agent, provided in one embodiment of this application, to generate and filter candidate creative packages based on corresponding standardized task cards, audience profiles, advertising platform material specifications, and compliance constraints.

[0013] Figure 3 This application provides a flowchart of the steps for a target audience and strategy agent to generate a candidate group targeting strategy and a candidate delivery execution strategy based on a corresponding standardized task card, and to combine and tailor the candidate group targeting strategy and the candidate delivery execution strategy with a candidate creative package to generate a candidate delivery combination list.

[0014] Figure 4 This is a flowchart illustrating the steps of an advertising agent, provided in one embodiment of this application, to create a cross-platform advertising plan based on a task order and a list of candidate advertising combinations, and to write the mapping relationship between the unique identifiers of each advertising platform and the unique identifiers of the candidate advertising combinations into a shared data space.

[0015] Figure 5This application provides an embodiment of an advertising agent that, based on a campaign order, pulls and processes real-time campaign data from various advertising platforms, performs multi-dimensional A / B testing on each candidate campaign combination in the candidate campaign combination list using a benchmark campaign combination as a reference, automatically adjusts bids and budgets within preset constraints, and writes the parameter adjustment records into a shared data space.

[0016] Figure 6 This application provides a flowchart of the steps for a monitoring and evaluation layer that collects relevant data from various advertising platforms based on a delivery task order, performs multi-level effect attribution analysis on the delivery-related data according to the dimensions of creative content, audience targeting strategy, delivery execution strategy, and advertising platform, generates attribution results, and writes them into a shared data space.

[0017] Figure 7 This is a flowchart illustrating the steps of each agent updating the knowledge base based on attribution results and updating the best-performing candidate deployment combination in the current deployment cycle as the benchmark deployment combination, according to an embodiment of this application.

[0018] Figure 8 This is a structural block diagram of a multi-agent collaborative advertising delivery optimization system provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, implementation methods, and advantages of this application clearer, exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these exemplary embodiments are provided to make the description of this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. It should be noted that the brief descriptions of terminology in this application are merely for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0020] In the description of this application, it should be understood that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include one or more features.

[0021] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Terminology Explanation: A / B testing, also known as a controlled experiment, refers to a method of conducting parallel campaigns within the same campaign period, placing a baseline campaign mix (control group) and multiple candidate campaign mixes (experimental groups) under the same campaign environment. Only the differences in single or combined variables in creative content, audience targeting strategies, campaign execution strategies, and advertising platforms are controlled. By comparing the core performance indicators such as impressions, clicks, conversions, costs, and ROI (return on investment) of each group, the impact of different variables on the campaign performance is quantitatively evaluated, thereby selecting a better campaign strategy.

[0024] Baseline Combination: Also known as the baseline combination, it is a unified benchmark for multi-dimensional A / B testing. It is initially determined by the operations manager's intelligent agent based on the delivery task list. After each round of delivery cycle, the candidate delivery combination that performs best in this round and meets the statistical significance requirements is iterated into a new baseline delivery combination to ensure the scientific nature of A / B testing and the accuracy of optimization direction.

[0025] Large Language Model: This is an artificial intelligence model based on a deep learning architecture, pre-trained on a large-scale text corpus, capable of understanding natural language, generating text, logical inference, and content creation. In this solution, it is mainly used to automatically generate and optimize advertising copy and creative content based on audience profiles, targeting objectives, and advertising platform specifications, thereby improving creative production efficiency and standardization.

[0026] Intelligent agents are intelligent functional entities with independent perception, autonomous decision-making, automatic execution, and continuous learning capabilities. In this solution, intelligent agents are given clear business responsibilities and can independently complete corresponding tasks in the entire advertising delivery process based on standardized task cards and shared data space information. They can also iteratively optimize behavioral strategies based on performance feedback. Multiple intelligent agents collaborate and share data to form an automated, closed-loop advertising delivery optimization system.

[0027] Shared Data Space: A centralized data carrier providing unified data interaction and storage for multi-agent collaborative operations. It centrally stores structured data across the entire process, including task orders, standardized task cards, candidate creative packages, campaign combinations, advertising platform identifier mappings, real-time campaign data, parameter tuning records, attribution results, and a knowledge base. Each agent and the monitoring and evaluation layer reads and writes this data according to unified permissions and format specifications, ensuring end-to-end data interoperability, traceability, and consistency, avoiding information silos, and providing a unified data foundation for collaborative execution, A / B testing, performance attribution, and strategy iteration.

[0028] The current advertising industry largely relies on human experience for strategy formulation, creative production, and campaign optimization, resulting in cumbersome cross-platform operations, low collaboration efficiency, and lagging strategy iteration. Traditional campaign models suffer from non-standard task breakdowns and fragmented data chains, making it difficult to achieve a systematic combination and refined comparison of creative, audience, and campaign execution strategies. Furthermore, the attribution of campaign performance is vague, and best practices are difficult to retain and reuse, leading to high costs, unstable results, and an inability to meet the demands of continuous automated optimization. As multi-platform campaigns and intelligent operations become industry trends, manual management models are no longer suitable for large-scale, high-concurrency, and time-sensitive campaign scenarios.

[0029] Based on this, this application proposes a multi-agent collaborative advertising optimization method and system. Through a clearly defined agent system, a unified shared data space, and a closed-loop self-iterative mechanism, an efficient, stable, explainable, and scalable advertising optimization system is constructed.

[0030] Reference Figure 1 , Figure 1 This is a flowchart of a multi-agent collaborative advertising delivery optimization method provided in an embodiment of this application, including but not limited to steps S110 to S180.

[0031] In step S110, the operations manager intelligent agent receives the goals and constraints of the advertising campaign, parses the goals and constraints into machine-readable indicators and constraints, generates a structured campaign task sheet, and writes it into the shared data space.

[0032] This step is the top-level target input and standardized parsing stage in the multi-agent collaborative advertising delivery optimization process. It is executed by the Operations Agent, a system component that uses a large language model as its core inference engine and combines it with tool invocation capabilities to undertake global target parsing, task coordination, and collaborative scheduling for advertising delivery. The core function of this step is to transform the unstructured delivery requirements described in natural language by the business side into machine-readable metrics and constraint rules that can be recognized and executed by all agents in the system. This generates standardized structured delivery task sheets and writes them into the shared data space, providing a unified and unique top-level execution basis for subsequent creative production, candidate targeting strategy design, cross-platform delivery execution, and other processes. This solves the technical problems of ambiguous target communication and inconsistent execution standards in traditional advertising delivery from the source.

[0033] The operations manager's intelligent agent first receives the advertising placement goals and constraints input by the business side / marketing personnel through the system's preset interactive interface. The placement goals are the advertising placement effect that the business side expects to achieve, and the constraints are the hard boundary rules and execution requirements that must be strictly followed during the placement process. Both are input in natural language and can be adapted to various advertising placement business scenarios such as credit, auto finance, and e-commerce.

[0034] Taking credit product advertising as an example, the goals input by the business side can include improving the customer acquisition conversion rate of credit products, reducing the cost of leads for auto finance business by 20%–32%, and ensuring that the overall return on investment ranks among the top 2 of the partner advertisers. Input constraints can include a daily total budget not exceeding 50,000 yuan, the proportion of advertising on a major short video platform not exceeding 50% of the total budget, the content of the advertising must comply with the risk control and compliance rules of the financial industry, and the data feedback delay from each advertising platform must not exceed 3 hours. If the goals and constraints received by the operations manager's intelligent agent have issues such as missing content, logical conflicts, or ambiguous wording, it will proactively initiate a response to the business side, requesting them to supplement, confirm, or correct the relevant content to ensure the completeness, rationality, and feasibility of the initial advertising requirements.

[0035] After obtaining the confirmed complete campaign objectives and constraints, the operations manager's intelligent agent invokes a Large Language Model (LLM) to perform structured parsing of natural language into machine-readable metrics and constraints. The parsing process follows the principle of "hierarchical objectives and categorized constraints," breaking down campaign objectives into primary optimization metrics and secondary optimization metrics. Primary optimization metrics are the core performance indicators assessed by the business side, while secondary optimization metrics are the process indicators that need to be monitored during the campaign. Simultaneously, constraints are broken down into resource constraints, compliance and risk control constraints, and data operation constraints. Resource constraints are the rules for allocating resources such as budget and channels during the campaign; compliance and risk control constraints are the industry regulations and business risk control requirements that must be followed; and data operation constraints are the technical rules for pulling and transmitting campaign data.

[0036] Taking the credit product placement scenario as an example, after analysis by the large language model, the main optimization indicators are: Return on Investment (ROI, the ratio of revenue to cost generated by the placement) ≥ 3.5, and Cost Per Acquisition (CPA, the cost of acquiring a valid business lead) ≤ 80 yuan / lead. The auxiliary optimization indicators are: Click-Through Rate (CTR, the ratio of ad clicks to impressions) ≥ 2%, and Conversion Rate (CVR, the ratio of valid leads to ad clicks) ≥ 5%. Resource constraints are: total daily budget ≤ 50,000 yuan, and short video platform channel allocation ≤ 50%. Compliance and risk control constraints are: placement materials must not contain false revenue statements, and targeting of individuals with poor credit history must be excluded. Data operation constraints are: data retrieval granularity 15 minutes / time, and data latency tolerance threshold 3 hours. After analysis, the operations manager's intelligent agent will logically verify all machine-readable indicators and constraints to ensure no conflicts between indicators and no contradictions between constraints. Only after successful verification can the next stage be proceeded.

[0037] The verified machine-readable metrics and constraints will be integrated by the operations manager agent to generate a standardized, structured campaign task sheet (GoalSpec). This task sheet is a machine-readable file with a fixed data structure and serves as the top-level basis for all subsequent operations by all agents in the system. Its core fields cover all dimensions of content, including campaign business objectives, quantitative metrics, constraint rules, and technical configurations. Each field has a clearly defined quantitative value or executable rule, without any ambiguity. The core fields of the campaign task sheet include at least the core business objectives, main optimization metrics and quantitative thresholds, auxiliary optimization metrics and quantitative thresholds, all-dimensional constraint rules, campaign cycle and performance evaluation window, list of advertising platforms, performance attribution window, real-time data retrieval granularity, basic A / B testing configuration, and anomaly fallback strategy. Each field precisely corresponds to subsequent multi-agent operations. For example, the basic A / B testing configuration provides direct basis for the candidate campaign combination design of the target audience and strategy agents, and for the cross-platform A / B testing execution of the advertising agent. The anomaly fallback strategy sets safety boundaries for the automated parameter tuning of the advertising agent.

[0038] The operations manager agent generates and verifies structured campaign task orders, writing them into the shared data space according to the system's pre-defined unified data read / write rules. This shared data space serves as a centralized data carrier for multi-agent collaborative operations, used for unified storage, reading, writing, and interaction of campaign data throughout the entire process. All agents and the monitoring and evaluation layer operate on this data according to unified rules. During the writing process, the operations manager agent assigns a unique global task identifier to each campaign task order. This identifier serves as the core identifier for subsequent data association, traceability, and auditing throughout the entire campaign process, ensuring a unique association between the campaign task order and all data across the entire chain, including creative packages, candidate targeting strategies, campaign execution data, and performance attribution results. After the campaign task order is written into the shared data space, the system automatically triggers a data update notification, synchronizing it to the creative production agent, the audience targeting and strategy agent, the ad placement agent, and the monitoring and evaluation layer. Each participant can accurately read the corresponding campaign task order content from the shared data space based on this unique global task identifier, ensuring that all agents in the system collaborate under unified goals, metrics, and constraints, achieving standardized and regulated control of the entire advertising campaign process from the top level.

[0039] In step S120, the operations manager agent determines the baseline delivery combination based on the delivery task order, and breaks down the delivery task into standardized task cards corresponding to the creative production agent, the audience targeting and strategy agent, and the advertising agent, and issues them out. The standardized task card includes at least the source of input data, output data fields, quantitative conditions for result acceptance, and task execution deadline.

[0040] This step is the core of the global task planning and distribution in the multi-agent collaborative advertising optimization method, and is executed by the operations manager agent. This step uses a structured delivery task sheet as the sole basis for execution, successively determining the baseline delivery combination, finely breaking down the global delivery tasks, and generating and distributing standardized task cards. It defines the operational boundaries and clarifies execution standards for the creative production agent, the target audience and strategy agent, and the advertising agent, forming the foundation for multi-agent collaborative operations. This solves the technical problems of unclear task division, inconsistent execution requirements, and lack of standardized basis for cross-role collaboration in traditional advertising.

[0041] The operations manager's AI agent first determines the baseline ad mix for this round of advertising based on core information recorded in the ad delivery task sheet, such as the delivery cycle, optimization metrics, total budget, and channel allocation constraints. This baseline ad mix serves as a unified benchmark for subsequent multi-dimensional A / B testing and is a reference standard for evaluating the performance of all candidate ad mixes. Its determination rules are adapted to the actual ad delivery scenario and are divided into two scenarios: If it is the first execution of this round of advertising, the operations manager's AI agent will generate an initial baseline ad mix algorithmically or manually, based on the optimization metrics in the ad delivery task sheet, industry experience, and historical ad delivery data in the feature layer. This mix includes basic creative content, audience targeting strategies, and execution strategies that are suitable for the ad delivery objectives. If it is an iterative execution of an ad delivery cycle, the operations manager's AI agent will directly use the candidate ad mix from the previous cycle that has been validated, meets preset statistical significance requirements, and has the best optimization metrics performance as the baseline ad mix for this round. Taking cross-platform advertising for credit products as an example, and considering the optimization target of "reducing the cost of auto finance leads by 20%-32%" in the campaign task sheet, if it is the first campaign, the benchmark campaign combination can be determined as "low-interest credit selling point creative + targeting strategy for young and middle-aged people aged 25-45 in first-tier and new first-tier cities with good credit + CPC (Cost Per Click) bidding, with an initial bid range of 1.0-1.5 yuan". If it is an iterative campaign, the candidate campaign combination with the lowest cost of acquiring auto finance leads and the target return on investment in the previous campaign will be used as the benchmark campaign combination for this round, ensuring the scientific nature, continuity and comparability of the multi-dimensional A / B test benchmark.

[0042] After determining the baseline campaign mix, the operations manager's intelligent agent first determines the experimental scale, boundary parameters, and overall campaign plan draft for this campaign cycle based on the campaign task sheet. Simultaneously, it initializes the candidate campaign mix list skeleton in the shared data space, defining the unique identifier rules for each candidate campaign mix. The experimental scale and boundary parameters include the number of candidate creatives, the number of candidate audience targeting strategies, and the maximum number of candidate campaign mixes per ad platform. The overall campaign plan draft clarifies core content such as the target audience profile, ad platform channel allocation scheme, and preliminary budget allocation scheme. The candidate campaign mix list skeleton is a basic data structure containing a unique identifier for each candidate campaign mix, associated ad platform, creative identifier to be filled, audience targeting strategy identifier to be filled, and campaign execution strategy identifier to be filled. The unique identifier rule can be set in the format of "campaign task unique identifier + version number + unit sequence number," laying the foundation for subsequent full-process data association and traceability.

[0043] After completing the above planning, the operations manager agent, based on the campaign task list and combining the exclusive business responsibilities of the creative production agent, the audience targeting and strategy agent, and the advertising agent, meticulously breaks down the overall advertising campaign task according to the principles of matching responsibilities, unifying goals, and quantifiable execution. Each sub-task revolves around the core optimization indicators of the campaign task list and precisely corresponds to the operational capabilities of each agent, with no overlapping or omissions in responsibilities. Based on the meticulously broken down sub-tasks, the operations manager agent generates a unique standardized task card for each of the three types of agents. The standardized task card serves as the direct execution basis for each agent to carry out its exclusive work. Its core fields include at least the input data source, output data field, quantitative conditions for result acceptance, and task execution deadline. Each field is set based on the campaign task list, with no ambiguity, and precisely adapts to the data structure of the shared data space, ensuring that each agent can directly read input data from the shared data space and directly write output data into the shared data space to achieve end-to-end information exchange.

[0044] The following examples illustrate the core configurations of standardized task cards for each AI agent in a real-world scenario of cross-platform deployment of credit products. Each configuration aligns with the core requirements of the deployment task order: "7-day deployment period, total budget of 50,000 yuan, short video platform share not exceeding 50%, and customer acquisition cost for auto finance leads ≤ 80 yuan / lead." The standardized task card for the creative production agent: Input data sources include structured delivery task sheets in the shared data space, credit product audience profiles in the data and feature layer, material specifications for various advertising platforms, and compliance constraints in the financial industry. Output data fields include unique candidate creative identifiers, creative format, main selling points of the creative, target audience tags, compliance verification results, performance prediction scores, and creative metadata. The quantitative criteria for acceptance are: generating no fewer than 25 candidate advertising creatives, achieving a 100% compliance verification pass rate, retaining no fewer than 20 after homogenization and deduplication, and ensuring that creatives adapted to short video, news feed, and search advertising platforms account for no less than 40%, 30%, and 20% respectively; the task execution deadline is 10 hours after the delivery task starts.

[0045] The standardized task cards corresponding to the audience targeting and strategy intelligence agents are as follows: Input data sources include structured delivery task sheets, baseline delivery combinations, audience profiles and feature libraries in the data and feature layers, and targeting configuration rules for each advertising platform from the shared data space. Output data fields include unique identifiers for candidate group targeting strategies, audience selection rules, audience exclusion rules, expected reach, cross-platform targeting configuration parameters, unique identifiers for candidate delivery execution strategies, bid type, initial bid upper and lower limits, exposure frequency control rules, delivery time period rules, and a list of candidate delivery combinations. The quantitative criteria for results acceptance are generating no fewer than 10 candidate group targeting strategies, each candidate group targeting strategy matching no fewer than 3 candidate delivery execution strategies, generating no fewer than 50 valid candidate delivery combinations after trimming, and all candidate group targeting strategies achieving cross-platform targeting configuration mapping. The task execution deadline is 8 hours after the creative creation intelligence agent completes the generation of candidate creative packages and writes them to the shared data space.

[0046] The standardized task card for the advertising agent consists of input data from structured campaign orders, baseline campaign combinations, candidate campaign combination lists, and API configuration rules for various advertising platforms, all sourced from the shared data space. Output data fields include a unique cross-platform campaign plan identifier, the mapping relationship between candidate campaign combinations and platform plans, initial bid configuration, initial budget configuration, parameter tuning records, and campaign data collection logs. The quantifiable criteria for acceptance are: creation of cross-platform campaign plans for all valid candidate campaign combinations with 100% configuration accuracy; initial bids and budgets not exceeding the constraints of the campaign order and candidate campaign execution strategies; and the first real-time data retrieval from each advertising platform within one hour of campaign creation. The task execution deadline is six hours after the AI ​​agent (both the target audience and strategy agent) generates the candidate campaign combination list and writes it to the shared data space.

[0047] After generating all standardized task cards, the operations manager agent will distribute the task cards according to the system's preset communication mechanism. Simultaneously, it will write the complete content of all standardized task cards into the shared data space and assign a unique identifier to each card, ensuring full traceability and auditability throughout the standardized task card process. Upon receiving their corresponding standardized task card, each agent will accurately read the complete task requirements from the shared data space using this unique identifier. They will strictly adhere to the input data sources to obtain the necessary data for the task, standardize the output format of the task results according to the output data fields, and complete the task before the deadline, meeting all requirements for quantitative acceptance of the results. This ensures that all agents collaboratively advance the advertising campaign under unified task standards and timelines, achieving standardized and refined management of the overall campaign and laying the foundation for efficient subsequent creative generation, strategy design, and cross-platform campaign execution.

[0048] In step S130, the creative production intelligent agent generates and filters candidate creative packages based on the corresponding standardized task cards, audience profiles, advertising platform material specifications and compliance constraints, and writes the unique identifier and metadata corresponding to the candidate creative packages into the shared data space.

[0049] This step involves the generation and selection of candidate ad creatives, executed by the creative production AI agent. This agent uses a large language model as its core inference engine, combined with tool-invoking capabilities, to automatically generate ad creatives, perform compliance verification, deduplication, and performance filtering. In this step, the creative production AI agent first reads the input data from the corresponding standardized task card. Combining this data with the user profiles from the feature layer, the material specifications of various external advertising platforms, and the compliance constraints of relevant industries such as finance and e-commerce, it generates multiple versions of candidate ad creatives using the large language model. Then, it performs multi-dimensional compliance verification on all candidate ad creatives, eliminating those that do not comply with industry regulatory requirements, advertising platform material specifications, business risk control rules, and the compliance constraints of the placement task sheet. Next, for the candidate ad creatives that pass the compliance verification, it calculates similarity based on the creative's core selling points, copy structure, visual elements, and target audience tags, eliminating duplicate creatives with similarity exceeding a preset threshold. Finally, it processes the deduplicated candidate ad creatives... The advertising creatives are evaluated for performance, resulting in estimated click-through rate, conversion rate, and ROI scores. Candidate advertising creatives that meet the standardized task card quantity requirements are selected based on their scores from highest to lowest, packaged into a candidate creative package, and assigned a unique identifier. Simultaneously, metadata such as creative format, main selling points, target audience tags, compliance verification results, and performance prediction scores are compiled and written into a shared data space (a centralized data carrier for multi-agent collaborative operations, used for unified storage, reading, writing, and interaction of data throughout the entire process). This provides a standardized and traceable creative material foundation for subsequent creative and strategy combinations by the target audience and strategy agents.

[0050] Reference Figure 2 , Figure 2 This application provides a flowchart of the steps for a creative production intelligence agent to generate and filter candidate creative packages based on corresponding standardized task cards, audience profiles, advertising platform material specifications and compliance constraints, including but not limited to steps S210 to S240.

[0051] In step S210, the creative production agent reads the input data of the corresponding standardized task card, combines the audience profile, advertising platform material specifications and compliance constraints, and generates multiple versions of candidate advertising creatives through a large language model.

[0052] This step is the core starting point for generating candidate creative packages, executed by the creative production intelligent agent. Its core function is the intelligent generation of multiple versions and formats of candidate ad creatives based on the needs of the advertising scenario, laying the foundation for subsequent creative selection. In this step, the creative production intelligent agent first reads all the input data from the standardized task card assigned to it through the shared data space, ensuring that the creative generation direction is highly aligned with the requirements of the advertising task. Then, it retrieves audience profiles (structured data depicting the age, region, consumption habits, and needs of the target audience), material specifications published by various external advertising platforms, and compliance rules of the industry to which the advertising business belongs from the data and feature layer. Using this information as core input, it calls the large language model to generate creatives. The generated candidate ad creatives are differentiated content in multiple versions and formats, covering mainstream ad creative types such as ad copy, image scripts, and video scripts. Each creative also includes basic annotation information such as the target audience, core selling points, and compatible advertising platforms, initially meeting the basic requirements of the standardized task card for the number and format of creatives.

[0053] Taking the cross-platform deployment scenario of credit products as an example, the creative production intelligence agent takes the demographic profile of "young and middle-aged people aged 25-45 in first-tier and new first-tier cities with small loan needs and good credit" as input, the platform material specifications of short video platforms ("copywriting should not exceed 50 words, video material length should be 15-30 seconds"), the platform material specifications of search platforms ("copywriting should highlight core selling points and be adapted to keyword search"), and the compliance constraints of the financial industry ("cannot exaggerate loan amounts, cannot promise fixed returns, and cannot target students") as input, and generates multiple versions of creative content through a large language model. Among them, the short video platform creative content... The app can generate short copy and a matching 15-second video script for "Low-interest, fast loans, instant approval for those with good credit, exclusively for urban professionals aged 25-45". The search platform can generate keyword-matched copy for "Low-interest loans for urban youth with good credit, fast disbursement, and no hidden fees". It can also generate long copy, poster scripts, and other content formats adapted to information flow platforms. Each version of the creative is marked with basic information such as "Suitable for urban youth aged 25-45", "Core selling points: low interest, fast disbursement", and "Compatible platforms: short video / search / information flow", ensuring accurate matching of creative with the target audience.

[0054] Step S220 involves performing multi-dimensional compliance checks on all candidate ad creatives to eliminate those that do not comply with industry regulatory requirements, ad platform material specifications, business risk control rules, and compliance constraints of the placement task sheet.

[0055] This step is the compliance screening of candidate ad creatives. Following the creative generation results of step S210, it is executed by the creative production AI. The core of this step is to eliminate non-compliant creatives through comprehensive rule verification, mitigating compliance risks, platform rule violations, and business risk control risks in ad placement from the source. This ensures that all creatives entering the screening stage meet full-scenario compliance requirements. In this step, the creative production AI activates a preset rule engine and keyword search mechanism for all multi-version, multi-format candidate ad creatives generated in step S210. It conducts a four-dimensional full compliance verification covering industry regulatory requirements, ad platform material specifications, business risk control rules, and compliance constraints of the placement task sheet. These four verification dimensions are independent and indispensable; any creative that fails any dimension will be directly eliminated, ultimately achieving a 100% compliance verification pass rate for candidate ad creatives in this step. Among them, industry regulatory requirements are the national and industry-level regulatory standards for the industry to which the advertising business belongs; advertising platform material standards are the hard rules such as the format, content, and specifications of the materials released by various external advertising platforms; business risk control rules are the advertising placement risk control guidelines formulated internally by the advertising entity; and the compliance constraints of the placement task order are the personalized compliance requirements specified by the operations manager's intelligent agent in the placement task order for this placement.

[0056] Taking the cross-platform advertising scenario of credit products as an example, the specific execution logic for the four-dimensional compliance verification of the candidate ad creatives generated in step S210 is as follows: In terms of industry regulatory requirements, creatives containing phrases such as "100% loan disbursement," "no credit check required," and "interest-free" that violate financial advertising regulations and exaggerate the efficacy of credit products are eliminated. In terms of advertising platform material specifications, creatives with copy exceeding 50 characters or video scripts exceeding 15-30 seconds in length on short video platforms, and creative copy on search platforms that does not conform to the platform's keyword search rules, are eliminated. In terms of business risk control rules, creatives targeting students or people with poor credit histories, as well as those implying high interest rates or hidden fees, violate the internal risk control principles of the credit business. In terms of compliance constraints of the advertising task list, creatives that violate the personalized compliance requirements of this advertising task list, such as "not targeting people in lower-tier markets without fixed income" and "the main selling point of the creative must highlight low interest rates and fast loan disbursement," are eliminated. Through the above multi-dimensional and meticulous compliance verification, the platform can completely avoid problems such as platform rejection, industry regulatory penalties, and business risk control violations that may occur during subsequent advertising campaigns, thus laying a compliant foundation for the smooth implementation of subsequent advertising campaigns.

[0057] Step S230: For candidate ad creatives that have passed the compliance verification, calculate the similarity based on the creative's core selling point, copy structure, visual elements, and target audience tags, and remove duplicate creatives with similarity exceeding a preset threshold.

[0058] This step is the differentiation screening of candidate ad creatives. Following the compliance verification results of step S220, it is executed by the creative creation intelligent agent. The core is to eliminate homogeneous and duplicate creatives through multi-feature dimension similarity calculation, so as to avoid traffic competition and resource waste caused by highly similar creatives in subsequent ad placement. This ensures the diversity and differentiation of candidate creatives and provides rich creative samples for subsequent multi-dimensional A / B testing. In this step, the creative production agent first extracts the core feature dimensions of all candidate advertising creatives that have passed the compliance verification in step S220. The extraction dimensions cover four key aspects: the core selling point of the creative, the copy structure, visual elements, and the target audience tags. These four features are the core dimensions that determine the presentation effect and the suitability of the advertising creative, and can accurately represent the differentiated attributes of the creative. Then, the preset similarity algorithm is called to quantify the similarity between the extracted features of any two creatives, and obtain the similarity value between each pair of creatives. The value is then compared with the preset similarity threshold of the system. If the similarity value exceeds the preset threshold, the two sets of creatives are determined to be homogeneous and duplicate creatives, and the one with the relatively lower quality score and effect prediction score is removed. Only the high-quality and differentiated creatives are retained to enter the next stage. The preset similarity threshold can be flexibly configured according to the creative diversity requirements of the campaign task. Under normal campaign scenarios, the threshold can be set to 80%.

[0059] Taking the cross-platform deployment of credit products as an example, for candidate creatives that have passed the compliance verification in step S220, the specific execution logic of this step is as follows: Extract the core features of Creative A "Low-interest, fast loan, instant approval for those with good credit, exclusive to urban office workers aged 25-45" and Creative B "Low-interest, fast loan, extremely fast disbursement for those with good credit, exclusive to urban young and middle-aged people aged 25-45". The core selling points of both are "low interest, fast disbursement, and approval for those with good credit", the copywriting structure is a short sentence structure of "selling point + target audience", the visual elements are both adapted to the minimalist poster style of short video platforms, and the target audience tag is urban people aged 25-45 with stable income. The similarity between the two is calculated to be 92%, which exceeds the preset threshold of 80%, and is judged as homogeneous and duplicate creatives. At this time, the creative production agent will retrieve the estimated performance scores of the two. If the estimated click-through rate of Creative A is 5.2% and the estimated click-through rate of Creative B is 4.8%, then Creative B will be removed, and only Creative A will be retained to ensure that the creatives entering the next stage have significant differentiation. As for Creative C, "Small Loans with Flexible Repayment, Low-Interest Solution Exclusively for Individual Businesses," its core selling point is "flexible repayment, exclusively for individual businesses," and its target audience is small and micro-sized business owners / individual businesses. Its core feature similarity to Creative A is only 35%, far below the preset threshold, so it will be retained. This will achieve a diversified distribution of candidate creatives in terms of core selling points, target audience, and presentation format, providing rich creative variables for subsequent creative and strategy combinations for audience targeting and strategic intelligence, as well as multi-dimensional A / B testing.

[0060] Step S240: Estimate the performance of the deduplicated candidate ad creatives, obtain the estimated click-through rate, estimated conversion rate, and estimated return on investment scores for each creative, and select the candidate ad creatives that meet the number requirements of the standardized task card according to the scores from high to low, and package them into a candidate creative package.

[0061] This step is the final screening stage for the candidate creative package generation. Following the homogenization and deduplication results of step S230, it is executed by the creative production intelligence agent. The core is to achieve accurate screening of high-quality creatives through quantitative prediction of the delivery effect, and finally form a candidate creative package that meets the requirements of the standardized task card. This provides high-potential, standardized creative materials for the subsequent creative and strategy combination of the target audience and strategy intelligence agents. In this step, the creative production agent first uses the massive creative effect data from the historical campaign log library in the data and feature layer to call the preset creative effect prediction model to conduct a comprehensive quantitative evaluation of the campaign effect of all candidate ad creatives after deduplication in step S230. Finally, it outputs three core scores for each creative: estimated click-through rate (CTR), estimated conversion rate (CVR), and estimated return on investment (ROI). All three scores are quantitative indicators on a 100-point scale. The higher the value, the stronger the potential of the creative in the corresponding dimension. The estimated click-through rate is the probability estimate of the user click after the creative is exposed. The estimated conversion rate is the probability estimate of the business conversion (such as lead generation or product order) after the creative is clicked. The estimated return on investment is the ratio of the revenue generated after the creative is launched to the cost of the campaign. The three indicators together constitute a comprehensive evaluation system for creative effect. Subsequently, the creative production agent comprehensively calculates the three scores according to the preset weight ratio to obtain the comprehensive effect score of each candidate creative. Then, it sorts all candidate creatives from high to low according to the comprehensive effect score, and finally selects the number of candidate advertising creatives that meet the quantitative conditions for result acceptance in the standardized task card. These are then integrated and packaged into a standardized candidate creative package. Each creative in the package is associated with its unique identifier, various effect scores, target audience, and compatible platform, etc., and complete metadata.

[0062] Taking a cross-platform ad campaign for credit products as an example, after deduplication in step S230, 30 compliant and differentiated candidate ad creatives are retained. The standardized task card requires that at least 20 candidate creatives be selected to form a creative package for this campaign. The creative creation agent evaluates these 30 creatives using a creative effect prediction model. Creative 1 has an estimated click-through rate of 8.5, an estimated conversion rate of 7.8, and an estimated return on investment of 8.2. Calculated using a weighted ratio of "30% click-through rate, 40% conversion rate, and 30% return on investment," its overall performance score is 8.01. Creative 2 has an estimated click-through rate of 7.2, an estimated conversion rate of 6.5, and an estimated return on investment of 7.0, with an overall performance score of 6.81. The remaining creatives are calculated using the same overall performance score. Subsequently, the 30 creatives were ranked from highest to lowest based on their overall performance scores. The top 20 creatives were then integrated and packaged to form the candidate creative package for this campaign. The package includes both high-click-through-rate short copy creatives adapted to short video platforms and high-conversion-rate keyword copy creatives adapted to search platforms. It also covers creative types that address different core selling points and cater to different audience segments. This approach not only meets the quantity requirements of the standardized task cards but also ensures the overall performance potential and diversity of the creative package, providing high-quality and abundant creative samples for subsequent multi-dimensional A / B testing.

[0063] In this embodiment, the creative production agent relies on a large language model to automatically generate multiple versions of advertising creatives. Combined with audience profiling, platform material specifications, and compliance constraints, the creatives are adapted to the placement scenarios and business requirements from the source. Then, through a progressive process of multi-dimensional compliance verification, homogenization deduplication, and effect prediction screening, on the one hand, a rule engine and keyword search are used to achieve full-dimensional compliance verification, ensuring that all generated creatives comply with industry regulations, platform specifications, business risk control, and placement task requirements, thus mitigating compliance risks in placement from the root. On the other hand, by calculating the similarity of the core features of the creatives, homogenized creatives are eliminated, ensuring the diversity and differentiation of candidate creatives and avoiding the loss of placement funds. This approach eliminates the ineffective waste of creative resources and enables quantitative scoring and precise screening of creative ideas based on historical data and performance prediction models. The resulting candidate creative package not only meets the quantity requirements of standardized task cards but also ensures that the creatives possess high-quality placement potential. This significantly improves the production efficiency and quality of advertising creatives, replacing the inefficient traditional manual creative production and screening model. It reduces labor costs and reliance on experience, and all creatives come with standardized basic annotation information. This provides a compliant, high-quality, standardized, and traceable creative material foundation for subsequent creative and strategy combinations for target audience and strategy agents, as well as the creation of cross-platform placement plans. This ensures the efficient implementation of subsequent multi-agent collaborative operations and multi-dimensional A / B testing.

[0064] In step S140, the target audience and strategy agent generates candidate group targeting strategies and candidate delivery execution strategies based on the corresponding standardized task cards. The candidate group targeting strategies and candidate delivery execution strategies are then combined and tailored with candidate creative packages to generate a candidate delivery combination list. This list is then output to the advertising agent and written into the shared data space. Each candidate delivery combination in the candidate delivery combination list is a one-to-one combination of a candidate creative package, a candidate group targeting strategy, and a candidate delivery execution strategy.

[0065] This step involves the design and optimization of the campaign strategy combination, executed by the Audience Targeting and Strategy Intelligence Agent. This agent, with a large language model as its core inference engine, is a system component responsible for audience targeting strategy design, campaign execution strategy matching, and creative and strategy combination optimization. In this step, the Audience Targeting and Strategy Intelligence Agent first reads the input requirements of the corresponding standardized task card, retrieves audience profiles and feature database data from the data and feature layers, and generates multiple sets of candidate group targeting strategies. Each strategy includes audience selection rules, audience exclusion rules, expected reach scale, and risk control compliance tags. Simultaneously, each candidate group targeting strategy is mapped to the native targeting configuration parameters that can be executed by each advertising platform, achieving cross-platform implementation of general audience targeting rules. Next, for each candidate group targeting strategy that has completed cross-platform mapping, at least one candidate delivery execution strategy is matched. This candidate delivery execution strategy includes at least the bidding type, initial bid upper and lower limits, exposure frequency control rules, delivery time rules, and budget consumption rhythm control rules. Then, each candidate ad creative in the candidate creative package generated by the creative production agent is systematically orthogonally combined with each candidate group targeting strategy and the corresponding matched candidate delivery execution strategy to form a creative and strategy candidate matrix. Then, based on the delivery task sheet's agreed delivery combination quantity threshold, coverage requirements, effect prediction score, and constraint feasibility rules, the creative and strategy candidate matrix is ​​pruned, and invalid combinations that exceed the constraint boundaries, fail to meet the predicted effect, or have overlapping coverage are eliminated. Each valid combination retained after pruning is assigned a unique identifier and a candidate delivery combination list is formed. Each candidate delivery combination in this list is a one-to-one correspondence between the candidate creative package, the candidate group targeting strategy, and the candidate delivery execution strategy. Finally, the candidate placement combination list is synchronously output to the advertising agent and completely written into the shared data space. This provides the advertising agent with a standardized and feasible minimum candidate placement combination for creating cross-platform placement plans and conducting multi-dimensional A / B testing, while also enabling full-process traceability of placement strategy combination data.

[0066] Reference Figure 3 , Figure 3This application provides a flowchart of the steps for a target audience and strategy agent to generate a candidate group targeting strategy and a candidate delivery execution strategy based on a corresponding standardized task card, and to combine and tailor the candidate group targeting strategy and the candidate delivery execution strategy with a candidate creative package to generate a candidate delivery combination list, including but not limited to steps S310 to S360.

[0067] In step S310, the target audience and strategy intelligence agent reads the input requirements of the corresponding standardized task card, retrieves the audience profile and feature database data, and generates multiple sets of candidate group targeting strategies. Each set of candidate group targeting strategies includes audience selection rules, audience exclusion rules, expected reach scale, and risk control compliance labels.

[0068] This step is the core initial stage of audience targeting and strategy combination design, executed by the audience targeting and strategy intelligent agent. This agent is a system component that uses a large language model as its core reasoning engine and combines audience feature analysis capabilities to design advertising audience targeting strategies. Its core function is to generate multiple sets of differentiated and standardized candidate group targeting strategies based on the requirements of the advertising task and audience data, laying the foundation for the subsequent cross-platform implementation of strategies and the combination of advertising execution strategies and creative content. In this step, the target audience and strategy intelligence agent first accurately reads all the input requirements from the standardized task cards assigned to it in the shared data space, clarifying the core requirements such as the target customer group orientation, audience coverage, and risk control compliance boundaries for this campaign. Then, it retrieves the full-volume structured data of the audience profile and feature library from the data and feature layer. This data covers multi-dimensional tag information such as the target audience's age, region, occupation, consumption habits, credit preferences, and risk control characteristics. Combining the standardized task card requirements and audience profile feature data, the target audience and strategy intelligence agent generates multiple sets of candidate group targeting strategies adapted to this campaign scenario through feature analysis and audience segmentation modeling. Each strategy fully includes four core components: audience selection rules, audience exclusion rules, expected reach scale, and risk control compliance tags. These four elements form a closed-loop audience targeting rule, ensuring that each strategy is executable, quantifiable, and compliant, meeting the requirements for subsequent cross-platform campaigns and strategy combinations.

[0069] The selection criteria are positive screening conditions for defining the target customer group for this campaign. The exclusion criteria are negative screening conditions to avoid non-target customer groups and high-risk customer groups. The expected reach is the number of target customers that this strategy can reach, calculated based on the customer profile and feature database data. The risk control and compliance labels are compliance attribute labels marked on this strategy based on business risk control rules and industry compliance requirements, used for subsequent strategy verification and campaign risk control. Taking the cross-platform campaign scenario for credit products as an example, combined with the core requirement of "reducing the cost of auto finance leads and reaching high-quality credit customers" in the standardized task card, the target audience and strategy intelligence agent can generate multiple sets of differentiated candidate group targeting strategies. One specific configuration of the strategy is as follows: the selection criteria are people aged 25-45 in first-tier and new first-tier cities who have stable jobs and good credit, and who have browsed / consumed car-related content in the past 6 months. The exclusion criteria are people with poor credit, no fixed income, students, and people with high debt. The expected reach is 800,000 target customers calculated based on the customer profile data. The risk control and compliance tags are "high-quality credit customers, non-students, non-high-debt, and car-demand oriented". Simultaneously, the target audience and strategy intelligence can generate candidate group targeting strategies for different segmented customer groups, such as small and micro-enterprise owners aged 30-50 in lower-tier markets and newlyweds in second-tier cities, based on the campaign requirements. Each strategy has four complete and distinct core elements, forming diverse audience targeting solutions and providing rich audience strategy variables for subsequent multi-dimensional A / B testing.

[0070] Step S320: Map each candidate group targeting strategy to the native targeting configuration parameters that each advertising platform can execute, so as to realize the cross-platform implementation of general audience targeting rules.

[0071] This step is the platform adaptation stage for audience targeting strategies. Following the candidate group targeting strategy generated in step S310, it is executed by the audience targeting and strategy intelligence agent. This step addresses the technical issue of inconsistent targeting field names, configuration logic, and interface specifications across different advertising platforms, which prevents the direct deployment and execution of general strategies. In this step, the audience targeting and strategy intelligence agent translates and converts the parameters of each generalized candidate group targeting strategy generated in step S310 according to the native targeting systems of each external advertising platform. It maps the uniformly described audience selection rules, exclusion rules, and feature tags to native targeting configuration parameters that can be directly recognized, invoked, and deployed on the corresponding advertising platform. This ensures consistent and accurate implementation of the same audience strategy across multiple advertising channels, guaranteeing uniformity in audience targeting scope and quality standards when delivering across platforms. The native targeting configuration parameters refer to the dedicated fields, enumeration values, and combination logic defined by each advertising platform in its delivery API or backend system for defining the target audience. Different platforms typically differ in their expression of region, age, behavioral interests, device, and excluded audiences.

[0072] Taking a credit product advertising scenario as an example, a general audience targeting strategy generated in step S310 is: "Selected first-tier and new first-tier cities, aged 25-45, with stable jobs, and having browsed car-related content in the past 30 days; excluding students, those with poor credit, and those with high debt." When the audience targeting and strategy intelligence agent maps this strategy across platforms, on a short video platform, it can be mapped as: Region = First-tier / New first-tier, Age = 25-45, Occupation tag = Corporate employee / Office worker, Behavioral interest = Car information / Car consumption, Exclusion package = Students + High-risk credit groups. On a search advertising platform, it is mapped as: Region filter + Keywords, Targeting = Car loans / Low-interest loans, Audience package = High-quality credit customers, Exclusion audience package = High debt / No fixed income. On a news feed platform, it is mapped to its platform-specific audience tag ID and behavioral combination rules. Through the above mapping process, a unified audience strategy can be accurately implemented in the native systems of various advertising platforms, laying a consistent audience foundation for subsequent ad combinations and cross-platform A / B testing.

[0073] Step S330: For each candidate group targeting strategy that has completed cross-platform mapping, match at least one candidate delivery execution strategy. The candidate delivery execution strategy shall at least include bid type, initial bid upper and lower limits, exposure frequency control rules, delivery time period rules, and budget consumption rhythm control rules.

[0074] This step involves configuring the campaign strategy parameters. Following step S320, which completed the cross-platform mapping of the candidate group targeting strategy, this step is executed by the audience targeting and strategy intelligence agent. The aim is to provide each audience targeting strategy with a readily applicable set of execution parameters, ensuring that the audience strategy synergizes with bidding, frequency control, and time-of-day targeting rules, meeting the diverse strategy variable requirements of multi-dimensional A / B testing. In this step, the audience targeting and strategy intelligence agent, based on the optimization metrics, total budget constraints, exploration and evaluation period settings in the standardized task card, as well as the quality characteristics and estimated conversion potential of the corresponding audience, matches at least one candidate campaign execution strategy for each candidate group targeting strategy that has completed cross-platform mapping. This strategy is the core parameter set controlling online ad delivery behavior, including at least the bidding type, initial bid upper and lower limits, exposure frequency control rules, time-of-day targeting rules, and budget consumption rhythm control rules. The bidding type refers to the advertising platform's billing and optimization method, including CPC (cost-per-click), OCPC (cost-per-click with conversion as the optimization goal), and CPA (cost-per-conversion). The initial bid upper and lower limits are the safe range for automatic system price adjustments. Exposure frequency control rules limit the number of times a single user sees the same ad within a certain time frame, preventing overexposure that could cause user annoyance and wasted budget. Targeting time rules define the time range during which the ad can be played online. Budget consumption rhythm control rules regulate the rate at which the budget is consumed within the campaign period, such as uniform consumption, peak acceleration, or stable control.

[0075] Taking the "high-quality customer base aged 25-45 in first-tier and new first-tier cities" in credit product targeting as an example, the target audience and strategy agent can match at least two differentiated execution strategies: The first strategy uses CPC (Cost Per Click) bidding, with an initial bid range of 1.0-1.5 yuan, an exposure frequency of no more than 3 times per user per day, a campaign period of 9:00-21:00, and a uniform budget consumption rhythm. The second strategy uses OCPC bidding, targeting conversion cost, with an initial bid range of 80-100 yuan, an exposure frequency of no more than 2 times per user per day, a campaign period of 10:00-22:00, and a budget consumption rhythm of moderately increasing volume during evening traffic peaks. By configuring multiple execution strategies with different execution logics for the same target audience, a rich combination of strategies can be formed, providing sufficient variables to support subsequent multi-dimensional comparative testing and optimal strategy selection.

[0076] Step S340: Systematically orthogonally combine each candidate ad creative in the candidate creative package with each candidate group targeting strategy and the candidate delivery execution strategy that matches the audience targeting strategy to form a creative and strategy candidate matrix.

[0077] This step involves generating a full-dimensional sample pool for the campaign mix. Following step S330, which completed the matching of candidate campaign execution strategies and candidate group targeting strategies, as well as the candidate creative package generated by the creative production agent, this step is executed by the audience targeting and strategy agents. The core of this step is to integrate the three independent variables—creative, audience targeting, and campaign execution—into a complete campaign candidate solution through systematic, full-dimensional matching. This provides a comprehensive initial sample pool for subsequent combination tailoring and screening. In this step, the audience targeting and strategy agents perform systematic orthogonal combination, that is, they perform one-to-one matching of the three types of variables across all dimensions. Each candidate ad creative in the candidate creative package is completely matched with each candidate group targeting strategy and each candidate campaign execution strategy corresponding to that audience targeting strategy, ensuring no potential combination is overlooked. After all pairings are completed, a creative and strategy candidate matrix is ​​formed. This matrix is ​​a structured set of initial candidate solutions, where each matrix unit corresponds to a set of independent and complete minimum deployment units, including the unique identifier of the creative, the unique identifier of the audience targeting strategy, the unique identifier of the deployment execution strategy, the compatible platform, the performance prediction score, and other complete core parameters, ensuring that all potential deployment solutions are included in the candidate range.

[0078] Taking the cross-platform advertising scenario of credit products as an example, after the aforementioned steps, the candidate creative package retains 20 compliant and high-potential candidate ad creatives, generating 10 sets of differentiated candidate audience targeting strategies. Each audience targeting strategy is matched with 3 sets of differentiated candidate execution strategies. In the orthogonal combination in this step, the agent performs full-dimensional matching of the 20 creatives, 10 audience strategies, and 30 execution strategies (each of the 10 audience strategies is matched with 3 sets), ultimately forming 20×10×3=600 independent initial advertising candidate schemes. These schemes together constitute the creative and strategy candidate matrix. Each scheme in the matrix covers the complete configuration of the three dimensions of creative, audience, and execution. It includes combinations of low-interest creatives and CPC bidding strategies suitable for urban young and middle-aged customers, as well as combinations of business loan creatives and OCPC bidding strategies suitable for micro and small business owners. This fully covers all potential advertising adaptation possibilities, avoiding the omission of high-potential advertising combinations due to the limitations of human experience, and providing a sufficient and comprehensive initial sample foundation for the subsequent candidate matrix trimming stage.

[0079] Step S350: Based on the threshold for the number of ad combinations, coverage requirements, effect prediction scores, and constraint feasibility rules agreed upon in the ad delivery task order, the creative and strategy candidate matrix is ​​pruned to remove invalid combinations that exceed the constraint boundaries, fail to meet the predicted effects, or have overlapping coverage.

[0080] This step involves precise selection of campaign combinations. Following the creative and strategy candidate matrix containing hundreds of initial options generated in step S340, this step is executed by the user and strategy intelligence agent. Its core purpose is to address the combination explosion problem that easily occurs after orthogonal combinations (i.e., an excessive number of candidate options after full-dimensional pairing, leading to excessive budget dispersion, insufficient single-group experimental samples, and A / B testing statistical failure if all are deployed). Through multi-dimensional rules, the initial matrix is ​​precisely trimmed, eliminating invalid combinations and retaining high-potential effective combinations that meet the campaign requirements. In this step, the user and strategy intelligence agent uses the four core rules agreed upon in the campaign task sheet to perform layer-by-layer screening and elimination of all initial combinations in the creative and strategy candidate matrix: The first rule is the constraint feasibility rule, used to verify whether the combination meets the hard constraints of the campaign, eliminating combinations that exceed the total budget and channel allocation constraints, platform campaign rule restrictions, and business risk control requirements. The second item is the performance prediction score, used to filter combinations with advertising potential. Combinations with a combined score of predicted click-through rate, predicted conversion rate, and predicted return on investment below a preset threshold are eliminated, meaning combinations whose predicted performance is not up to standard. This filters out low-potential solutions with no advertising value in advance. The third item is the coverage requirement, used to optimize the rationality of audience coverage. Combinations with excessively overlapping audience coverage are eliminated to avoid internal traffic competition and budget drain between different advertising combinations. The fourth item is the threshold for the number of advertising combinations, used to control the number of experimental units launched in the final run. This avoids insufficient budget for individual units and slow accumulation of experimental data due to too many combinations, ensuring the statistical validity of multi-dimensional A / B testing.

[0081] Taking the cross-platform advertising scenario for credit products as an example, the creative and strategy candidate matrix generated in step S340 contains 600 initial advertising combinations. The specific trimming process in this step is as follows: First, based on the constraint feasibility rules, 120 combinations that exceed the budget limit of the short video platform or whose bids exceed the platform rules are eliminated, leaving 480 combinations. Then, based on the performance prediction score, 200 low-potential combinations that predict the CPA of auto finance leads to exceed the target threshold of 80 yuan are eliminated, leaving 280 combinations. Next, based on the coverage requirements, 80 redundant combinations that have an overlap of more than 70% in audience coverage and are likely to cause internal traffic competition are eliminated, leaving 200 combinations. Finally, based on the threshold of a maximum of 50 ad combinations per platform as agreed in the ad campaign task sheet, the remaining 200 combinations were sorted from high to low according to the estimated comprehensive performance score. Only the top 50 high-scoring combinations were retained, and the remaining 150 low-scoring combinations were removed. This completed the precise trimming of the initial candidate matrix, ensuring that the retained combinations met all ad campaign constraints, had high ad campaign potential, and were controllable in number and had reasonable audience coverage, thus providing controllable candidate ad combinations for subsequent ad campaign plans.

[0082] Step S360: Assign a unique identifier to each valid combination retained after cropping to form a candidate deployment combination list.

[0083] This step is the final stage of the user targeting and strategy agent combination design process. Following the effective candidate campaign combinations retained after the trimming in step S350, this step is executed by the user targeting and strategy agent. The core of this step is to integrate the scattered effective combinations into a unified structured list through standardized unique identifier allocation. This provides a standardized foundation for subsequent cross-platform campaign creation, end-to-end data association, and performance attribution. In this step, the user targeting and strategy agent first follows the unique identifier rules for candidate campaign combinations agreed upon by the operations manager agent in the shared data space (i.e., a globally unified coding rule of "unique identifier for campaign task + version number + unit sequence number"). A unique global identifier is assigned to each effective combination retained after trimming. This identifier is a unique code that is not repeated throughout the entire campaign system, used to achieve data association and alignment across agents, advertising platforms, and all stages of the process, solving the problems of inconsistent data standards and inability to associate data across different stages and platforms. After assigning identifiers, the agent populates the previously initialized candidate campaign combination list skeleton with the globally unique identifier for each valid combination, as well as the corresponding unique identifiers for creative content, audience targeting strategies, campaign execution strategies, advertising platforms, and overall performance prediction scores. This ultimately forms a standardized candidate campaign combination list. This list is a structured, ordered data set, where each entry corresponds to a complete, independently executable minimum experimental unit for campaign execution. All entries conform to a unified data format specification and can be directly read and accessed by the advertising agent.

[0084] Taking the cross-platform deployment scenario of credit products as an example, after the steps of trimming, a total of 50 valid candidate deployment combinations are retained. The unique identifier of this deployment task is AD-TASK-20250325-001, and the iteration version number is V1.0. Then, the agent assigns the globally unique identifier AD-TASK-20250325-001-V1.0-001 to the first valid combination, assigns AD-TASK-20250325-001-V1.0-002 to the second combination, and so on until the 50th combination is assigned AD-TASK-20250325-001-V1.0-050. Simultaneously, the creative ID, audience targeting strategy ID, campaign execution strategy ID, compatible platform, and estimated score of each combination are bound to the corresponding identifiers and filled into the preset list skeleton, ultimately forming a candidate campaign combination list containing 50 standardized items. This list will be synchronously output to the advertising agent and fully written into the shared data space, providing a standardized candidate campaign combination basis for the creation of subsequent cross-platform campaign plans and the conduct of multi-dimensional A / B testing, and also providing core correlation basis for cross-platform data alignment in the subsequent effect attribution stage.

[0085] In this embodiment, the audience targeting and strategy intelligence agent achieves a systematic combination and optimization of audience targeting strategies, campaign execution strategies, and advertising creatives through steps S310 to S360. On one hand, multiple sets of differentiated audience targeting strategies are generated based on audience profiles and feature libraries, and cross-platform native configuration mapping is completed to solve the problem of inconsistent targeting rules across different advertising platforms, ensuring that the strategies can be accurately implemented on each platform. At the same time, multiple sets of campaign execution strategies are matched for each audience targeting strategy, enriching the variable dimensions of the campaign strategy. On the other hand, sufficient creative and strategy samples are formed through orthogonal combination, and then precise tailoring is performed by combining various constraints and effect prediction scores of the campaign task sheet. This avoids the loss of control over the campaign caused by the explosion of strategy combinations, and ensures that the retained candidate campaign combinations meet compliance constraints, budget requirements, and have high-quality campaign potential. At the same time, a unique global identifier is assigned to each group of valid candidate campaign combinations, enabling standardized management and full-process traceability of campaign combinations. Furthermore, the final generated candidate campaign combination list is a one-to-one combination of "candidate creative + candidate group targeting strategy + candidate campaign execution strategy". This combination can be directly used as the minimum candidate campaign combination for the advertising agent to create cross-platform campaign plans. This can greatly simplify the subsequent campaign execution process, improve the efficiency of multi-agent collaborative work, and provide standardized and diversified experimental samples for subsequent multi-dimensional A / B testing. This ensures the scientific nature and reference value of A / B test results and lays a solid strategic foundation for optimizing campaign performance.

[0086] In step S150, the advertising agent creates a cross-platform advertising plan based on the advertising task order and the candidate advertising combination list, and writes the mapping relationship between the unique identifier of each advertising platform and the unique identifier of the candidate advertising combination into the shared data space.

[0087] This step involves the creation and implementation of a cross-platform campaign, executed by the advertising agent. This agent, powered by a large language model as its core inference engine and leveraging platform API calls, is a system component responsible for building cross-platform campaigns, configuring parameters, and mapping data. In this step, the advertising agent first reads the structured campaign task sheet and the candidate campaign combination list generated by the target audience and strategy agents from the shared data space. Based on the total budget and channel allocation constraints in the campaign task sheet, it plans a campaign hierarchy structure that matches the native rules of each external advertising platform. Simultaneously, it determines the platform campaign hierarchy mapping relationship for each combination in the candidate campaign combination list. Then, it converts each candidate campaign combination into executable configuration parameters for the corresponding advertising platform's campaign hierarchy. Finally, it performs compliance checks, risk control checks, budget allocation checks, frequency control, and deduplication checks on each mapped parameter, eliminating candidate campaign combinations that fail the checks. Finally, it performs checks on all combinations that pass all the checks. The system calculates and configures initial bids and initial budgets by combining the optimization metrics of the campaign task with the corresponding campaign execution strategies. Then, it calls the standardized API interfaces of each advertising platform to create corresponding platform campaign plans for the candidate campaign plans with completed parameter configurations. Finally, it obtains the unique identifier of the advertising platform corresponding to the campaign plan returned by each advertising platform, establishes a one-to-one mapping relationship between this identifier and the unique identifier of the candidate campaign plan, and writes this mapping relationship completely into the shared data space. This provides the core basis for cross-platform data association for subsequent real-time campaign data retrieval, multi-dimensional A / B testing, and performance attribution analysis, realizing the standardized connection of campaign plans from strategy design to platform implementation.

[0088] Reference Figure 4 , Figure 4 This application provides an embodiment of an advertising agent that creates a cross-platform advertising plan based on an advertising task order and a candidate advertising combination list, and writes the mapping relationship between the unique identifier of each advertising platform and the unique identifier of the candidate advertising combination into a shared data space, including but not limited to steps S410 to S450.

[0089] In step S410, the advertising agent reads the placement task list and candidate placement combination list in the shared data space, and based on the total budget and channel allocation constraints in the placement task list, plans a placement plan hierarchy structure that matches the platform's native rules for each advertising platform, and determines the platform placement hierarchy mapping relationship corresponding to each candidate placement combination in the candidate placement combination list.

[0090] This step is the initial stage of the cross-platform campaign creation process, executed by the ad placement agent. Its core purpose is to resolve the inconsistency in the hierarchical rules of different external advertising platforms. By adapting the hierarchical planning to the native rules of each platform, a unified candidate placement combination is mapped to the corresponding platform's organizational structure, laying a foundation for adaptability in subsequent campaign creation. In this step, the ad placement agent first reads the structured placement task sheet from the shared data space, as well as the candidate placement combination list generated by the audience targeting and strategy agents. It extracts the total budget and channel allocation constraints from the placement task sheet (i.e., the proportion of budget allocated to each advertising platform in the total budget, a hard constraint used to control the scale of investment in different channels). Then, for each external advertising platform, it retrieves the platform's native rules (i.e., the platform's own exclusive rules for its own placement system, including requirements such as placement hierarchy structure, hierarchy capacity limits, and budget configuration rules), and plans a placement plan hierarchy structure that matches the platform's native rules. This means adapting to the platform's organizational structure and building a hierarchical framework from the top-level plan to the bottom-level creatives, while ensuring that the total budget of each platform's hierarchical framework complies with the channel allocation constraints. After completing the hierarchical planning, the agent determines the platform deployment hierarchy mapping relationship for each candidate deployment combination in the candidate deployment combination list. That is, it clarifies the specific position of each candidate combination in the corresponding platform hierarchy framework, such as which top-level plan or which intermediate unit it belongs to, to ensure that the affiliation of all combinations conforms to the platform's hierarchical rules.

[0091] Taking a cross-platform campaign for credit products as an example, the total budget for this campaign is 50,000 yuan. The channel allocation constraints stipulated in the campaign task are 50% for short video platforms, 30% for search advertising platforms, and 20% for information flow platforms. The candidate campaign combination list contains 50 valid combinations, of which 30 are adapted to short video platforms, 15 to search platforms, and 5 to information flow platforms. In this step, the advertising agent plans different levels for different platforms: For short video platforms, the native level rule is a three-level structure of "promotion plan - ad group - creative". The advertising agent plans a top-level promotion plan for it, "Credit Campaign - Short Video Platform - V1.0", with a total budget of 25,000 yuan (meeting the 50% channel allocation). Under the plan, intermediate ad groups are divided according to the audience targeting strategy, such as "urban young and middle-aged customer group ad group" and "small and micro enterprise owner customer group ad group". Then, the 30 candidate campaign combinations adapted to short video platforms are mapped to the creative level under the corresponding customer group ad group. For search advertising platforms, the native hierarchical structure is a four-level hierarchy of "Account - Campaign - Unit - Creative". The ad agent planned a top-level campaign, "Credit Placement - Search Platform - V1.0", with a total budget of 15,000 yuan (compliant with a 30% channel allocation). Intermediate units such as "CPC Bidding Unit" and "OCPC Bidding Unit" were categorized by bidding type, and 15 candidate placement combinations adapted to the search platform were mapped to the creative level under the corresponding unit. For feed platforms, the native hierarchical structure is a two-level hierarchy of "Campaign - Creative". The ad agent planned a corresponding top-level campaign with a total budget of 10,000 yuan, and directly mapped 5 adapted candidate placement combinations to the creative level under the campaign. Through this planning and mapping, all candidate placement combinations were adapted to the native hierarchy of each platform, providing clear guidance for subsequent parameter conversion and campaign creation.

[0092] Step S420: Map each candidate placement combination in the candidate placement combination list to the configuration parameters that can be executed under the corresponding advertising platform placement level, and perform compliance verification, risk control verification, budget allocation verification, frequency control and deduplication verification on each set of mapped configuration parameters in sequence to remove candidate placement combinations that fail the verification.

[0093] This step involves parameter adaptation and pre-verification for cross-platform campaigns. Following the platform-level mapping completed in step S410, it is executed by the ad agent. The core function is to convert the system's unified candidate combination configuration into executable parameters recognizable by the platform. Pre-verification through multi-dimensional checks eliminates non-compliant and invalid combinations, preventing issues such as violations, budget overruns, and traffic waste in subsequent campaigns. In this step, the ad agent first converts the system's unified combination configuration (including creative information, audience targeting, bidding rules, frequency control rules, etc.) into executable configuration parameters at the corresponding ad platform's campaign level for each candidate combination. This translates the system's common fields and identifiers into native fields and parameter values ​​directly recognizable by the corresponding platform's API, achieving platform adaptation. After parameter mapping, the agent performs four progressive checks on each mapped configuration parameter. Any candidate combination failing any check is immediately eliminated: the first check is compliance verification, verifying whether the parameters conform to the ad platform's content specifications and configuration rules, eliminating configurations containing non-compliant content or parameters exceeding the platform's value range. The second step is risk control verification, which verifies whether the configuration complies with business risk control rules and removes configurations targeting high-risk groups or violating internal risk control requirements. The third step is budget allocation verification, which verifies whether the total budget of the corresponding platform still complies with channel allocation constraints after the budget configuration of this combination is added, and removes configurations that would lead to budget overruns. The fourth step is frequency control and deduplication verification, which verifies whether the exposure frequency control rules of the combination meet the requirements, and at the same time checks for duplicate placement configurations and redundant configurations that may cause internal traffic competition, removing combinations that violate regulations or are duplicates.

[0094] Taking a cross-platform advertising scenario for credit products as an example, for a candidate combination adapted to a short video platform, the advertising agent first converts the unified configurations within the system—such as creative copy, audience segment identifiers, OCPC bid of 90 yuan, and a limit of no more than 2 daily impressions per user—into native parameters recognizable by the short video platform's API. These parameters include creative titles, platform-specific audience segment IDs, ocpc_bid bid fields, and frequency_control frequency control fields. Then, these parameters are validated: In compliance validation, it is confirmed that the creative has no illegal keywords and the bid of 90 yuan is within the legal range of 50-200 yuan for the platform's OCPC bid, thus passing the validation; in risk control validation, it is confirmed that the corresponding audience is a high-quality customer group that is neither students nor highly indebted, meeting business risk control requirements, thus passing the validation; in budget allocation validation, it is confirmed that after adding a daily budget of 500 yuan for this combination, the total budget for all combinations on the short video platform is 24,000 yuan, which does not exceed the channel budget limit of 25,000 yuan, thus passing the validation; in frequency control and deduplication validation, it is confirmed that the frequency control rules meet the requirements and there are no duplicate creative-audience targeting strategy combinations, thus passing the validation and the combination is retained. For another group of candidate combinations, compliance verification after parameter mapping revealed that the creative contained the illegal statement "no credit requirements," and it was directly eliminated. Another combination was eliminated because budget verification showed that adding the budget would cause the total budget of the short video platform to exceed the channel limit of 25,000. Yet another combination was eliminated because frequency control verification showed that its configuration of 5 daily exposures per user exceeded the system rule. Through the above pre-verification, it was ensured that all combinations entering the subsequent stages were legal, compliant, and valid configurations.

[0095] Step S430 involves calculating and configuring the initial bid and initial budget for the candidate delivery combinations that have passed all verifications, based on the optimization metrics of the delivery task order and the delivery execution strategy corresponding to the candidate delivery combinations.

[0096] This step involves configuring the launch parameters for the cross-platform campaign. Following step S420, which validated all valid candidate campaign combinations, this step is executed by the advertising agent. The core objective is to configure appropriate initial bid and budget parameters for each valid combination, ensuring cost control and reasonable budget allocation during the campaign launch phase. This prevents issues such as bid deviations from targets and budget imbalances during the launch phase. In this step, the advertising agent first extracts the optimization metrics from the campaign task list (i.e., the core optimization objectives of this campaign, such as lowest lead acquisition cost and highest ROI, which are the core guiding principles for all parameter configurations), as well as the corresponding campaign execution strategy for each candidate campaign combination. Based on the guidance of the optimization metrics, combined with the combination's performance prediction score and the initial bid upper and lower limits agreed upon in the execution strategy, the agent calculates and configures the initial bid for that combination. This initial bid parameter is used for platform traffic bidding during the campaign launch phase. This bid always remains within the upper and lower limits agreed upon in the execution strategy, ensuring the safety of the bidding. Meanwhile, the advertising agent allocates an initial budget to each combination based on the total budget constraint and the performance prediction score of each combination. This initial budget is the upper limit of the combination's budget during the campaign launch phase. The budget is tilted towards high-potential combinations with high performance prediction scores to ensure that high-quality combinations can obtain sufficient traffic support. At the same time, the total budget of all combinations still complies with the constraints of channel allocation and total budget.

[0097] Taking a cross-platform campaign for credit products as an example, the optimization metric for this campaign was "Auto Finance Lead CPA ≤ 80 RMB". After verification in step S420, 45 valid candidate campaign combinations were retained. For one combination adapted to short video platforms and corresponding to the OCPC campaign execution strategy, the initial bid range stipulated in the execution strategy was 80-100 RMB, with an estimated overall performance score of 8.2 (high potential). The advertising agent, considering the cost requirements of the optimization metric, configured an initial bid of 80 RMB, perfectly matching the target lead cost, and allocated an initial daily budget of 600 RMB to ensure traffic supply for this high-potential combination. For another combination targeting small and micro-enterprise owners with a corresponding CPC campaign execution strategy, the bid range stipulated in the execution strategy was 1.0-1.5 RMB, with an estimated score of 7.5. The advertising agent configured an initial bid of 1.2 RMB (between the upper and lower limits) and allocated an initial daily budget of 400 RMB. For the OCPC campaign targeting customers in lower-tier markets, the bid range for execution strategy is 70-90 yuan, with an estimated score of 7.0. The ad agent is configured with an initial bid of 75 yuan and an initial daily budget of 300 yuan. The initial bids for all campaigns did not exceed the execution strategy's upper and lower limits, and the total budget for all campaigns also conformed to the budget allocation constraints of each channel, providing precise launch parameters for subsequent campaign creation.

[0098] Step S440: Call the standardized API interfaces of each advertising platform to create corresponding campaign plans for candidate campaign combinations that have passed verification and completed initial parameter configuration.

[0099] This step is the execution phase of the cross-platform campaign. Following step S430, which completed the initial parameter configuration of valid candidate campaign combinations, this step is executed by the ad agent. The core of this step is to synchronize the system's configuration to the external advertising platform through standardized interface calls, completing the actual creation of the campaign. This automates the execution from strategy configuration to platform implementation, replacing the inefficient traditional method of manually creating campaigns on each platform's backend. In this step, the ad agent calls the standardized API interfaces of each advertising platform. APIs, or Application Programming Interfaces, are interactive interfaces provided by external advertising platforms that allow external systems to remotely call their campaign management functions. The standardized API interfaces are adaptation interfaces that the system uniformly encapsulates the native APIs of each platform, shielding the differences in request formats, parameter naming, and calling logic between different platforms, thus achieving unified cross-platform calls. The advertising agent takes each candidate campaign combination that has passed verification and completed initial parameter configuration and passes the native configuration parameters, initial bid, initial budget, and other information, and passes them into the standardized API interface according to the requirements of the corresponding platform. The platform then completes the actual creation of the campaign plan, that is, it generates executable promotion plans, ad units, creatives and other campaign units in the advertising platform's campaign system, and completes the implementation of the configuration.

[0100] Taking the cross-platform deployment scenario of credit products as an example, after the aforementioned steps, a total of 45 valid candidate deployment combinations are retained, of which 30 are adapted to short video platforms and 15 are adapted to search advertising platforms. In this step, the advertising agent first calls the encapsulated standardized deployment creation API of the short video platform for the 30 combinations. It batches the previously planned top-level promotion plan, customer group ad groups, and the creative parameters, OCPC / CPC bids, daily budgets, and other configurations of the 30 combinations into the interface. After receiving the request, the short video platform completes the creation of the corresponding promotion plan, ad group, and creative in its own deployment system and returns a successful creation result. Subsequently, for the 15 combinations of the search platform, it calls the standardized creation API of the search platform, passing in the promotion plan, bid unit, creative configuration, and other parameters of the search platform to complete the creation of the deployment plan on the search platform side. The entire process does not require manual login to the backend of each platform. Through batch calls of standardized APIs, 45 cross-platform campaigns can be created in just a few minutes, which can greatly improve the efficiency of campaign creation and avoid configuration errors caused by manual operation. After creation, each advertising platform will return its own unique identifier for the campaign unit, providing a data foundation for subsequent identifier mapping.

[0101] Step S450: Obtain the unique identifier of the advertising platform corresponding to the delivery plan returned by each advertising platform, establish a one-to-one mapping relationship between the unique identifier of the advertising platform and the unique identifier of the candidate delivery combination, and write the mapping relationship into the shared data space.

[0102] This step is the final stage of the cross-platform campaign creation process. Following the campaign creation results returned by each advertising platform in step S440, it is executed by the advertising agent. The core of this step is establishing the correspondence between the system's global identifier and the platform's private identifier, resolving the issues of inconsistent data definitions and lack of alignment across platforms. This provides crucial data correlation for subsequent campaign data retrieval, multi-dimensional A / B testing, and performance attribution analysis. In this step, the advertising agent first obtains the unique identifier of each advertising platform returned by each platform. This is the private identifier assigned by each external advertising platform to its own campaign unit (campaign, ad group, creative). This identifier is only valid within its corresponding platform, and the identifiers of different platforms are independent and lack a unified rule. Subsequently, the advertising agent binds this platform's private identifier to the unique identifier of the corresponding candidate campaign combination within the system (i.e., the globally unified identifier previously assigned to each combination, a system-specific code that runs throughout the entire campaign process), establishing a one-to-one mapping relationship between the two. This ensures that each combination within the system can accurately match the campaign unit on the platform side. After establishing the mapping relationship for all combinations, the agent writes all the mapping relationships into the shared data space, realizing the sharing of identification data across agents and stages.

[0103] Taking the cross-platform advertising scenario for credit products as an example, this campaign created a total of 45 effective advertising plans, of which 30 were adapted to short video platforms and 15 were adapted to search advertising platforms. For the 30 combinations on the short video platform, their globally unique system identifiers are AD-TASK-20250325-001-V1.0-001 to AD-TASK-20250325-001-V1.0-030. The corresponding creative unique identifiers returned by the short video platform are 100001 to 100030. The advertising agent establishes a mapping relationship: AD-TASK-20250325-001-V1.0-001→100001, AD-TASK-20250325-001-V1.0-002→100002, and so on to complete the mapping of the 30 groups. For the 15 combinations from the search platform, their globally unique system identifiers range from AD-TASK-20250325-001-V1.0-031 to AD-TASK-20250325-001-V1.0-045. The corresponding creative unique identifiers returned by the search platform range from 200001 to 200015, establishing a one-to-one mapping relationship. All 45 mapping relationships are ultimately written into the shared data space. When subsequently pulling platform campaign data, this mapping can be used to convert the platform's private identifiers into the system's globally unified identifiers, achieving accurate cross-platform data alignment and ensuring the accuracy of A / B testing and performance attribution.

[0104] In this embodiment, the advertising agent first addresses the technical pain point of inconsistent placement levels and interface rules across different advertising platforms by adapting to the hierarchical planning and parameter mapping of the native rules of each platform. Combined with the encapsulated standardized API interface, it enables automated batch creation of cross-platform campaigns, replacing the inefficient traditional method of manually creating campaigns one by one on each platform's backend. This reduces the manual operation that would have taken hours to minutes, significantly improving campaign execution efficiency and avoiding configuration errors caused by manual operation. Secondly, through a four-tiered progressive verification process involving compliance, risk control, budget, and frequency control, invalid combinations that violate regulations, exceed budgets, or are redundant are eliminated before campaign creation. This mitigates the risks of platform rejection, regulatory penalties, budget overruns, and internal traffic consumption during subsequent campaigns, ensuring compliance and controllability. Simultaneously, based on the initial bid and budget configuration according to the core optimization metrics and predicted combination performance scores, traffic and budget resources are allocated to high-potential, high-quality combinations. This ensures controllable costs and reasonable budget allocation during the campaign launch phase, avoiding issues such as bid deviations from targets and budget imbalances during the launch phase. Finally, by establishing a globally unified identifier mapping relationship, the core issues of inconsistent data standards and inability to correlate and align data across platforms can be resolved. This provides accurate data correlation basis for subsequent real-time data retrieval, multi-dimensional A / B testing, and multi-level effect attribution, achieving standardized and automated connection of the entire process from strategy design to platform implementation of the campaign mix, laying a solid foundation for subsequent campaign performance optimization.

[0105] In step S160, the advertising agent pulls and processes real-time advertising data from various advertising platforms based on the advertising task order. Using the benchmark advertising combination as a reference, it conducts multi-dimensional A / B testing on each candidate advertising combination in the candidate advertising combination list, automatically adjusts the bid and budget within preset constraints, and writes the parameter adjustment records into the shared data space.

[0106] This step involves real-time data processing and dynamic optimization during the online campaign phase. It is executed by the advertising agent, which, in conjunction with the monitoring and evaluation layer, first periodically retrieves real-time campaign data from various advertising platforms by calling their standardized API interfaces, according to the data granularity agreed upon in the campaign task. Then, it performs standardized processing on the acquired cross-platform raw data, including time zone alignment, data deduplication, transmission latency correction, and outlier filtering. This integrates the scattered data from different platforms into a unified statistical dataset. Next, using the previously determined benchmark campaign combination as a unified reference, and based on this effective dataset, it optimizes the campaign from the perspectives of creative content, audience targeting strategy, and campaign execution strategy. From both the strategic and advertising platform perspectives, parallel A / B testing is conducted on each candidate campaign combination in the candidate campaign combination list. Based on the statistical results of the A / B testing, and in accordance with the core optimization metrics agreed upon in the campaign task sheet, the bids and budgets of each candidate campaign combination are automatically iterated and adjusted under preset constraints such as the upper limit of single parameter adjustment range and the upper limit of parameter adjustment frequency per unit time. Traffic and budget resources are tilted towards combinations with better performance. Finally, the entire process of parameter adjustment operations, the values ​​before and after parameter changes, and the basis for adjustment are compiled into a standardized parameter adjustment record, which is then completely written into the shared data space to provide real-time process data support for subsequent performance attribution and strategy iteration.

[0107] Reference Figure 5 , Figure 5 The flowchart of the steps provided in one embodiment of this application is as follows: Based on the delivery task order, the advertising agent pulls and processes the real-time delivery data of each advertising platform in a unified manner, and performs multi-dimensional A / B testing on each candidate delivery combination in the candidate delivery combination list with the benchmark delivery combination as a reference. Within the preset constraints, the agent automatically adjusts the bid and budget and writes the parameter adjustment record into the shared data space. The flowchart includes, but is not limited to, steps S510 to S540.

[0108] In step S510, the advertising agent, together with the monitoring and evaluation layer, retrieves real-time delivery data from various advertising platforms according to the data granularity agreed upon in the delivery task order. The real-time delivery data is then subjected to standardized processing, including time zone alignment, data deduplication, transmission delay correction, and outlier filtering, to obtain a valid delivery dataset with unified statistical standards.

[0109] This step marks the beginning of real-time data processing during the online campaign phase. It is jointly executed by the advertising agent and the monitoring and evaluation layer. Its core purpose is to address the issues of inconsistent data standards and varying data quality across advertising platforms. Through standardization, it integrates fragmented raw data from various platforms into a clean, unified, and effective dataset, providing a reliable data foundation for subsequent multi-dimensional A / B testing. In this step, the advertising agent first periodically calls the standardized API interfaces of each advertising platform to retrieve real-time raw data according to the data granularity agreed upon in the campaign task (i.e., the time interval for retrieving campaign data, used to control the frequency of data updates, such as hourly or daily). Then, it performs four standardization operations on the acquired cross-platform raw data: The first is time zone alignment, converting the reporting times of different platforms to the system's standard time zone to resolve inconsistencies in time zone standards. The second is data deduplication, removing duplicate data caused by interface retries and repeated platform reporting to avoid duplicate statistics. The third step is transmission delay correction. Addressing the industry-wide issue of delayed conversion data reporting on advertising platforms, this step supplements the delayed conversion data into the corresponding historical time window, correcting data discrepancies caused by the delay. The fourth step is outlier filtering. This removes extreme anomalies in metrics such as impressions, clicks, and costs that significantly deviate from the normal range, preventing anomalies from interfering with subsequent statistical analysis. After all processing is complete, a valid dataset with standardized statistical criteria is obtained.

[0110] Taking a cross-platform credit product deployment scenario as an example, the agreed data granularity for this deployment is hourly. The intelligent agent pulls real-time deployment data from three platforms—short video, search, and news feed—every hour. First, the reporting time in UTC time zone of the short video platform is uniformly converted to East Eighth Zone standard time to complete time zone alignment. Then, two duplicate data entries caused by retries during the retrieval process are removed. Next, three lead conversion data entries from the previous hour that were delayed in reporting are added to the dataset of the corresponding time window to correct transmission delays. Finally, one test data entry with an abnormally sudden increase in exposure is filtered out, resulting in a unified hourly effective deployment dataset. All data has completely consistent time and metric standards, and can be directly used for subsequent A / B testing analysis.

[0111] Step S520: Using the benchmark campaign combination as a unified benchmark, and based on the effective campaign dataset, conduct parallel comparative A / B testing on each candidate campaign combination from the dimensions of creative content, audience targeting strategy, campaign execution strategy, and advertising platform.

[0112] This step is the core of the online campaign performance comparison and analysis. Continuing from the effective campaign dataset generated in step S510, it is executed jointly by the advertising agent and the monitoring and evaluation layer. The core objective is to simultaneously verify the campaign performance of different variables through multi-dimensional parallel comparative experiments. This solves the problems of low efficiency and inability to simultaneously decompose the effects of multiple variables in traditional single-dimensional A / B testing, providing accurate experimental basis for subsequent bid budget adjustments. In this step, the advertising agent first uses the previously determined benchmark campaign combination (i.e., a unified benchmark throughout the entire process, used for effect reference in all experiments, ensuring the comparability of all test results) as a unified control. Based on the standardized effective campaign dataset, it conducts parallel comparative A / B testing (i.e., conducting comparative experiments on multiple sets of variables simultaneously, without testing them one by one, significantly improving experimental efficiency). Testing is carried out simultaneously from four core dimensions: The first is the creative dimension, breaking down the campaign performance of different creatives and comparing the click-through rate, conversion rate, customer acquisition cost, and other indicators of each creative with the benchmark creative to verify the differences in the effectiveness of creative variables. The second dimension is audience targeting strategy, breaking down the performance of different audience targeting strategies, comparing the performance of each audience group with the benchmark audience, and verifying the effectiveness of audience variables. The third dimension is campaign execution strategy, breaking down the performance of different bidding, frequency control, and other execution strategies, comparing the performance of each execution strategy with the benchmark strategy, and verifying the effectiveness of execution strategy variables. The fourth dimension is advertising platform, breaking down the performance of different advertising platforms, comparing the performance of each platform with the benchmark platform, and verifying the effectiveness of platform variables. All dimensions of testing are completed simultaneously and in parallel, with all experimental results output at once.

[0113] Taking a cross-platform campaign for credit products as an example, the benchmark campaign combination had a lead CPA of 90 yuan and a ROI of 2.5. The effective dataset contained hourly campaign data for 45 candidate combinations. In this step, the advertising agent simultaneously conducted tests across four dimensions, using the benchmark combination as a control: In the creative dimension, the lead CPA for low-interest selling point creatives was verified to be 75 yuan, significantly better than the benchmark creative; in the audience targeting strategy dimension, the lead CPA for urban young and middle-aged customers was verified to be 72 yuan, outperforming the benchmark audience; in the campaign execution strategy dimension, the lead CPA for the OCPC bidding strategy was verified to be 78 yuan, outperforming the benchmark CPC strategy; and in the platform dimension, the lead CPA for short video platforms was verified to be 76 yuan, outperforming search and news feed platforms. All test results were output simultaneously, eliminating the need for separate experiments for each dimension, significantly improving the efficiency of performance analysis and providing comprehensive experimental basis for subsequent parameter adjustments.

[0114] Step S530: Based on the statistical results of A / B testing, and in accordance with the optimization indicators agreed upon in the campaign task order, under preset constraints, automatically iterate and adjust the bids and budgets of each candidate campaign combination. The preset constraints include at least the upper limit of the single parameter adjustment range and the upper limit of the parameter adjustment frequency per unit time.

[0115] This step is the real-time optimization execution phase of the online campaign. Following the multi-dimensional A / B test statistics output in step S520, it is executed by the advertising agent. The core is to automatically iteratively adjust the bids and budgets of each candidate campaign combination according to the optimization goals agreed upon in the campaign task list (i.e., the agent automatically and cyclically adjusts campaign parameters based on real-time performance data, continuously optimizing campaign performance through an automated optimization mechanism). Simultaneously, preset constraints are used to avoid campaign fluctuations caused by excessive parameter adjustments, achieving dynamic optimization of campaign performance. In this step, the advertising agent first uses the core optimization indicators agreed upon in the campaign task list as adjustment targets, combined with the real-time performance data of each candidate combination, to perform differentiated parameter adjustments for combinations with different performance: For combinations with performance exceeding the optimization goals and high potential, bids are appropriately increased to obtain more exposure opportunities, while increasing their budget share to amplify the acquisition of high-quality traffic. For combinations with performance below the optimization goals and low consumption efficiency, bids are appropriately reduced to control customer acquisition costs, while compressing their budget share to avoid ineffective budget consumption. For combinations with performance severely below targets and far exceeding optimization goals, their campaigns are directly suspended, and the budget is released to other high-quality combinations. To ensure campaign stability and avoid campaign fluctuations caused by excessive parameter tuning, this step sets preset constraints to limit the boundaries of parameter tuning. The core constraints fall into two categories: The first is the upper limit of a single parameter adjustment, i.e., the maximum percentage change in parameters when adjusting bids or budgets in a single instance. For example, in this solution, the bid adjustment can be set to no more than 20% and the budget adjustment to no more than 30% in a single instance, preventing drastic fluctuations in impressions and consumption rates caused by sudden parameter changes. The second is the upper limit of parameter tuning frequency per unit of time, i.e., the maximum number of parameter adjustments that can be performed on the same combination per unit of time. For example, in this solution, a single combination can be adjusted a maximum of once per hour to prevent excessively frequent parameter tuning, which could lead to repeated adjustments before the platform data is updated in time, causing parameter fluctuations and ensuring stable convergence of campaign performance.

[0116] Taking a cross-platform campaign for credit products as an example, the core optimization metric for this campaign is a CPA of ≤ 80 yuan for auto finance leads. Preset constraints include a single bid adjustment ≤ 20%, a single budget adjustment ≤ 30%, and a maximum of one adjustment per ad combination per hour. For different candidate ad combinations, the ad agent can perform the following differentiated adjustments: For the high-quality combination of low-interest creative content, targeting urban young and middle-aged demographics, and OCPC bidding, the current actual lead CPA is 72 yuan, significantly better than the optimization target of 80 yuan, and the ROI reaches 3.2. The agent adjusts its initial bid from 78 yuan to 85 yuan (an adjustment of 9%, not exceeding the 20% limit) and its daily budget from 600 yuan to 750 yuan (an adjustment of 25%, not exceeding the 30% limit). Moreover, the interval since the last adjustment is 1.2 hours, which meets the frequency constraint. By adjusting and allocating more budget to this high-quality combination, its customer acquisition capabilities can be amplified.

[0117] For the inefficient combination of lower-tier market demographics and CPC bidding, the current actual lead CPA is 92 yuan, exceeding the optimization target of 80 yuan. The agent adjusts the bid from 1.5 yuan to 1.2 yuan (an adjustment of 20%, reaching the single adjustment limit) and adjusts the daily budget from 400 yuan to 300 yuan (an adjustment of 25%, not exceeding the limit). By compressing the consumption of this combination, the waste of ineffective budget is controlled.

[0118] For a particular portfolio that is significantly underperforming, with an actual CPA of 120 yuan, far exceeding the optimization target, the agent directly adjusts its daily budget to 0, suspends the campaign for that portfolio, and reallocates the freed-up budget to other high-quality portfolios.

[0119] This step replaces traditional manual parameter tuning with a constrained automated parameter tuning mechanism, which can significantly improve the response speed of campaign optimization. At the same time, by avoiding parameter tuning risks through constraints, it can ensure the stability of the campaign, effectively improve the overall budget utilization efficiency, and promote the overall campaign effect to converge quickly towards the optimization goal.

[0120] Step S540: Generate a parameter tuning record by recording the parameter tuning operations, the values ​​before and after parameter changes, and the basis for adjustment throughout the entire process, and write the parameter tuning record into the shared data space.

[0121] This step is the final stage of the parameter tuning process in the online campaign phase, executed by the advertising agent. Its core function is to structurally retain all information from the parameter adjustment process and synchronize it to a shared data space, providing comprehensive historical data support for subsequent performance attribution, knowledge base updates, and issue tracing. In this step, the advertising agent first structures all parameter tuning operations from step S530, generating standardized tuning records. These records fully cover three core types of information: The first type is the tuning operation itself, i.e., the specific operations performed on the candidate combination, such as bid adjustment, budget adjustment, and campaign pause. The second type is the numerical values ​​before and after the parameter change, i.e., the original values ​​of parameters such as bid and budget before the adjustment, and the new parameter values ​​after the adjustment, completely recording the parameter change process. The third type is the adjustment basis, i.e., the performance data and decision logic that triggered this parameter adjustment, such as the combination's real-time customer acquisition cost, ROI, and other performance indicators, as well as the corresponding optimization decision reasons. After the parameter tuning records are generated, the agents write all the structured parameter tuning records into the shared data space (i.e., the unified data interaction carrier of the multi-agent collaborative system, in which all agents can read and write the data, and realize cross-agent information exchange), ensuring that all agents can synchronously obtain the historical information of parameter tuning.

[0122] By writing parameter tuning records into a shared data space, end-to-end traceability of parameter adjustments can be achieved, solving the problems of traditional manual parameter tuning lacking complete traceability and subsequent review failing to trace the reasons for decisions. This provides complete historical data for subsequent multi-level effect attribution and intelligent agent knowledge base updates, while also improving the interpretability and maintainability of the entire deployment system and ensuring information synchronization and transparency during multi-agent collaboration.

[0123] In this embodiment, through the collaborative efforts of the advertising agent and the monitoring and evaluation layer, firstly, standardized data processing integrates fragmented raw data from across platforms into a unified statistical dataset, ensuring the reliability of subsequent analysis data from the outset. Secondly, multi-dimensional parallel A / B testing allows for the simultaneous verification of effects across four dimensions: creative, audience targeting strategy, campaign execution strategy, and platform performance. This eliminates the need for independent experiments in each dimension, significantly improving the efficiency of performance analysis and enabling precise decomposition of multi-variable effects, avoiding the omission of high-potential optimization directions. Furthermore, a constrained automated iterative parameter tuning mechanism replaces traditional manual parameter tuning, dynamically tilting the budget towards high-performance combinations while ensuring campaign stability. This significantly improves budget utilization efficiency and drives the overall campaign performance towards the optimization goal. Finally, structured retention and sharing of full-process parameter tuning records enable end-to-end traceability of parameter adjustments, providing complete historical data for subsequent performance attribution and agent knowledge base updates. This also enhances the system's interpretability and maintainability, ensuring information synchronization and transparency during multi-agent collaboration.

[0124] In step S170, the monitoring and evaluation layer collects relevant data from various advertising platforms based on the delivery task order, and conducts multi-level effect attribution analysis on the delivery-related data according to the creative dimension, audience targeting strategy dimension, delivery execution strategy dimension and advertising platform dimension, generates attribution results and writes them into the shared data space.

[0125] This step, the global performance review and attribution phase within the campaign period, is executed by the system's monitoring and evaluation layer. Its core is to accurately break down the performance contribution of different campaign variables through the aggregation of full data and multi-level attribution, solving the problem that traditional overall attribution cannot pinpoint specific optimization variables. This provides accurate performance data for subsequent updates and iterative optimization of the intelligent agent knowledge base. In this step, the monitoring and evaluation layer first aggregates full campaign logs and conversion feedback data from various advertising platforms, as well as back-end quality data from corresponding business systems (such as real demand verification data for leads in credit scenarios), according to the evaluation window (the time range used to evaluate the overall performance of this campaign) and attribution window (the time range used to statistically analyze user conversion touchpoints, covering industry characteristics of conversion delays), according to the campaign task agreement. This forms the raw attribution dataset. Subsequently, the raw dataset undergoes preprocessing, including user identification unification, structured construction of campaign touchpoints, effective touchpoint extraction, and deduplication, eliminating invalid and redundant data to obtain a standardized attribution dataset with unified statistical standards. Furthermore, based on a standardized dataset, and using candidate ad combinations as the smallest statistical unit, multi-level performance attribution analysis is conducted from the dimensions of creative content, audience targeting strategy, ad execution strategy, and advertising platform. This involves breaking down the contribution of ad variables from different dimensions to the final conversion effect layer by layer, generating structured attribution results that include unique identifiers of candidate combinations, weighted conversion numbers, conversion value, return on investment, customer acquisition cost, and performance confidence. Finally, the complete attribution results are written into a shared data space and synchronized to all agents to provide data support for subsequent optimization iterations.

[0126] Reference Figure 6 , Figure 6 The flowchart of the monitoring and evaluation layer provided in one embodiment of this application is as follows: it collects relevant data from various advertising platforms based on the delivery task order, performs multi-level effect attribution analysis on the relevant data according to the creative dimension, audience targeting strategy dimension, delivery execution strategy dimension and advertising platform dimension, generates attribution results and writes them into the shared data space, including but not limited to steps S610 to S640.

[0127] In step S610, the monitoring and evaluation layer, according to the evaluation window and attribution window agreed upon in the delivery task order, gathers the full delivery logs and conversion feedback data of each advertising platform, as well as the back-link quality data of the corresponding business system, to form the original attribution dataset.

[0128] This step is the preliminary data preparation stage for multi-level performance attribution analysis, executed by the monitoring and evaluation layer. Its core is to unify and aggregate end-to-end campaign data scattered across various advertising platforms and business systems, providing a complete and comprehensive raw data source for subsequent attribution analysis. This addresses the attribution bias caused by traditional attribution methods relying solely on platform-side data and lacking post-campaign quality data. In this step, the monitoring and evaluation layer first reads the evaluation and attribution windows agreed upon in the campaign task order. Based on these two time windows, it aggregates three types of core data: The first type is the full campaign log, i.e., the raw campaign behavior data reported by each advertising platform, including full campaign behavior records such as impressions, clicks, display time, and user identifiers. The second type is conversion feedback data, i.e., conversion event data reported by each advertising platform, including platform-side conversion records such as user conversion behavior, conversion time, and user identifiers. The third type is post-campaign quality data, i.e., the quality verification data of the post-conversion process stored in the business system, used to supplement the real business effects that platform-side conversion data cannot cover. For example, in a credit scenario, data such as the verification of the actual demand for leads and the final loan disbursement result are used to distinguish between effective and ineffective conversions. After extracting the three types of data, all data are integrated into a unified attribution raw dataset, which serves as the basis for subsequent attribution preprocessing and analysis.

[0129] Taking the cross-platform advertising scenario of credit products as an example, the evaluation window for this advertising task was set at 7 days and the attribution window at 30 days. Based on this time range, the monitoring and evaluation layer pulled the full advertising logs from three advertising platforms: short video, search, and information flow. This included all exposure and click behavior records within 7 days. At the same time, it pulled the lead conversion feedback data from the three platforms within 30 days. In addition, it also pulled the post-link quality data of the corresponding leads from the credit business system, including information such as sales follow-up results, verification of users' real needs, and final loan disbursement status. All these scattered data were integrated into a unified attribution raw dataset, which can provide complete end-to-end data support for subsequent attribution analysis.

[0130] Step S620 involves sequentially performing preprocessing operations on the original attribution dataset, including user identification unification, delivery touchpoint structure construction, effective touchpoint extraction, and deduplication, to remove invalid and redundant data and obtain a standardized attribution dataset with unified statistical standards.

[0131] This step is the data cleaning and standardization stage for attribution analysis. It follows the raw attribution dataset generated in step S610 and is executed by the monitoring and evaluation layer. Its core purpose is to address the issues of inconsistent data definitions across platforms, fragmented user behavior, and data redundancy. It processes the scattered raw data into a standardized dataset usable for attribution analysis, providing a clean and unified data foundation for subsequent multi-level attribution. In this step, the monitoring and evaluation layer performs four preprocessing operations on the raw attribution dataset: The first is user identifier unification. Addressing the issue of inconsistent user identifier formats across different advertising platforms, it maps the private user identifiers of each platform (such as device ID, encrypted phone number, IMEI, etc.) to a globally unified unique user identifier, connecting user behavior data across platforms and enabling the association of the same user's advertising behavior on different platforms, thus resolving the problem of fragmented user identities across platforms. The second is the structured construction of advertising touchpoints. For each unified user, all their advertising touchpoints (i.e., user interactions with the advertisement, such as impressions, clicks, etc.) are sorted chronologically and organized into a structured user touchpoint sequence, clearly reconstructing the entire user behavior path from ad exposure to final conversion. The third step is effective touchpoint extraction, which involves capturing effective touchpoints within the attribution window specified in the campaign task agreement and removing expired touchpoints that fall outside the attribution window to avoid interference with the attribution results. The fourth step is touchpoint deduplication, which removes duplicate touchpoints from the same user on the same ad (e.g., multiple clicks on the same ad within a short period) to avoid interference with attribution weight calculations and eliminate redundant data. After completing all preprocessing operations, a standardized attribution dataset is obtained—a clean dataset with consistent data definitions, unified user identities, and structured touchpoints, which can be directly used for subsequent multi-level attribution analysis.

[0132] Taking the cross-platform advertising scenario of credit products as an example, for the previously aggregated raw dataset, the device ID of the short video platform, the encrypted mobile phone number of the search platform, and the IMEI of the information flow platform were first uniformly mapped to the system's global user ID. For example, the different identifiers of user A on the three platforms were uniformly mapped to the global user ID U10086, thus connecting the user's cross-platform behavior. Then, all the user's advertising touchpoints were sorted by time and organized into a structured touchpoint sequence: 2025-03-20 14:23 low-interest creative exposure on the short video platform, 2025-03-22 10:15 business loan creative click on the search platform, and 2025-03-25 09:47 lead conversion. Next, the expired touchpoints of the user on 2025-02-18 (exceeding the 30-day attribution window) were removed, and the user's two duplicate clicks on the same short video ad on March 20 were also removed. Finally, standardized attribution data was obtained, and all user behaviors were connected and organized, which can be directly used for subsequent attribution analysis.

[0133] Step S630: Based on the standardized attribution dataset, using candidate ad combinations as the smallest statistical unit, conduct multi-level effect attribution analysis from the dimensions of creative content, audience targeting strategy, ad execution strategy, and advertising platform to generate structured attribution results. The attribution results include at least the unique identifier of the candidate ad combination, weighted conversion number, conversion value, return on investment, customer acquisition cost, and effect confidence level.

[0134] This step is the core execution phase of attribution analysis. Continuing from the standardized attribution dataset generated in step S620, it is executed by the monitoring and evaluation layer. Its core function is to precisely quantify the contribution of each campaign variable to conversion performance through multi-dimensional, hierarchical decomposition. This addresses the problem of traditional attribution methods, which only provide overall results and cannot pinpoint specific optimization variables, thus providing accurate performance data for subsequent agent optimization. In this step, the monitoring and evaluation layer first uses candidate campaign combinations as the smallest statistical unit. Each independent candidate campaign combination is used as the smallest granularity for performance statistics because each candidate combination corresponds to a specific set of creatives, audience targeting strategies, campaign execution strategies, and platform configurations, precisely corresponding to our experimental variables. Based on this, multi-level performance attribution analysis is conducted from four core dimensions, decomposing the contribution of different dimensions of campaign variables to the final conversion performance layer by layer: The first dimension is the creative dimension, quantifying the contribution of different creatives to conversion and statistically analyzing the conversion performance of each creative. The second dimension is the audience targeting strategy dimension, quantifying the contribution of different audience targeting strategies to conversion. The third dimension is the campaign execution strategy dimension, quantifying the contribution of different bidding, frequency control, and other execution strategies to conversion. The fourth dimension is the advertising platform dimension, quantifying the contribution of different advertising platforms to conversions. Through multi-level decomposition, not only can the overall effect of each combination be obtained, but also the individual effect of each dimension variable can be broken down, accurately identifying the optimization direction. After the analysis is completed, a structured attribution result is generated, which contains at least six core types of information: The first type is a unique identifier for the candidate ad combination, used to associate the corresponding ad configuration. The second type is the weighted conversion count, which is the number of effective conversions after considering the contribution weight of touchpoints, which is different from simple conversion counting and more accurately reflects the true conversion contribution. The third type is the conversion value, which is the value of the corresponding post-conversion business, such as the expected loan value corresponding to a valid lead in a credit scenario. The fourth type is the return on investment, which is the input-output ratio, reflecting the profitability of the combination. The fifth type is the customer acquisition cost, which is the unit conversion cost of the corresponding combination. The sixth type is the effect confidence, which is the statistical confidence of the attribution result, used to identify the stability and reliability of the result and avoid the bias caused by small sample data.

[0135] Taking cross-platform ad placement for credit products as an example, for a standardized attribution dataset, using 45 candidate ad combinations as the smallest statistical unit, four-dimensional attribution was conducted. The results showed: 186 weighted conversions for low-interest creative ads, a lead CPA of 72 yuan, and a 98% confidence level. 172 weighted conversions for urban young adults, a lead CPA of 70 yuan, and a 97% confidence level. 158 weighted conversions for the OCPC execution strategy, a lead CPA of 75 yuan, and a 96% confidence level. 162 weighted conversions for short video platforms, a lead CPA of 73 yuan, and a 97% confidence level. For a single combination, such as low-interest creative content + urban young and middle-aged demographics + OCPC strategy + short video platform, the attribution results are as follows: unique identifier T20250325001-V1-001, weighted conversion count 24, conversion value 480,000 yuan, return on investment 3.2%, customer acquisition cost 72 yuan, and performance confidence level 95%. All results are structured and can provide accurate quantitative basis for subsequent optimization.

[0136] Step S640: Write the attribution results into the shared data space, and generate targeted delivery optimization suggestions based on the attribution results.

[0137] This step is the final stage of the multi-level performance attribution process. Following the structured attribution results generated in step S630, it is executed by the monitoring and evaluation layer. Its core function is to synchronize the attribution results to all agents and generate targeted optimization directions based on these results, providing direct decision-making references for subsequent iterative optimization. In this step, the monitoring and evaluation layer first writes the complete structured attribution results into the shared data space, ensuring that the creative production agent, the audience targeting and strategy agent, the ad placement agent, and the operations manager agent can all synchronously obtain the global performance data for this round of campaigns, eliminating the need for separate data transmission across agents and eliminating information gaps. After the results synchronization is complete, the monitoring and evaluation layer generates targeted campaign optimization suggestions based on the multi-dimensional performance data in the attribution results. Specifically, it outputs specific optimization directions based on the different business scenarios of each agent and the performance differences across various dimensions, helping each agent quickly locate optimization points without requiring agents to extract optimization directions from massive amounts of attribution data, significantly improving the efficiency of subsequent optimization.

[0138] Taking a cross-platform campaign for credit products as an example, the monitoring and evaluation layer first writes the complete attribution results, including 45 sets of combined effects, into a shared data space, which can be read synchronously by all agents. Then, based on the attribution results, targeted optimization suggestions are generated: For the creative production agent, it is recommended to prioritize generating creatives with low-interest selling points and reduce the output of long-cycle creatives for business loans with poor performance. For the audience targeting and strategy agent, it is recommended to focus on expanding the target audience of similar demographics among urban young adults and appropriately reduce the proportion of advertising to lower-tier markets. For the advertising agent, it is recommended to prioritize the OCPC bidding strategy and gradually replace the poor-performing CPC bidding strategy. For the operations manager agent, it is recommended to increase the budget allocation for short video platforms from 50% to 55% and decrease the budget allocation for information flow platforms from 20% to 15%, tilting the budget towards high-performance platforms. All suggestions are based on the attribution results of this round, providing direct decision-making references for the next round of campaign optimization.

[0139] This application's embodiments effectively address industry pain points in traditional cross-platform advertising attribution through a multi-level, end-to-end performance attribution process. These pain points include a single data source, fragmented user identities across platforms, the inability to pinpoint optimization variables while only outputting overall performance, and information asynchrony between agents. Specifically, through end-to-end processing at the monitoring and evaluation layer, it first unifies and integrates the campaign logs, conversion data, and post-processing quality data from various advertising platforms and business systems by aggregating all data. This supplements the missing real-world performance data from traditional attribution, resolving attribution bias caused by relying solely on platform-side data. Furthermore, through standardized data preprocessing, it unifies user identities across platforms, organizing scattered user touchpoints into a structured end-to-end behavioral sequence and eliminating invalid and redundant data. This resolves the problem of fragmented user behavior across platforms, providing clean, unified, and standardized data for attribution analysis. Building upon this foundation, multi-level attribution analysis, using candidate campaign combinations as the smallest unit, decomposes the conversion contribution of each variable layer by layer across four dimensions: creative content, audience targeting strategy, campaign execution strategy, and platform. This not only outputs accurate overall results but also pinpoints specific optimization variables. Furthermore, by filtering out biases from small sample data through performance confidence scores, the accuracy and reliability of the attribution results are significantly improved. Finally, by sharing and synchronizing attribution results and generating targeted optimization suggestions, information gaps between multiple agents can be eliminated, helping each agent quickly identify optimization directions without having to extract optimization logic from massive amounts of data, thus greatly improving the efficiency of subsequent iterative optimization.

[0140] In step S180, each agent updates its knowledge base based on the attribution results, updates the candidate deployment combination with the best performance in this deployment cycle as the benchmark deployment combination, and then returns to step S120.

[0141] This step is the final iteration and cycle start-up phase of the multi-agent collaborative campaign. It is executed collaboratively by four agents: Creative Production, Audience Targeting and Strategy, Ad Targeting, and Operations Manager. The core objective is to accumulate the performance experience of this round of campaigns into the system, update the benchmark for iteration, and initiate the next round of campaign optimization, achieving continuous self-optimization of campaign performance. In this step, each agent first reads the attribution results of this round from the shared data space, filters out stable and valid conclusions that meet the minimum sample size and performance confidence standards, and then updates its respective business domain's dedicated knowledge base (i.e., each agent's dedicated data set storing business rules and historical performance experience to guide subsequent campaign decisions). Specifically, the Creative Production agent updates the performance and compliance rules of creative materials, the Audience Targeting and Strategy agent updates the performance rules of audience targeting and execution strategies, the Ad Targeting agent updates the cross-platform participation and risk control rules, and the Operations Manager agent updates the campaign task management and performance evaluation rules. All updated knowledge bases are synchronized to the shared data space, achieving global accumulation of experience. After the knowledge base is updated, the candidate campaign combination with the best performance and statistical significance among all candidate combinations in this round is updated as the new benchmark campaign combination, replacing the original benchmark. This new benchmark will serve as the unified reference benchmark for the next round of A / B testing. After completion, the system will automatically return to step S120, where the operations manager agent will start the task breakdown and optimization for the next campaign cycle based on the updated knowledge base and the new benchmark, thus achieving continuous iteration of campaign performance.

[0142] Reference Figure 7 , Figure 7 This is a flowchart of the steps provided in one embodiment of the present application for each agent to update the knowledge base based on the attribution results and update the candidate deployment combination with the best performance in the current deployment cycle as the benchmark deployment combination, including but not limited to steps S710 to S740.

[0143] In step S710, each agent obtains the attribution results of the current deployment cycle from the shared data space and selects stable and valid conclusions that simultaneously meet the preset minimum sample size threshold and statistical confidence threshold.

[0144] This step is a preliminary screening process for knowledge base updates. It is executed collaboratively by four agents: Creative Production Agent, Audience and Strategy Agent, Ad Caster Agent, and Operations Manager Agent. The core objective is to filter reliable and stable valid conclusions from a massive amount of attribution results, avoiding interference from small-sample, low-confidence noise data and ensuring the accuracy of accumulated experience. In this step, each agent first reads the complete attribution results for the current campaign cycle from the shared data space. Then, it filters all attribution conclusions according to two preset thresholds: The first is the minimum sample size threshold, which is the minimum number of samples required to obtain a statistically stable conclusion. In this solution, this can be set to a weighted conversion count of no less than 10 to filter conclusions with too small a sample size, as results with such a small sample size are likely due to random fluctuations and lack general applicability. The second is the statistical confidence threshold, which is the minimum requirement for the statistical confidence of the attribution results. In this solution, this can be set to no less than 90% to filter statistically insignificant conclusions and ensure their reliability. After screening, stable and valid conclusions are obtained, which are reliable results that meet the sample size and confidence requirements and can be used to update the knowledge base. These conclusions will serve as the basis for subsequent knowledge base updates.

[0145] Taking the cross-platform deployment scenario of credit products as an example, each agent reads the attribution results of 45 candidate combinations in this round from the shared space. The results are then filtered according to the preset minimum sample size of 10 weighted conversions and a confidence level of 90%. Among them, 8 combinations have fewer than 10 weighted conversions and a confidence level of less than 90%. These conclusions are judged as noise data and are removed. The conclusions of the remaining 37 combinations meet the threshold requirements. These include core conclusions such as low-interest creative (186 weighted conversions, 98% confidence level), urban young and middle-aged population (172 weighted conversions, 97% confidence level), OCPC execution strategy (158 weighted conversions, 96% confidence level), and short video platform (162 weighted conversions, 97% confidence level). These stable and effective conclusions will serve as the basis for subsequent knowledge base updates.

[0146] In step S720, each intelligent agent updates its dedicated knowledge base for its corresponding business area based on the stable and effective conclusions after screening. Specifically, the creative production intelligent agent updates the knowledge base for creative material effects and compliance rules, the audience targeting and strategy intelligent agent updates the knowledge base for audience targeting strategies and campaign execution strategies, the advertising agent updates the knowledge base for cross-platform campaign coordination and risk control rules, and the operations manager intelligent agent updates the knowledge base for campaign task management and effect evaluation rules. All updated knowledge base content is synchronized to the shared data space.

[0147] This step is the core execution phase for experience accumulation. Following the stable and effective conclusions selected in step S710, four agents update their knowledge bases for their respective business domains. The core objective is to accumulate the reliable performance experience from this round of campaigns into the system, guiding subsequent campaign decisions and enabling the system's self-learning and self-optimization. In this step, each agent updates its own business domain-specific knowledge base based on the selected stable and effective conclusions (i.e., each agent stores a dedicated dataset of historical performance experience and business rules for its own business scenario, used to guide subsequent campaign decisions; different agents' knowledge bases focus on different business domains, ensuring the accuracy of decisions). Specifically, the creative production agent updates the knowledge base for creative material effects and compliance rules, updating the knowledge base with the creative effect conclusions and compliance verification rules from this round to guide subsequent creative generation and selection. The audience targeting and strategy agent updates the knowledge base for audience targeting strategies and campaign execution strategies, updating the knowledge base with the conclusions of audience performance and execution strategy performance from this round to guide the generation of subsequent audience strategies and execution strategies. The advertising agent updates its cross-platform campaign tuning and risk control rule knowledge base, incorporating the current round's cross-platform tuning results and risk control verification rules to guide subsequent campaign creation and automatic tuning. The operations manager agent updates its campaign task management and performance evaluation rule knowledge base, updating the current round's task management and performance evaluation conclusions to guide subsequent campaign planning. After completing their respective knowledge base updates, all updated knowledge base content is synchronized to the shared data space, ensuring all agents have access to the latest rules and experience, achieving global sharing of expertise.

[0148] Taking cross-platform advertising of credit products as an example, each agent updates its knowledge base based on stable conclusions after screening: The creative creation agent updates the conclusion that "the CPA of leads from low-interest selling point creatives is 72 yuan, which is significantly better than other creatives" to the creative knowledge base, and will prioritize generating low-interest related creatives in the future. The target audience and strategy agent updates the conclusion that "the CPA of leads from urban young people is 70 yuan, and the CPA of leads from OCPC bidding strategy is 75 yuan, which is more effective" to the strategy knowledge base, and will prioritize expanding into this audience and adopting the OCPC strategy in the future. The advertising agent updates the parameter tuning rules of the OCPC strategy and the budget adjustment rules of short video platforms to the parameter tuning knowledge base, and will prioritize following these rules in future parameter tuning. The operations manager agent updates the conclusion that "the budget ratio of short video platforms can be increased to 55%" to the task management knowledge base, and will adjust the budget allocation when planning tasks in the future. All updated knowledge bases are synchronized to the shared space, and all agents can access the latest experience rules simultaneously.

[0149] Step S730: Using the optimization indicators agreed upon in the campaign task as the evaluation standard, compare all candidate campaign combinations in this campaign cycle with the benchmark campaign combination in all dimensions to select the optimal candidate campaign combination that has the best performance in optimization indicators and whose performance data meets the statistical significance requirements in this campaign cycle.

[0150] This step is a preliminary screening process for updating the benchmark campaign mix. Led by the operations manager's intelligent agent, its core objective is to select the truly optimal and reliable mix from the current round of candidate mixes. This will serve as the new benchmark for the next round, ensuring the iteration direction is positive and preventing benchmark degradation due to random fluctuations. First, using the core optimization metrics agreed upon in the campaign task as evaluation standards, the performance data of all candidate campaign mixes in this round are compared across all dimensions with the original benchmark mix. This comparison considers all core performance dimensions, including customer acquisition cost, ROI, and conversion rate, rather than just a single metric, ensuring the selected mix is ​​truly globally optimal. After the comparison, the candidate mix with the best performance in the optimization metrics is selected. Simultaneously, the performance data of this mix is ​​verified to meet statistical significance requirements, i.e., whether the confidence level and sample size of the mix meet preset thresholds. This ensures that the optimal performance is not due to random fluctuations but is stable and reliable, ultimately yielding the optimal candidate campaign mix for this round, which will serve as a candidate for the new benchmark.

[0151] Taking the cross-platform deployment scenario of credit products as an example, the optimization metric this time is the lead CPA. The original benchmark combination has a CPA of 90 yuan and an ROI of 2.5. The 45 candidate deployment combinations in this round are compared with the original benchmark deployment combination in all dimensions. Among them, the combination of low-interest creative + urban young and middle-aged target audience + OCPC strategy + short video platform has a lead CPA of 72 yuan and an ROI of 3.2, which is significantly better than the original benchmark deployment combination. At the same time, the weighted conversion number of this combination is 24 and the effect confidence level is 95%, which meets the requirements of minimum sample size and statistical significance. Therefore, this combination is selected as the optimal candidate deployment combination in this round for subsequent benchmark updates.

[0152] Step S740: The optimal candidate delivery combination selected is determined as the new benchmark delivery combination, and the benchmark version change nodes, corresponding performance data and experimental budget configuration rules are recorded simultaneously and written into the shared data space as a unified reference benchmark for multi-dimensional A / B testing in the next delivery cycle.

[0153] This step is the final stage of the iterative cycle. Following the optimal candidate campaign combination selected in step S730, it is executed by the operations manager agent. Its core function is to iteratively update the baseline campaign combination, providing a new benchmark for the next round of campaign testing and driving continuous iteration of campaign performance. In this step, the selected optimal candidate campaign combination is first determined as the new baseline campaign combination, replacing the old benchmark, and serving as the unified benchmark for multi-dimensional A / B testing in the next campaign cycle. Then, the benchmark version change node is recorded synchronously, including the benchmark update time, version number, and other information, to trace the benchmark's iteration history and facilitate subsequent review and problem localization. Simultaneously, the performance data corresponding to this new benchmark, as well as the experimental budget configuration rules, i.e., the experimental rules such as budget allocation and bidding configuration based on this benchmark, are recorded to guide the experimental configuration for the next round of campaigns. After recording, all updated information is written to the shared data space and synchronized to all agents, ensuring that all agents can use the new benchmark and configuration in the next round of campaigns. After completion, the system automatically returns to step S120 to start the next round of campaign optimization cycle.

[0154] Taking the cross-platform deployment scenario of credit products as an example, the optimal combination of low-interest creative content, urban young and middle-aged demographics, OCPC strategy, and short video platform is selected as the benchmark deployment combination for version V2, replacing the original version V1 benchmark. The benchmark change node is recorded as 2025-03-25, with corresponding performance data of lead CPA of 72 yuan and ROI of 3.2. The experimental budget configuration rules are 55% for short video platform, 30% for search, and 15% for information flow. OCPC bidding is prioritized. This information is written into the shared data space as a unified benchmark for the next round of A / B testing. The system then returns to step S120 to start the next round of deployment optimization.

[0155] This application's embodiments effectively address the industry pain points of traditional cross-platform advertising, such as the inability to achieve self-learning, the lack of experience accumulation, and the uncontrollable direction of iteration, through a full-process process of experience accumulation and benchmark iteration. Specifically, through multi-agent collaborative processing, firstly, by pre-screening effective conclusions, small-sample, low-confidence noise data is filtered out, ensuring the accuracy of experience accumulation from the source and avoiding interference from randomly fluctuating erroneous data with the system's learning logic. Then, by updating the dedicated knowledge base of each agent, reliable performance experience from this round of advertising is accumulated into the rule base of each business domain, enabling the system's self-learning capability. Simultaneously, the updated knowledge base is globally synchronized, enabling global sharing of experience, allowing all agents to make decisions based on the latest experience, significantly improving the accuracy of subsequent advertising decisions. Building on this, through comprehensive optimal combination screening, the truly stable and optimal advertising combination for this round is found, ensuring positive iteration of the benchmark advertising combination and avoiding benchmark degradation. Finally, by updating the benchmark and initiating the iteration cycle, the next round of testing benchmarks can be elevated to a new level, forming a closed-loop self-optimization cycle that drives continuous iterative improvement in advertising performance.

[0156] In some embodiments, refer to Figure 8 , Figure 8 This is a structural block diagram of a multi-agent collaborative advertising delivery optimization system provided in one embodiment of this application. This application also proposes a multi-agent collaborative advertising delivery optimization system for use with the multi-agent collaborative advertising delivery optimization method provided in any embodiment of this application. The system includes a business and configuration layer 810, a multi-agent collaboration layer 820, a data and feature layer 830, a monitoring and evaluation layer 840, and an external advertising platform layer 850. Wherein: The business and configuration layer 810 is used by business personnel to configure deployment targets and constraints, store risk control and compliance rules, and provide input and constraints for the multi-agent collaboration layer.

[0157] The multi-agent collaboration layer 820 includes an operations manager agent, a creative production agent, a target audience and strategy agent, and an advertising agent. Each agent collaborates to complete task decomposition, creative generation, strategy design, and campaign execution, and achieves information exchange through a shared data space.

[0158] The data and feature layer 830 includes a user profile and feature library, a historical deployment log library, an attribution result library, and a knowledge base, providing decision-making support for the multi-agent collaboration layer and storing full-process data.

[0159] The monitoring and evaluation layer 840 is used to perform data retrieval, preprocessing, A / B testing, multi-level attribution analysis, and optimization suggestion generation to feed data back to the multi-agent collaboration layer. The external advertising platform layer 850 includes multiple cross-channel advertising platforms, which connect with the advertising agent via API to receive campaign configurations and send back real-time campaign data.

[0160] The multi-agent collaborative advertising optimization system provided in this application embodiment can implement all the method steps implemented in the above method embodiment and achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0161] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for optimizing advertising delivery through multi-agent collaboration, characterized in that, The method includes: The operations manager intelligent agent receives the goals and constraints of advertising placement, parses the goals and constraints into machine-readable indicators and constraints, generates a structured placement task sheet, and writes it into the shared data space. Based on the delivery task order, the operations manager intelligent agent determines the benchmark delivery combination and breaks down the delivery task into standardized task cards corresponding to the creative production intelligent agent, the audience targeting and strategy intelligent agent, and the advertising delivery intelligent agent, and issues them out. The standardized task card includes at least the input data source, output data field, result acceptance quantitative conditions, and task execution deadline. The creative production agent generates and filters candidate creative packages based on the corresponding standardized task cards, audience profiles, advertising platform material specifications and compliance constraints, and writes the unique identifier and metadata corresponding to the candidate creative packages into the shared data space. The target audience and strategy agent generates candidate group targeting strategies and candidate delivery execution strategies based on the corresponding standardized task cards. It then combines and trims the candidate group targeting strategies and candidate delivery execution strategies with the candidate creative package to generate a candidate delivery combination list. The candidate delivery combination list is output to the advertising agent and written into the shared data space. Each candidate delivery combination in the candidate delivery combination list is a one-to-one correspondence of candidate creative package, candidate group targeting strategy, and candidate delivery execution strategy. The advertising agent creates a cross-platform advertising plan based on the advertising task order and the candidate advertising combination list, and writes the mapping relationship between the unique identifier of each advertising platform and the unique identifier of the candidate advertising combination into the shared data space; The advertising agent pulls and processes real-time advertising data from various advertising platforms based on the advertising task order. Using the benchmark advertising combination as a reference, it conducts multi-dimensional A / B testing on each candidate advertising combination in the candidate advertising combination list, automatically adjusts the bid and budget within preset constraints, and writes the parameter adjustment records into the shared data space. The monitoring and evaluation layer collects relevant data from various advertising platforms based on the delivery task order, and performs multi-level effect attribution analysis on the relevant data according to the creative dimension, audience targeting strategy dimension, delivery execution strategy dimension and advertising platform dimension, generates attribution results and writes them into the shared data space; Each agent updates its knowledge base based on the attribution results, and updates the candidate campaign combination with the best performance in this campaign cycle as the benchmark campaign combination. Then, the operations manager agent returns to the campaign task order to determine the benchmark campaign combination, break down the campaign task into standardized task cards corresponding to each agent, and distributes the standardized task cards to the corresponding creative production agent, audience targeting and strategy agent, and ad placement agent.

2. The method according to claim 1, characterized in that, The campaign task order should include at least the campaign period, a list of advertising platforms, total budget and channel allocation constraints, optimization metrics, evaluation window, attribution window, data retrieval granularity, probing period configuration, evaluation period configuration, and anomaly fallback strategy.

3. The method according to claim 2, characterized in that, After the operations manager intelligent agent generates the delivery task order, the method further includes: Based on the delivery task order, the operations manager intelligent agent determines the experimental scale, boundary parameters, and overall delivery plan draft for the current delivery cycle of advertising. Initialize the candidate campaign combination list skeleton in the shared data space, and agree on the unique identification rules for the candidate campaign combination. The candidate campaign combination list skeleton is a basic data structure that includes the unique identifier of the candidate campaign combination, the associated advertising platform, the creative identifier to be filled, the audience targeting strategy identifier to be filled, and the campaign execution strategy identifier to be filled.

4. The method according to claim 2, characterized in that, The creative production agent generates and filters candidate creative packages based on the corresponding standardized task cards, audience profiles, advertising platform material specifications, and compliance constraints, including: The creative creation agent reads the input data of the corresponding standardized task card, combines the audience profile, advertising platform material specifications and compliance constraints, and generates multiple versions of candidate advertising creatives through a large language model; All candidate ad creatives undergo multi-dimensional compliance verification to eliminate those that do not comply with industry regulatory requirements, ad platform material specifications, business risk control rules, and compliance constraints of the campaign task sheet. For candidate ad creatives that pass the compliance verification, similarity calculation is performed based on the core selling points of the creative, copy structure, visual elements, and target audience tags, and duplicate creatives with similarity exceeding the preset threshold are eliminated. The performance of the deduplicated candidate ad creatives is estimated, and the estimated click-through rate, estimated conversion rate, and estimated return on investment scores are obtained for each creative. Candidate ad creatives that meet the quantity requirements of the standardized task card are selected from high to low scores and packaged into a candidate creative package.

5. The method according to claim 2, characterized in that, The target audience and strategy agent generate candidate group targeting strategies and candidate delivery execution strategies based on the corresponding standardized task cards. These strategies are then combined and tailored with the candidate creative package to generate a candidate delivery combination list, including: The target audience and strategy intelligence agent reads the input requirements of the corresponding standardized task card, retrieves the audience profile and feature database data, and generates multiple sets of candidate group targeting strategies. Each set of candidate group targeting strategies includes audience selection rules, audience exclusion rules, expected reach scale, and risk control compliance labels. Each candidate group targeting strategy is mapped to the native targeting configuration parameters that each advertising platform can execute, so as to realize the cross-platform implementation of general audience targeting rules; For each candidate group targeting strategy that has completed cross-platform mapping, at least one candidate delivery execution strategy is matched. The candidate delivery execution strategy includes at least the bid type, initial bid upper and lower limits, exposure frequency control rules, delivery time period rules, and budget consumption rhythm control rules. The candidate ad creatives in the candidate creative package are systematically orthogonally combined with the candidate group targeting strategies and the candidate delivery execution strategies that match the audience targeting strategies to form a creative and strategy candidate matrix. Based on the threshold for the number of ad combinations, coverage requirements, effect prediction scores, and constraint feasibility rules agreed upon in the ad delivery task sheet, the creative and strategy candidate matrix is ​​pruned to remove invalid combinations that exceed the constraint boundaries, fail to meet the predicted effects, or have overlapping coverage. Each valid combination retained after cropping is assigned a unique identifier, forming a list of candidate deployment combinations.

6. The method according to claim 2, characterized in that, The advertising agent creates a cross-platform advertising plan based on the advertising task order and the candidate advertising combination list, and writes the mapping relationship between the unique identifier of each advertising platform and the unique identifier of the candidate advertising combination into the shared data space, including: The advertising agent reads the placement task list and the candidate placement combination list in the shared data space, and based on the total budget and channel allocation constraints in the placement task list, plans a placement plan hierarchy structure that matches the platform's native rules for each advertising platform, and determines the platform placement hierarchy mapping relationship corresponding to each candidate placement combination in the candidate placement combination list. Each candidate placement combination in the candidate placement combination list is mapped to the configuration parameters that can be executed under the corresponding advertising platform placement level. For each set of mapped configuration parameters, compliance verification, risk control verification, budget allocation verification, frequency control and deduplication verification are performed in sequence to remove candidate placement combinations that fail the verification. For candidate delivery combinations that pass all validations, calculate and configure the initial bid and initial budget based on the optimization metrics of the delivery task order and the delivery execution strategy corresponding to the candidate delivery combination; Call the standardized API interfaces of various advertising platforms to create corresponding campaign plans for candidate campaign combinations that have passed verification and completed initial parameter configuration; Obtain the unique identifier of the advertising platform corresponding to the campaign plan returned by each advertising platform, establish a one-to-one mapping relationship between the unique identifier of the advertising platform and the unique identifier of the candidate campaign combination, and write the mapping relationship into the shared data space.

7. The method according to claim 2, characterized in that, The advertising agent, based on the advertising task order, pulls and processes real-time advertising data from various advertising platforms. Using the benchmark advertising combination as a reference, it conducts multi-dimensional A / B testing on each candidate advertising combination in the candidate combination list. Within preset constraints, it automatically adjusts the bid and budget, and writes the parameter adjustment records into the shared data space, including: The advertising agent, in conjunction with the monitoring and evaluation layer, retrieves real-time delivery data from various advertising platforms according to the data granularity agreed upon in the delivery task order. It then performs standardized processing on the acquired real-time delivery data, including time zone alignment, data deduplication, transmission delay correction, and outlier filtering, to obtain a valid delivery dataset with unified statistical standards. Using the benchmark campaign combination as a unified benchmark, and based on the effective campaign dataset, parallel comparative A / B tests were conducted on each candidate campaign combination from the dimensions of creative content, audience targeting strategy, campaign execution strategy, and advertising platform. Based on the statistical results of A / B testing, and in accordance with the optimization indicators agreed upon in the campaign task sheet, the bids and budgets of each candidate campaign combination are automatically iteratively adjusted under preset constraints. The preset constraints include at least the upper limit of the single parameter adjustment range and the upper limit of the parameter adjustment frequency per unit time. The entire parameter tuning process, the values ​​before and after parameter changes, and the basis for adjustment are used to generate a parameter tuning record, which is then written into the shared data space.

8. The method according to claim 2, characterized in that, The monitoring and evaluation layer collects relevant data from various advertising platforms based on the delivery task order, and performs multi-level effect attribution analysis on the delivery-related data according to the creative dimension, audience targeting strategy dimension, delivery execution strategy dimension, and advertising platform dimension, generating attribution results and writing them into the shared data space, including: The monitoring and evaluation layer, in accordance with the evaluation window and attribution window agreed upon in the aforementioned delivery task order, gathers the full delivery logs and conversion feedback data from various advertising platforms, as well as the back-link quality data from the corresponding business systems, to form the original attribution dataset. The original attribution dataset is subjected to preprocessing operations in sequence, including user identification unification, delivery touchpoint structure construction, effective touchpoint extraction and deduplication, in order to remove invalid and redundant data and obtain a standardized attribution dataset with unified statistical caliber. Based on the standardized attribution dataset, and taking the candidate ad combination as the smallest statistical unit, multi-level effect attribution analysis is carried out from the dimensions of creative content, audience targeting strategy, ad execution strategy, and advertising platform to generate structured attribution results. The attribution results include at least the unique identifier of the candidate ad combination, weighted conversion number, conversion value, return on investment, customer acquisition cost, and effect confidence. The attribution results are written into the shared data space, and targeted delivery optimization suggestions are generated based on the attribution results.

9. The method according to claim 2, characterized in that, Each agent updates its knowledge base based on the attribution results and updates the best-performing candidate delivery combination in this round of delivery to the benchmark delivery combination, including: Each agent obtains the attribution results of the current deployment cycle from the shared data space and selects stable and valid conclusions that simultaneously meet the preset minimum sample size threshold and statistical confidence threshold. Based on the stable and effective conclusions after screening, each intelligent agent updates its dedicated knowledge base for its corresponding business domain. Specifically, the creative production intelligent agent updates the knowledge base for creative material effects and compliance rules, the audience targeting and strategy intelligent agent updates the knowledge base for audience targeting strategies and campaign execution strategies, the advertising agent updates the knowledge base for cross-platform campaign coordination and risk control rules, and the operations manager intelligent agent updates the knowledge base for campaign task management and effect evaluation rules. All updated knowledge base content is synchronized to the shared data space. Using the optimization indicators agreed upon in the aforementioned campaign task as the evaluation standard, all candidate campaign combinations in this campaign cycle are compared with the benchmark campaign combination in all dimensions to select the optimal candidate campaign combination that performs best in the aforementioned optimization indicators and whose performance data meets the statistical significance requirements in this campaign cycle. The selected optimal candidate campaign combination is determined as the new benchmark campaign combination, and the benchmark version change nodes, corresponding performance data and experimental budget configuration rules are recorded simultaneously and written into the shared data space as a unified benchmark for multi-dimensional A / B testing in the next campaign cycle.

10. A multi-agent collaborative advertising delivery optimization system, characterized in that, The system is used to perform the method according to any one of claims 1-9, the system comprising a business and configuration layer, a multi-agent collaboration layer, a data and feature layer, a monitoring and evaluation layer, and an external advertising platform layer; The business and configuration layer is used for business personnel to configure deployment targets and constraints, store risk control and compliance rules, and provide input and constraints for the multi-agent collaboration layer. The multi-agent collaboration layer includes an operations manager agent, a creative production agent, a target audience and strategy agent, and an advertising agent. Each agent collaborates to complete task decomposition, creative generation, strategy design, and campaign execution, and achieves information exchange through a shared data space. The data and feature layer includes a user profile and feature library, a historical deployment log library, an attribution result library, and a knowledge base, which provide decision-making basis for the multi-agent collaboration layer and store full-process data. The monitoring and evaluation layer is used to perform data retrieval, preprocessing, A / B testing, multi-level attribution analysis, and optimization suggestion generation to feed back data to the multi-agent collaboration layer. The external advertising platform layer includes multiple cross-channel advertising platforms, which connect with the advertising agent via API to receive placement configurations and transmit real-time placement data.

Images