Commercial data processing method based on multiple agents

Abstract

The invention provides a business data processing method based on multiple agents. A business ecology construction agent collects and preprocesses business source data, builds a prediction model, generates a construction scheme and screens participating business bodies; the central collaborative agent distributes the scheme and participated business body information to the artificial intelligence enabling agent, the task scheduling agent and the fund management agent, and a collaborative process is started; the artificial intelligence enabling agent calls the scenarized model library, generates task difficulty pre-judgment and human skill matching data, and outputs the task difficulty pre-judgment and human skill matching data to the task scheduling agent; the method comprises the following steps: constructing a participated commercial body human skill portrait, disassembling a task and pushing a sub-task to a matching terminal; the fund management agent executes fund management according to the progress node and dynamically matches task progress and fund; and the income distribution agent integrates fund and task quality data after the task is completed, accounts the income through a contribution degree quantification model, triggers intelligent contract distribution, and synchronizes a result to each agent and a participating commercial body.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a commercial data processing method based on multiple agents. BACKGROUND

[0002] With the deep development of digital economy, the complexity of commercial activities continues to increase, covering multiple core links such as industry ecosystem construction, project fund management, task execution landing, human resource allocation, and technology resource integration. Commercial data processing has become a key support for enterprises to improve operational efficiency and achieve sustainable growth. Currently, various industries generally manage commercial data through digital tools, trying to optimize business processes through data-driven optimization.

[0003] In particular, in recent years, with the explosive development of artificial intelligence (AI) technology, the processing of commercial business processes and data through artificial intelligence technology has become an important means to improve the efficiency of commercial business.

[0004] In some related technologies disclosed, such as the patent entitled "A profit distribution marketing system and method based on an agent large model" (application number CN202510503280.6), an implementation scheme is given for effectively improving the accuracy of marketing and the rationality of profit distribution using AI technology. The profit distribution marketing system and method based on the agent large model provided by the patent are used to solve the problem of user cold start when performing profit distribution marketing. It includes real-time collection of user behavior data on different platforms, construction of user portraits based on behavior data, generation of optimal marketing strategies based on large models and reinforcement learning algorithms, use of optimal marketing strategies for users, optimization of resources and dynamic balance of interests based on marketing results through multi-agent game algorithms and multi-agent collaboration algorithms, custom setting of corresponding profit distribution ratios according to business needs, obtaining of final profit distribution rules, obtaining of product attributes, user intent, and context relationships through visual-language models, intelligent recommendation and marketing content generation, and execution of profit distribution rules and revenue settlement according to the final profit distribution rules.

[0005] However, with the in-depth development of commercial activities of commercial entities, current AI technology applications in commercial data processing are mostly focused on single links, such as AI empowerment for revenue distribution only. At the same time, the depth of integration of existing AI applications and business scenarios is insufficient, resulting in AI technology being unable to fully support data analysis and decision-making. The efficiency of AI in commercial data processing is greatly limited. Therefore, a commercial data processing method that can realize AI multi-link empowerment has become a problem to be solved in the commercial field. SUMMARY

[0006] The embodiment of the present application provides a kind of intelligent contract data processing method, to realize the collaborative ability of commercial data processing in business system of multiple agents.

[0007] In the first aspect, the embodiment of the present application provides a business data processing method based on multiple agents, running in a business system, comprising: a business ecosystem construction agent collects commercial source data for preprocessing, builds a business ecosystem prediction model, and generates business ecosystem construction scheme data based on the commercial source data, and correspondingly screens out participating commercial bodies; the central coordination agent distributes the business ecosystem construction scheme data and participating commercial body information to the artificial intelligence enabled agent, the task scheduling agent and the fund management agent, triggering the start of the agent collaboration process; the artificial intelligence enabled agent generates task difficulty prediction and human skill matching data based on the business ecosystem construction scheme data, and outputs them to the task scheduling agent; the task scheduling agent combines the task difficulty prediction and human skill matching data to build a human skill portrait of the participating commercial body, and completes task decomposition, and pushes subtasks to the terminals of the matched participating commercial bodies; the fund management agent executes fund management at the progress node according to the progress node pushed by the task scheduling agent, dynamically matches the task progress with the required funds; the revenue distribution agent integrates the fund data of the fund management agent and the task quality data of the task scheduling agent after the task is completed, calculates the revenue through a contribution quantification model, triggers the execution of the smart contract for distribution, and synchronizes the distribution results to each agent and participating commercial body.

[0008] Optionally, in the business data processing method based on multiple agents, the business ecosystem construction agent collects commercial source data for preprocessing and builds a business ecosystem prediction model, comprising: collecting commercial source data including industry data of a business data platform, production and sales data of an enterprise resource planning system, and trend data of an industry database; preprocessing the commercial source data, using a hybrid strategy of forward filling and mean completion to handle missing data, and eliminating outliers; building a long short-term memory network trend prediction model based on a tensor flow machine learning framework, inputting the preprocessed commercial source data to generate an ecosystem construction scheme.

[0009] Optionally, in the business data processing method based on multiple agents, the corresponding screening of participating commercial bodies comprises: calculating the scores of participating commercial bodies based on their enterprise qualifications and / or historical business cooperation records and production capacity and / or enterprise service capabilities using a weighted scoring method, and screening participating commercial bodies with scores exceeding a first threshold.

[0010] Optionally, in the multi-agent-based business data processing method, the artificial intelligence empowered agent calls a scenario-based model library to generate task difficulty prediction and human skill matching data based on business ecosystem construction scheme data, including: based on industry labels and / or business task attribute data in the business ecosystem construction scheme data, adopting a rule engine to match corresponding business scenario models in a business scenario-based artificial intelligence model library; generating business task difficulty prediction data according to at least one of the data volume, process node number, cross-agent collaboration frequency, urgency, task delivery deadline and current remaining period of the business task in the business application scenario model; obtaining skill level adaptation data and technical resource requirement list data of the participating business entity as human skill matching data according to the business application scenario model.

[0011] Optionally, in the multi-agent-based business data processing method, the task scheduling agent combines task difficulty prediction and human skill matching data to construct a human skill portrait of the participating business entity, including: receiving task difficulty coefficients, core technical requirement data in the task difficulty prediction data, and skill labels, skill adaptation suggestion data in the human skill matching data, converting unstructured text data into numerical features through word vector mapping, and mapping discrete skill level data to a fixed interval through normalization processing; dividing human skill dimensions based on a knowledge graph, filtering information from the numerical features and / or the fixed interval data, and constructing a skill tree, and dynamically adjusting the weights of features in the skill tree according to task requirements; combining at least one quantitative indicator of skill mastery level, task execution proficiency, and cross-scenario adaptation ability, calculating skill scores through a weighted summation algorithm, and mapping to the human skill level of the participating business entity.

[0012] Optionally, in the multi-agent-based business data processing method, further comprising: triggering human skill portrait update of the participating business entity when at least one of task execution feedback data, skill improvement proof data or task difficulty prediction update data is received; performing corresponding feature update and model training update on the update data; setting an accuracy evaluation index of the human skill portrait of the participating business entity, and adjusting feature extraction rules and quantitative model parameters when the accuracy evaluation index is lower than a threshold value, to realize optimization processing of the human skill portrait of the participating business entity.

[0013] Optionally, in the multi-agent-based business data processing method, the task decomposition includes: using a business process model and a marking method to analyze the process structure of the original task, and converting it into a directed acyclic graph containing node task steps and dependency edges; extracting strongly connected components in the directed acyclic graph through a graph theory algorithm, identifying atomic task units that cannot be decomposed as the smallest granularity task units for decomposition.

[0014] Optionally, in the multi-agent-based business data processing method, the fund management at the execution progress node and the dynamic matching of the task progress and the required funds include: based on the dependency relationship in the directed acyclic graph of sub-tasks, using the critical path method to identify the fund allocation node, and splitting the total funds to each task node according to the sub-task weight; collecting sub-task progress data in real time to generate sub-task progress value; calculating the amount of funds to be allocated based on the preset sub-task progress-fund function; when the sub-task fund pool balance is less than the amount of funds to be allocated, allocating part of the funds according to the influence degree of the sub-task on the total target task.

[0015] Optionally, in the multi-agent-based business data processing method, further comprising: the ecological evolution agent analyzes the process data in the business system to generate process optimization strategy data, and reversely synchronizes to the ecological construction agent and / or the task scheduling agent to iterate the rules of the ecological construction agent and / or the task scheduling agent.

[0016] Optionally, in the multi-agent-based business data processing method, the calculating the revenue through the contribution quantification model includes: constructing a contribution index model according to the dimensions of fund input, task quality, technical support, and ecological recommendation; using a dimension comprehensive quantification algorithm to map and normalize the index data of each dimension, and calculating the contribution score of the participating business by a weighted summation model; defining a structured smart contract template, combining the contribution quantification result with the allocation rule to output the revenue data of the business.

[0017] The technical scheme of the application breaks the problem of coordination lag in the traditional business process through the unified distribution and triggering mechanism of the central coordination agent: after the business ecological construction agent preprocesses data and generates a scheme, the central agent accurately pushes the data to the AI empowerment agent, task scheduling agent and other agents, realizes seamless connection of "data-decision-execution", avoids artificial transmission and repeated processing across links, and significantly improves the response speed of the whole process. By enhancing the accuracy of task matching and resource allocation, relying on the invocation of the scenario-based model library of the AI empowerment agent, combining task difficulty prediction and human skill matching data, the task scheduling agent provides scientific decision basis. The task scheduling agent constructs a fine human skill portrait and disassembles the task accordingly, ensures that the subtasks are highly adapted to the skills and resources of the participating business body, solves the problem of "task and ability mismatch" in traditional allocation, and improves resource utilization efficiency. By realizing the dynamic coordination of funds and task progress, the risk fund management agent executes dynamic fund management according to the task progress node, avoids idle or insufficient funds through the "progress-fund" real-time matching mechanism. At the same time, the income distribution agent ensures the objectivity and fairness of the distribution result based on the contribution quantification model of multi-source data, reduces the process blockage caused by fund distribution disputes, and reduces the risk of business cooperation. The construction of an extensible business ecological closed loop forms a complete closed loop of "ecological construction-task execution-fund management-income distribution" through standardized data interaction of each agent, and the output data of each link is used as the input of other links, supporting the continuous iteration of the business process. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A flowchart of a business data processing method based on multiple agents provided by the first embodiment of the present application is provided. Figure 2 A flowchart of agent cooperation provided by the second embodiment of the present application is provided. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0021] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0022] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0023] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0024] In this embodiment of the invention, a unified distribution and triggering mechanism of the central collaborative intelligent agent breaks through the problem of lag in collaboration in traditional business processes: after the business ecosystem building intelligent agent preprocesses data and generates solutions, the central intelligent agent accurately pushes the data to AI-enabled, task scheduling, and other intelligent agents, achieving seamless connection of "data-decision-execution," avoiding manual transmission and repetitive processing across links, and significantly improving the response speed of the entire process. By enhancing the accuracy of task matching and resource allocation, relying on the scenario-based model library of the AI-enabled intelligent agent, combined with task difficulty prediction and human skill matching data, a scientific decision-making basis is provided for the task scheduling intelligent agent. Based on this, the task scheduling intelligent agent constructs a detailed human skill profile and decomposes tasks, ensuring that sub-tasks are highly compatible with the skills and resources of participating business entities, solving the problem of "task and ability mismatch" in traditional allocation, and improving resource utilization efficiency. By realizing dynamic collaboration between funds and task progress, the risk of fund management is reduced. The fund management intelligent agent performs dynamic fund management based on task progress nodes, and avoids idle funds or insufficient supply through a real-time "progress-fund" matching mechanism. Meanwhile, the revenue distribution agent, based on a contribution measurement model using multi-source data, ensures the objectivity and fairness of the distribution results, reduces process blockages caused by disputes over fund allocation, and lowers the risks of business cooperation. A scalable business ecosystem closed loop is constructed, with each agent forming a complete closed loop of "ecosystem building - task execution - fund management - revenue distribution" through standardized data interaction. The output data of each link serves as the input for other links, supporting the continuous iteration of business processes.

[0025] Example One Figure 1This is a flowchart illustrating a multi-agent-based business data processing method provided in Embodiment 1 of the present invention. Embodiment 1 describes the application of the multi-agent-based business data processing method of the present invention on the "YJXS" ​​platform and the "ABCD" e-commerce small business. This embodiment uses the "YJXS" ​​business platform (hereinafter referred to as "YJXS") as a carrier and takes the small e-commerce enterprise "ABCD" (mainly selling home furnishing and cultural creative products, such as wooden storage boxes and hand-painted decorative paintings, with annual sales of 5-8 million yuan and 20 employees) participating in the entire process of the "YJXS 618 Home Furnishing and Cultural Creative Special Promotion" as a scenario to reveal in detail the implementation process of the multi-agent-based business data processing method.

[0026] In this embodiment, the hardware deployment of "YJXS" ​​is preferably configured with a dual-machine hot standby server Intel Xeon Gold 6426Y and Kafka / Flink cluster nodes for the central collaborative intelligent agent, and the software environment uses Python 3.10, TensorFlow 2.12.0, Hyperledger Fabric 2.5, etc. "ABCD" as a participating commercial entity accesses the platform through the "YJXS" ​​merchant backend and Enterprise Resource Planning (ERP) system interface, and its commercial activities cover the entire chain of "ecosystem access - task execution - fund management - revenue distribution".

[0027] Step 100: The intelligent agent for building the business ecosystem collects and preprocesses business source data, builds a business ecosystem prediction model, and generates business ecosystem construction plan data based on the business source data, thereby selecting participating business entities accordingly. In this embodiment, for example, the "YJXS" ​​business ecosystem building intelligence agent serves as the "planning and screening core" of the "618 Home Furnishing and Cultural Creativity Special" ecosystem. Through multi-source e-commerce data collection, standardized preprocessing, LSTM trend prediction, and weighted scoring screening, it completes the generation of ecosystem solutions and the screening of participants (including "ABCD"). The specific implementation process is as follows: 1.1 Implementation of Commercial Source Data Collection The multi-source data acquisition module for building the intelligent agent of the business ecosystem is developed based on Python 3.10 and Pandas 2.0, and deployed on the "YJXS" ​​edge computing node (Intel Xeon Silver 4410Y CPU, 64GB DDR5 memory). It collects four types of core business source data related to "home furnishing and cultural creative e-commerce" through a hybrid approach of "API calls + direct system connection + web crawling." Specific details are as follows: Industry data from the business data platform: This involves calling the API of the National Bureau of Statistics' "Culture and Related Industries Statistical Monitoring Platform" (interface address: https: / / data.stats.gov.cn / api / culture, call frequency: 1 time / day) to collect 12 indicators, including "Annual market size of home furnishing and cultural creative products (120 billion yuan in 2023, a year-on-year increase of 18%)", "Penetration rate of sub-categories (12% penetration rate of wooden storage boxes, 8% penetration rate of hand-painted decorative paintings)", and "Regional consumption heat (35% in East China, 28% in South China)". Each data collection session is approximately 300MB, data transmission uses HTTPS encryption, and the interface response time is ≤2.5 seconds. Production and sales data from the Enterprise Resource Planning (ERP) system: Through the "YJXS" ​​merchant ERP interface, directly connect to the "ABCD" Yonyou U8 Cloud ERP system (version V6.1) to collect its production data (daily production capacity of 200 wooden storage boxes, raw material inventory of 3000 sets), sales data (average daily orders of 45 in the past 3 months, average order value of 89 yuan, repurchase rate of 15%), and inventory data (inventory of 500 hand-painted decorative paintings, inventory turnover rate of 2.3 times / month). The data push delay is ≤1 second, and the average daily data collection volume for a single enterprise is approximately 150MB. Industry database trend data: The iResearch Consulting "Home Furnishing E-commerce Industry Database" API (version V4.0, call frequency 1 time / hour) was used to collect 8 trend indicators, including "618 Home Furnishing and Cultural Creative Promotion Historical Conversion Rate (12.5% ​​in 2023)", "Changes in Consumer Preferences (Search volume for minimalist style products increased by 25%)", and "Price Sensitivity (Products with an average order value of 80-150 yuan had the highest order rate)". The data volume collected per time was about 50MB, and the data update delay was ≤10 minutes. Historical operational data within the platform: Activity data of home furnishing and cultural and creative merchants over the past two years were extracted from the "YJXS" ​​own database (MySQL 8.0 master-slave architecture), including "average ROI of merchants during the 618 special event: 1:4.2", "average customer complaint rate during the event: 0.8%", and "correlation between inventory preparation cycle and stockout rate (stockout rate < 5% if inventory preparation cycle < 15 days)". The data was extracted once a day, with each extraction containing approximately 200MB of data.

[0028] All collected data is stored in real time to the “YJXS” MongoDB 6.0 sharded cluster (3 sharded nodes, each with 16 cores and 32GB of memory). Each data entry is labeled with data_source (data source, such as “NBS_API” or “YUEJU_ERP”) and collect_time (collection timestamp, accurate to milliseconds), supporting full-chain traceability.

[0029] 1.2 Implementation of Commercial Source Data Preprocessing In response to the characteristics of e-commerce data—namely, high real-time performance, low missing rate, but a high likelihood of outliers—the intelligent agent for building the business ecosystem employs a hybrid strategy of "forward imputation + mean completion" to handle missing data, and removes outliers using the 3σ principle. The specific operation process is as follows: Data format standardization: The original data is uniformly formatted. Date data is converted to "YYYY-MM-DDHH:MM:SS" (e.g., "20240501" is converted to "2024-05-01 00:00:00"). Numeric data is retained to 4 decimal places (e.g., "45 orders / day" is converted to "45.0000 orders / day"). Text data (e.g., "product style tags") is deduplicated and filtered for stop words using the jieba word segmentation library (version V0.42.1) (retaining core words such as "minimalist style" and "retro style"). Missing data handling: Discrete data (such as the enterprise qualification level and product style tag of "ABCD") adopts the forward filling method: if the "new product style tag" of "ABCD" is missing on May 10, 2024, and the previous valid data (May 9) is "minimalist style", then fill it with "minimalist style" to ensure the consistency of non-continuous attributes; For continuous data (such as the daily average order volume and inventory turnover rate of "ABCD"), the mean completion method is used: calculate the arithmetic mean of the valid data of the field in the past 30 days. For example, if the inventory turnover rate of "ABCD" on May 12 is missing, and the average effective turnover rate in the past 30 days is 2.1 times / month, then fill it with "2.1000 times / month". Special scenario handling: For "618 pre-sale deposit data" (no data in the first 7 days of the event), fill in the amount using "data from the same period last year × this year's growth coefficient" (the average daily pre-sale deposit in the same period of 2023 was 12,000 yuan, and the home furnishing and cultural creative industry grew by 15% in 2024, so fill in 12,000 × 1.15 = 13,800 yuan). After preprocessing, the data integrity was verified, with a missing rate of ≤1.2% for each field and a data accuracy of ≥98.8%. Outlier removal: The 3σ principle is used to detect outliers in continuous data (such as the daily order peak and average order value of "ABCD"). The steps are as follows: Calculate the average daily order volume of "ABCD" over the past 3 months: μ = 45 orders, σ = 8 orders. The threshold range was set as [μ-3σ, μ+3σ]=[45-24, 45+24]=[21, 69]. On May 8th, there were 92 orders in a single day (due to a mistaken operation resulting in duplicate orders), which exceeded the threshold and was verified as abnormal, so they were removed. On May 15th, there were 72 orders (due to a temporary flash sale event in the store), which was verified as real data and marked as "event abnormal" and retained.

[0030] All preprocessed data is stored in the “YJXS” Hadoop 3.3.6 HDFS system (5 DataNodes, 4TB HDD per node), using GZIP compression (62% compression rate), with a storage block size of 128MB and 3 replicas (storage across data centers).

[0031] 1.3 Construction of Business Ecosystem Prediction Model and Generation of "618 Special" Plan The business ecosystem building agent uses TensorFlow 2.12.0 to construct a Long Short-Term Memory Network (LSTM) trend prediction model. Inputting preprocessed data generates an ecosystem building plan for the "618 Home Furnishings and Cultural Creativity Special". Specific parameters and execution details are as follows: LSTM model parameter configuration (adapted for home and cultural creative scenarios): Model Level Core Parameter Configuration Input Layer Input Feature Dimension 55 dimensions (including industry growth rate, "ABCD" type small business order fluctuation, consumer preference, etc.), Min-Max normalized to [0,1] Hidden Layer 3-layer hidden layer, 128 neurons per layer, activation function ReLU, Dropout rate 0.2 (to prevent overfitting) Output Layer Fully connected layer, 8 neurons (corresponding to "number of participating merchants", "total inventory", "promotion budget", etc. 8 core scheme indicators) Training Parameters Adam optimizer (initial learning rate 0.001, decay coefficient 0.9), training rounds 100 rounds, batch size 32, MSE loss function Model training results: Using data from the "YJXS" ​​618 home furnishing campaign (1825 samples) from 2019 to 2023 as the training set and data from January to April 2024 (120 samples) as the test set, the model's prediction error after training was <4.2%, with "prediction accuracy of participating merchant order volume" at 95.3% and "prediction accuracy of promotion budget utilization" at 94.1%. "618 Home Furnishings and Cultural Creativity Special" Ecosystem Building Plan (including "ABCD" compatible content): Types and number of ecosystem participants: 30 home furnishing and cultural creative merchants (including small businesses of categories “ABCD”), 3 logistics service providers (covering East China / South China), 2 customer service outsourcing teams, and 1 data analysis service provider; Core cooperation model: Merchants adopt a dual model of "pre-sale + spot goods" (pre-sale inventory ≥ 60% of estimated sales), the platform bears 40% of the promotional advertising costs, and the merchant bears 60%; logistics adopts "regional warehousing" (East China warehouse covers the main customer groups "ABCD"). Key performance indicators (KPIs) require: merchant inventory preparation period ≤ 12 days, defect rate ≤ 0.8%; logistics delivery time ≤ 48 hours (≤ 24 hours in core cities); customer service response time ≤ 5 minutes. Customized requirements for "ABCD": Stock quantity ≥ 1500 units (based on a 120% increase in historical sales), ≥ 30 SKUs listed, and coupon settings must cover the unified platform rules for "¥15 off for purchases over ¥99" and "¥40 off for purchases over ¥199".

[0032] The scheme data is stored in JSON format, with fields including scheme_id (SS_618_2024_001), generate_time (2024-05-05 09:30:00), participant_demand (participant requirements), and key_indicators (key indicators), and is synchronized to the central collaborative intelligent agent.

[0033] 1.4 Participation in the screening and implementation of business entities The intelligent agent for building the business ecosystem, targeting the "618 Special Event," uses a weighted scoring method to select participating businesses based on four dimensions: "company qualifications, historical cooperation, production capacity, and service capabilities." The selection process for "ABCD" is as follows: Screening criteria and weight settings: Indicator Category Specific indicators (home creative scene) Weight Scoring criteria (full score 100 points) Enterprise Qualification Business License / Food Business License (if containing creative food), platform residence time, no violation record 40% 40 points for complete qualification, 10 points for residence ≥2 years, 10 points for no violation in the past 1 year (total score ≤60 points, 40 points after 40% conversion) Historical Cooperation Record Number of "YJXS" activity participations in the past 1 year, activity satisfaction score, order fulfillment rate 30% Participation ≥2 times 15 points, satisfaction ≥90 points 15 points, fulfillment rate ≥98% 10 points (total score ≤40 points, 30 points after 30% conversion) Production Capacity Replenishment cycle, maximum daily production capacity, inventory turnover rate 15% Replenishment cycle ≤12 days 10 points, daily production capacity ≥100 pieces 8 points, turnover rate ≥2 times / month 7 points (total score ≤25 points, 15 points after 15% conversion) Enterprise Service Capability Customer service response time, after-sales problem solving rate, customization service capability (such as personalized packaging) 15% Response time ≤5 minutes 8 points, solving rate ≥95% 7 points, with customization capability 5 points (total score ≤20 points, 15 points after 15% conversion) “ABCD” rating calculation: Enterprise Qualification: Complete qualifications (40 points), 2.5 years of operation (+10 points), no violations in the past year (+10 points) --> 40 points after conversion (60×40%+10×40%+10×40%? Correction: Calculated according to the total score of the indicator category, the actual score of the qualification category is 50 points --> 50×(40 / 60)=33.33 points); Historical cooperation: Participation in 1 activity (10 points), satisfaction score of 92 (15 points), fulfillment rate of 99% (10 points) --> actual score of 35 points --> 35 × (30 / 40) = 26.25 points; Production capacity: 10-day lead time (10 points), 120 units per day (8 points), turnover rate 2.1 times / month (7 points) --> Actual score 25 points --> 25 × (15 / 25) = 15 points; Service capabilities: Response time 4 minutes (8 points), resolution rate 96% (7 points), customized packaging available (5 points) --> Actual score 20 points --> 20 × (15 / 20) = 15 points; Overall score = 33.33 + 26.25 + 15 + 15 = 89.58 points; Threshold setting and screening results: The first threshold is set at 80 points. A score of 89.58 for "ABCD" is ≥80 points, and the merchant is successfully selected as a participant in the "618 Special". Among the 30 merchants screened at the same time, small businesses account for 60% (all of which meet the score ≥80 points). The screening time is 28 seconds per merchant, and the results are synchronized to the central collaborative intelligent agent.

[0034] Step 101: The central collaborative intelligent agent distributes the business ecosystem construction plan data and participating business entity information to the AI-enabled intelligent agent, the task scheduling intelligent agent, and the fund management intelligent agent, triggering the start of the intelligent agent collaborative process. In this embodiment, the central collaborative intelligent agent of "YJXS" ​​acts as a "data hub and collaborative brain," accurately distributing the "618 Special Event" plan and information on participants such as "ABCD," triggering multi-agent collaboration. The specific implementation process is as follows: 2.1 Hardware and Software Deployment of the Central Collaborative Intelligent Agent Hardware configuration: Dual-machine hot standby server (Intel Xeon Gold 6426Y, 24 cores and 48 threads, 256GB DDR5-4800 memory, 10TB NVMe SSD RAID 10); Software environment: Operating system: CentOS Stream 9 (kernel 5.14.0-362.el9.x86_64); Data interaction protocol: JSON-LD standardized protocol, extending the "YJXS" ​​exclusive field, with the following format example: { "agent_id": "CENTRAL_AGENT_SSY_001", "data_type": "ecological_scheme", "timestamp": 1714867800000, "data_content": {"scheme_id": "SS_618_2024_001", "participant_list": ["ABCD", ...]}, "digital_signature": "RSA-2048 signature value", "ssy_extend": {"business_scene": "618_home_culture", "data_expire":3600}} Distributed coordination: ZooKeeper 3.8.1 cluster (3 nodes), state update frequency 100ms / time, coordination latency <50ms; Caching and Database: Redis 7.0 cluster (3 masters and 3 slaves), cache expiration time is 1 hour; MySQL 8.0 master-slave architecture, master-slave synchronization latency is <1 second.

[0035] 2.2 Data Preprocessing and Signature Generation After receiving the solution data (3.2MB) and participant information (1.8MB) pushed by the commercial ecosystem building agent, the central collaborative agent executes: Format validation: Validate the integrity of JSON-LD fields (e.g., whether agent_id "data_type" exists), the validity of the "ssy_extend" field (whether business_scene is "618_home_culture"), and the data size (single record ≤ 10MB). Validation pass rate is 100%. Digital signature: The data_content is hashed using SHA-256 (generating a 32-byte hash value), and then encrypted with the central private key (RSA-2048, stored in the “YJXS” KMS key system) to generate a signature, ensuring that the data cannot be tampered with.

[0036] 2.3 Data Distribution Implementation The central collaborative intelligent agent distributes data to the three target intelligent agents according to the principle of "caching first + offline persistence", and the specific path is as follows: Target agent information: Agent Name agent_id Deployment Address Interface Protocol Response Requirements Artificial Intelligence Empowered Agent AI_EMPOWER_SSY_002 192.168.50.100:8080 gRPC Response Time <20ms Task Scheduling Agent TASK_SCHEDULE_SSY_003 192.168.50.101:8081 RESTful Response Time <20ms Fund Management Agent FUND_MANAGE_SSY_004 192.168.50.102:8082 RESTful Response Time <20ms Distribution process: Data is transmitted to the target agent via TLS 1.3 encryption with a timeout of 3 seconds and a retrieval period of 2 times (500ms interval). The data is first written to the Redis cache (key format: agent_id:data_type:timestamp), and then asynchronously synchronized to MySQL; Taking the distribution of "ABCD" information as an example: the response time for pushing to the AI-enabled intelligent agent is 11ms, for pushing to the task scheduling intelligent agent is 13ms, and for pushing to the fund management intelligent agent is 12ms, all of which meet the requirements.

[0037] 2.4 Triggering the Start of the Intelligent Agent Collaboration Process When the central coordinating agent detects that all of the following conditions are met, the coordination process is triggered to start: All three target agents returned a "successful reception" response (including response_time "agent_status:ready"); In the ZooKeeper cluster, the state of each agent is synchronized to "pending collaboration"; The resource utilization rate of the central intelligent agent is ≤65% (CPU 62%, memory 60%, bandwidth 58%). The start command is sent via ZooKeeper. The command includes flow_id (FLOW_618_2024_001), start_time (1714867800000), and task_desc ("Based on the 618 Home Furnishing and Cultural Creative Solution, complete the task difficulty prediction, manpower skill matching and fund management"). After receiving the command, each agent enters the data processing stage.

[0038] Step 102: AI-enabled intelligent agents, based on data from the business ecosystem construction plan, call upon the scenario-based model library to generate task difficulty prediction and human skill matching data, and output them to the task scheduling intelligent agent. In this embodiment, the AI-enabled intelligent agent of "YJXS" ​​generates difficulty prediction and skill matching data for the "ABCD" 618 event task through scenario-based model library matching and multi-dimensional quantification. The specific implementation process is as follows: 3.1 Construction and Matching of Scenario-Based Model Library Model library configuration (adapted for home and cultural creative scenarios): Deployed on the “YJXS” GPU cluster (2 NVIDIA A100 40GB), it includes 8 types of home furnishing e-commerce scenario models, with the core models as follows: Model Name Model Type Training Data Volume Prediction Accuracy Adaptation Task Type Home e-commerce task difficulty model XGBoost 1.7.5 300,000 historical task data ≥92% Replenishment, listing, promotion, order processing Merchant skill matching model Knowledge graph + logistic regression 250,000 merchant data ≥91% Human skill level adaptation Model matching execution: The task attributes of “ABCD” were extracted from the “618 Plan”: industry_tag=home furnishing and cultural creation, task_type=pre-sale preparation + product listing + coupon setting + order processing. The rule engine (Drools 8.44.0) matched the “home furnishing e-commerce task difficulty model” and the “merchant skill matching model”, with a matching time of 180ms.

[0039] 3.2 Generation of Task Difficulty Prediction Data Artificial intelligence-enabled intelligent agents, based on a matching model, quantify the four core tasks of "ABCD" in multiple dimensions to generate a difficulty coefficient (1-10 points, with 10 points being the highest): Quantitative Indicators and Weights (Customized Tasks for Home Furnishing E-commerce): Quantitative dimension Specific indicators (take "pre-sale replenishment task" as an example) Weight Scoring criteria (1-10 points) Data volume Replenishment SKU quantity, estimated pre-sale order quantity 25% SKU≥30 gets 8 points, order quantity≥500 gets 8 points (take the average) Process node number Preparation process steps (purchase of raw materials---> production---> quality inspection---> warehousing) 25% Steps≥4 gets 7 points Cross-agent collaboration times Collaboration times with ecological construction (demand confirmation), fund management (preparation fund), task scheduling (progress synchronization) 15% Collaboration≥3 times gets 6 points Emergency degree The ratio of preparation deadline to current remaining period (deadline May 20, current May 5, remaining 15 days) 20% Remaining period / deadline period = 15 / 12=1.25---> gets 5 points (ratio>1 gets 5 points, 1-0.8 gets 7 points, <0.8 gets 10 points) Task complexity Whether to customize (such as personalized packaging), whether to contain cross-category preparation 15% Contain customization gets 8 points, cross-category (storage box + decorative picture) gets 6 points (take the average) Difficulty calculation of the "ABCD" pre-sale inventory preparation task: Data volume score: SKU=35 (8 points), order volume=580 (8 points) --> 8 points; Process node score: 4 steps --> 7 points; Collaboration frequency score: 3 times --> 6 points; Urgency score: 5 points; Complexity score: (8+6) / 2=7 points; Difficulty level = (8×25% +7×25% +6×15% +5×20% +7×15%) = 2 + 1.75 + 0.9 + 1 + 1.05 = 6.7 points (Difficulty level: Medium-High); Similarly, the difficulty coefficients for other tasks are calculated as follows: "Product Listing Task" 4.2 points, "Coupon Setting Task" 3.8 points, and "Order Processing Task" 5.3 points. The difficulty prediction data format is as follows: { "task_id": "TASK_YUEJU_001", "task_name": "ABCD618 Pre-sale Stock Preparation", "task_type": "stock preparation", "difficulty_coefficient": 6.7, "difficulty_level": "Medium to high (6-8 points)", "quantitative_details": {"data volume":8,"process nodes":7,"number of collaborations":6,"urgency level":5,"complexity":7}, "risk_tips": "Inventory preparation needs to be completed 3 days in advance to avoid affecting shelf listing." 3.3 Generation of Human Skills Matching Data AI-enabled intelligent agents generate skill level matching data and a list of technical resource requirements in "ABCD" categories based on a "merchant skill matching model": Skill level adaptation data (for teams "ABCD"): Post Core skill dimension Required skill level (1-5 level) Level description (adapt to task requirements) Operation clerk Product listing efficiency, coupon setting ≥3 level 3 level: daily average listing SKU≥20, coupon setting error rate≤1% Warehouse manager Preparation progress control, inventory checking ≥4 level 4 level: preparation period≤12 days, checking accuracy≥99.5% Customer service representative Presale order consultation response, after-sales processing ≥3 level 3 level: response time≤5 minutes, after-sales problem solving rate≥95% List of technical resource requirements: Resource type Specific demand (adapt to "ABCD") Hardware resources Preparation management computer (i5 CPU, 16GB memory) 2, handheld code scanner (scanning speed≥200 times / minute) 2, warehouse shelves (bearing≥50kg) 10 groups Software resources "YJXS" merchant background V5.8, ERP system (U8 Cloud V6.1 of use friend), customer service chat system (support automatic reply) Tool resources Product packaging materials (customized paper box≥2000), logistics slip printer (printing speed≥100 / hour) 1 The above data is pushed to the task scheduling agent through the central collaborative agent, with a push delay of 14ms and a reception confirmation rate of 100%.

[0040] Step 103: The task scheduling agent combines task difficulty prediction and human skill matching data to construct a human skill profile of the participating business entities, completes task decomposition, and pushes sub-tasks to the matching terminals of the participating business entities. In this embodiment, the task scheduling agent of "YJXS" ​​completes the construction of human skill profiles, dynamic optimization, and task decomposition and push for "ABCD". The specific implementation process is as follows: 3.1 Construction of Human Skills Profiles Data preprocessing and feature transformation: Unstructured data transformation: The "customer service specialist skill description" (such as "familiar with pre-sale order consultation scripts") is transformed into numerical features (such as [0.182, 0.456,..., 0.721]) using the Word2Vec word vector model (gensim 4.3.2, 200-dimensional). Discrete data normalization: Map "Warehouse Manager Skill Level 4" to [0,1] using Min-Max normalization -- (4-1) / (5-1)=0.75; Skill Dimension Classification and Skill Tree Construction: Based on the "YJXS" ​​e-commerce knowledge graph, and divided into "ABCD" human resource skill dimensions, taking "warehouse manager" as an example, the following skill tree graph is constructed (see the graph code illustration below): graph TD A [Core skills in warehouse management (weight 0.5)] --> B [Inventory preparation and control (weight 0.4)] A --> C [Inventory Management (Weight 0.3)] A --> D [Device operation (weight 0.3)] B --> B1 [Stock Preparation Progress Control (Quantitative Value 0.82)] B --> B2 [Quality Inspection Collaboration (Quantitative Value 0.78)] C --> C1 [Inventory Count (Quantitative Value 0.91)] C --> C2 [Out of Stock Alert (Quantitative Value 0.85)] D --> D1 [Bar scanner operation (quantization value 0.88)] D --> D2 [Shelf Management (Quantitative Value 0.76)] Skill score calculation and level mapping: We selected "skill mastery (weight 0.5), task proficiency (weight 0.3), and cross-scenario adaptability (weight 0.2)" as quantitative indicators, taking "warehouse manager" as an example: Skill mastery level: Weighted average of skill tree quantification values ​​= (0.82×0.4 +0.78×0.4 +0.91×0.3 +0.85×0.3 +0.88×0.3 +0.76×0.3) / (0.4+0.4+0.3+0.3+0.3+0.3) = 0.83; Task proficiency: Completion rate of the last 3 stock preparation tasks: 98% --> 0.98; Cross-scenario adaptation: Experience in preparing inventory for both "in-stock" and "pre-sale" scenarios --> 0.85; Skill score = 0.83×0.5 +0.98×0.3 +0.85×0.2 = 0.415 + 0.294 + 0.17 = 0.879; Level mapping: 0.879 --> Level 4 (rule: ≥0.9 --> Level 5, 0.8 -0.89 --> Level 4), matching the requirement of "stocking task ≥ Level 4"; Example of a "ABCD" warehouse manager's skills profile: { "participant_id": "YUEJU_STAFF_001", "name": "Zhang XX (Warehouse Manager)", "role": "Warehouse Management", "update_time": "2024-05-06 14:30:00", "skill_tree": {...}, / / Skill tree structure "quantitative_indicators": {"mastery":0.83,"proficiency":0.98,"adaptability":0.85}, "skill_score": 0.879, "skill_level": 4, "matching_tasks": ["ABCD618 Pre-sale Stock Preparation Tasks"]} 3.2 Update and Optimization of Human Resource Skill Profiles Update trigger condition: After "ABCD" completes the "stock preparation rehearsal task" on May 10, the task scheduling agent receives feedback data (stock preparation cycle 11 days, 1 day shorter than expected, inventory accuracy 99.8%). Update process: Feature Updates: The quantitative value for "Stock Preparation Progress Control" has been increased from 0.82 to 0.88, and the quantitative value for "Inventory Count" has been increased from 0.91 to 0.95. Incremental model training: Add the pre-training data to the training set and update the Word2Vec model (training 20 rounds). Image reconstruction: Skill score = (0.86×0.5 +0.99×0.3 +0.85×0.2) = 0.43 + 0.297 + 0.17 = 0.897 --> still level 4 (close to level 5); Accuracy assessment optimization: The "Profile Matching Accuracy" was set to ≥92%. On May 12th, it was detected that this rate had dropped to 89% (due to the addition of one customer service representative for "ABCD" whose skills were not recorded in time). Adjustments were made accordingly. Feature extraction rules: Add "New employee training record" feature (weight 0.1); Model parameters: The weight of "Task Proficiency" was adjusted from 0.3 to 0.25, and the weight of "Training Records" was adjusted to 0.05; After optimization, the matching accuracy rebounded to 93.5%.

[0041] Preferably, in this embodiment, the updating and optimization of the human skill profile can be carried out as follows: when at least one of task execution feedback data, skill improvement proof data, or task difficulty prediction update data is received, the updating of the human skill profile of the participating business entity is triggered; corresponding feature updates and model training updates are performed on the updated data; an accuracy evaluation index for the human skill profile of the participating business entity is set, and when the accuracy evaluation index is lower than a threshold, the feature extraction rules and quantization model parameters are adjusted to achieve the optimization of the human skill profile of the participating business entity.

[0042] 3.3 Task Decomposition and Implementation BPMN (Business Process Model and Notation) process analysis and DAG (Directed Acyclic Graph) transformation: Camunda Modeler 5.22.0 was used to parse the "ABCD 618 Total Task", the process of which is "Pre-sale preparation --> Product listing --> Coupon setting --> Order receipt --> Order processing --> After-sales service", and it was converted into a directed acyclic graph (DAG). The nodes are the task steps and the edges are the dependencies (such as "Preparation of goods --> Listing"). Atomic task unit identification: The Kosaraju algorithm is used to extract strongly connected components of the DAG, identify indivisible atomic tasks, and finally decompose them into 8 subtasks: Subtask ID Subtask name Dependent subtask Execution subject Execution time limit SUB_YUEJU_001 Raw material procurement None ABCD procurement clerk May 6-8 SUB_YUEJU_002 Cultural and creative product production SUB_YUEJU_001 ABCD production team May 9-15 SUB_YUEJU_003 Product inspection and warehousing SUB_YUEJU_002 ABCD warehouse manager May 16-17 SUB_YUEJU_004 Product information input and listing SUB_YUEJU_003 ABCD operation clerk May 18-19 SUB_YUEJU_005 618 coupon creation and configuration None ABCD operation clerk May 20-21 SUB_YUEJU_006 Presale order receiving and statistics SUB_YUEJU_004 YJXS order system June 1-15 SUB_YUEJU_007 Order sorting and delivery SUB_YUEJU_006 ABCD warehouse team June 1-18 SUB_YUEJU_008 After-sales problem handling (refund / exchange) SUB_YUEJU_007 ABCD customer service team June 1-20 Preferably, in this embodiment, the scheme for decomposing subtasks into subtasks can be carried out as follows: the process structure of the original task is parsed using a business process model and a labeling method, and it is converted into a directed acyclic graph containing nodes, task steps, nodes, and dependency edges; the strongly connected components in the directed acyclic graph are extracted using graph theory algorithms, and indivisible atomic task units are identified as the smallest granularity task units for decomposition.

[0043] Subtask push: Pushed via WebSocket protocol to the merchant backend (PC) and employee APP (mobile) of "ABCD", the push data is encrypted (AES-256), and the push delay is less than 1 second. Example of a "product listing task" push received by "ABCD" operations specialist: { "sub_task_id": "SUB_YUEJU_004", "task_name": "Product Information Entry and Listing", "description": "Enter details for 35 SKUs (including size, material, and after-sales policy) and complete the listing process in the YJXS backend." "deadline": "2024-05-19 24:00:00", "dependency": "SUB_YUEJU_003 (Quality Inspection and Warehousing Completed)", "required_skill": "Operations Specialist ≥ Level 3", "operation_guide": "Log in to the YJXS merchant backend --> Product Management --> Add Product --> Fill in the form according to the template"} Step 104: The fund management agent executes fund management at the progress nodes pushed by the task scheduling agent, dynamically matching the task progress with the required funds. In this embodiment, the "YJXS" ​​fund management intelligent agent, targeting the "ABCD" 618 activity funds (total budget of 800,000 yuan), achieves matching of funds with task progress through critical path identification, progress tracking, and dynamic allocation. The specific implementation process is as follows: 4.1 Identification of Fund Disbursement Nodes and Breakdown of Total Funds Implementation of the Critical Path Method: Based on the "ABCD" subtask DAG, calculate the earliest / latest start / completion time of each subtask, identify the critical path as "raw material procurement --> production --> quality inspection and warehousing --> product shelving --> order processing --> after-sales service", and determine 5 fund disbursement nodes: Disbursement node ID Corresponding sub-task completion node Disbursement trigger condition FUND_NODE_001 Raw material procurement completion Upload of purchase invoice and acceptance of qualified products FUND_NODE_002 Product quality inspection and warehousing completion Confirmation of warehousing order and next product rate ≤0.8% FUND_NODE_003 Product listing completion and 10-day promotion Listing SKU≥30 and promotion period over half (June 10) FUND_NODE_004 618 activity completion Order completion rate≥95% and after-sales rate≤5% FUND_NODE_005 After-sales completion All after-sales work orders closed and no unresolved disputes Total funds breakdown: The total fund of 800,000 yuan was divided using the "subtask weighting method". The weights were calculated based on "task cost ratio (0.5), task impact (0.3), and timeliness requirement (0.2)". The results are as follows: Disbursement node Split amount Proportion Corresponding sub-task cost details FUND_NODE_001 24 million yuan 30% Raw material procurement cost (wood materials 18 million + packaging materials 6 million) FUND_NODE_002 16 million yuan 20% Production and processing fee 12 million + quality inspection fee 4 million FUND_NODE_003 16 million yuan 20% Operation and promotion fee 8 million + platform service fee 8 million FUND_NODE_004 16 million yuan 20% Logistics fee 10 million + labor cost 6 million FUND_NODE_005 8 million yuan 10% After-sales reserve fund (refund / exchange cost) 4.2 Real-time collection of subtask progress data The funds management agent collects the progress of sub-tasks "ABCD" through three channels: Merchants proactively reported: "ABCD" purchasing specialists uploaded purchase invoices through the "YJXS" ​​funds module (completed on May 8, progress 100%), and warehouse managers uploaded inbound delivery slips (completed on May 17, progress 100%). The system automatically synchronizes the product listing progress from the "YJXS" ​​merchant backend (35 SKUs completed on May 19, progress 100%) and the promotional period order data from the order system (order completion rate 65% on June 10, progress 65%). IoT device data collection: Inventory changes are collected through the smart shelf sensors in the "ABCD" warehouse (1,520 items were received on May 17, progress 100%). Progress data is stored in real time to the Redis 7.0 cache, updated every minute, and a "progress value" (e.g., "FUND_NODE_002" progress value 100%) is generated.

[0044] 4.3 Calculation of Amount Due and Disbursement of Funds Progress - Funding Function Settings: The amount to be allocated using a linear function = node-specific amount × progress value (progress value is 0-100%), as shown in the example below: FUND_NODE_001 (240,000 yuan): Progress on May 8th: 100% --> Amount to be disbursed = 240,000 × 100% = 240,000 yuan; FUND_NODE_003 (160,000 yuan): Progress as of June 10th: 65% --> Amount to be disbursed = 160,000 × 65% = 104,000 yuan; Disbursement Execution: The funds management agent called the "YJXS" ​​payment interface (connected to UnionPay Corporate Payment V3.0) and transferred 240,000 yuan to the "ABCD" corporate account on May 8, with a confirmation time of 8 minutes; and transferred 104,000 yuan on June 10, with a confirmation time of 7 minutes. Handling insufficient fund pool balance: On June 10, when “FUND_NODE_003” was supposed to receive 104,000 yuan, the fund pool balance was only 80,000 yuan (due to the platform’s temporary budget adjustment). 80,000 yuan was allocated based on the “sub-task impact” (order processing tasks had an impact of 90% > after-sales tasks had an impact of 60%). The difference of 24,000 yuan was paid out on June 12 after the balance was replenished, ensuring that the core tasks were not affected.

[0045] In this embodiment, the solution for dynamically matching task progress with required funds at execution progress nodes can preferably be implemented as follows: Based on the dependencies in the directed acyclic graph of subtasks, the critical path method is used to identify fund disbursement nodes, and the total funds are distributed to each task node according to the weight of the subtasks; subtask progress data is collected in real time to generate subtask progress values; the amount to be disbursed for a subtask is calculated based on a preset subtask progress-funds function; when the balance of the subtask fund pool is less than the amount to be disbursed, a portion of the funds is disbursed according to the subtask's impact on the overall target task.

[0046] Step 105: After the task is completed, the revenue distribution agent integrates the financial data of the fund management agent and the task quality data of the task scheduling agent, calculates the revenue through a contribution measurement model, triggers the smart contract to execute the distribution, and synchronizes the distribution results to each agent and participating business entity. In this embodiment, the revenue distribution agent of "YJXS" ​​calculates and distributes the revenue for "ABCD" after the 618 event ends (June 20th). At the same time, it combines the ecological evolution agent to optimize the process. The specific implementation process is as follows: 5.1 Integration of Revenue Accounting Data The revenue-sharing agent obtains the core data of "ABCD" from the two main agents: Data from the fund management intelligent agent: Total investment of 800,000 yuan (amount at each disbursement node: 24 + 16 + 16 + 16 + 8), actual expenditure of 782,000 yuan (saving 18,000 yuan due to reduced raw material procurement costs), and platform subsidy of 120,000 yuan (40% advertising expense subsidy). Task scheduling agent data: Task quality score 92 points (out of 100, inventory completion rate 100%, listing accuracy rate 99%, after-sales resolution rate 96%), ecosystem recommendation contribution (recommended 2 similar small businesses to join, and received a recommendation reward).

[0047] 5.2 Contribution Measurement Model Accounting Contribution index model construction: Indicator dimension Specific indicators (for merchants) Weight Indicator value (ABCD) Normalization processing (mapped to 0-1) Fund investment Actual investment amount / platform average investment amount (platform home furnishing merchant average investment 60 million yuan) 40% 78.2 / 60≈1.303 1.303 / 2=0.651 (upper limit 2) Task quality Task quality score / 100 30% 92 / 100=0.92 0.92 Technical support Whether to use platform recommended tools (such as intelligent customer service), tool usage rate 20% Use and usage rate 90% 0.9 Ecological recommendation Recommended number of registered merchants / platform recommendation threshold (threshold 3) 10% 2 / 3≈0.667 0.667 Contribution score calculation: Contribution score = 0.651×40% +0.92×30% +0.9×20% +0.667×10% = 0.2604 + 0.276 + 0.18 + 0.0667 = 0.7831; Revenue Calculation: The platform's "618 Home Furnishing Special" has a total revenue pool of 5 million yuan, which is distributed according to "contribution score / total score of all merchants". The revenue of "ABCD" is 500 × (0.7831 / 28.56) = 500 × 0.0274 ≈ 137,000 yuan (total score of all merchants is 28.56). With the platform subsidy of 120,000 yuan, the total revenue is 257,000 yuan.

[0048] In this embodiment, preferably, the calculation of revenue through the contribution quantification model can also be implemented in the following way: a contribution index model is constructed based on the dimensions of capital investment, task quality, technical support, and ecosystem recommendation; a dimensional comprehensive quantification algorithm is used to normalize the mapping of the index data of each dimension, and the contribution score of the participating business entity is calculated through a weighted summation model; a predefined structured smart contract template is used to combine the contribution quantification results with the allocation rules to generate the revenue data of the business entity.

[0049] 5.3 Smart Contract Execution Allocation Smart contract template: Predefined structured smart contracts (based on Hyperledger Fabric 2.5), with core terms including: Distribution rules: Profit = Basic profit + Subsidy profit, with basic profit calculated based on contribution score; Payment method: Transferred to the merchant's corporate account within T+3 business days; Tax withholding: Value-added tax withheld at 13% (25.7 × 13% ≈ 33,410 yuan, actual payment 223,590 yuan); Contract Triggering and Execution: The revenue distribution agent calls the contract interface, passes in the revenue data of "ABCD" (257,000 yuan) and account information. After the contract is verified, it is executed and the fund transfer is completed on June 23. The arrival confirmation time is 12 minutes. Results synchronized: The allocation results are synchronized to: Each intelligent agent: Funds management intelligent agent (updates fund flow records), task scheduling intelligent agent (updates merchant credit scores); “ABCD”: The merchant’s backend “Revenue Center” displays the details (total revenue 257,000, deductions 33,410, actual payment 223,590), and sends SMS notifications simultaneously.

[0050] 5.4 Optimization of Ecological Evolutionary Intelligent Agent Process After the activity concluded, the ecological evolutionary agent analyzed the process data of "ABCD" and identified two major areas for optimization: Question 1: The raw material procurement cycle of "ABCD" (3 days) is longer than the platform's average cycle (2 days), resulting in a tight buffer time for inventory preparation; Question 2: During the product listing process, "ABCD" made mistakes in filling in the information for two SKUs due to unfamiliarity with the new template, affecting listing efficiency; Data on the generation and optimization strategies of ecological evolutionary intelligent agents: Reverse synchronization to the business ecosystem to build an intelligent agent: optimize the selection rules for participants, and add "raw material procurement cycle ≤ 2 days" as a bonus item (weight 5%). Reverse synchronization to the task scheduling agent: Add a "template tutorial pop-up" to the "Product Listing Task" and push task reminders one day in advance; The optimization strategy took effect on July 1st, and the average procurement cycle for subsequent merchants was shortened to 2.1 days, with the listing error rate reduced by 60%.

[0051] Preferably, in this embodiment, the ecological evolution agent can be configured to further analyze the process data in the business system, generate process optimization strategy data, and back-synchronize it to the ecological building agent and / or task scheduling agent to iterate the rules of the ecological building agent and / or task scheduling agent.

[0052] Embodiment two: To more clearly reveal the implementation details of the collaborative work of various intelligent agents in the commercial system involved in this invention, see [link to relevant documentation]. Figure 2 , Figure 2 A flowchart of the intelligent agent collaborative operation provided in Embodiment 2 of the present invention is given.

[0053] This second embodiment of the invention uses the "Cross-border E-commerce Ecosystem Supply Chain Collaboration Project" as a specific application scenario. Led by the e-commerce platform enterprise XYZ, the project aims to integrate various business entities, including upstream suppliers, cross-border logistics service providers, customs clearance agents, and overseas warehousing companies, to achieve end-to-end collaboration from product preparation, cross-border transportation, customs declaration to overseas delivery. The following details the complete process of business data processing involving multiple intelligent entities within the business entity system.

[0054] Process 1: The ecosystem building intelligent agent outputs a resource requirement list to the central collaborative intelligent agent. 1.1 Commercial Source Data Acquisition and Preprocessing The intelligent agent for ecosystem building first initiates a commercial source data collection process, clearly defining the scope of collection to cover three core types of data: Industry data from the business data platform: By connecting with the official cross-border e-commerce import and export statistics platform and iResearch's cross-border e-commerce industry database, it collects 12 types of macro-level industry data, including the North American cross-border e-commerce market size in 2024 (approximately US$1.2 trillion), the growth rate of cross-border circulation of 3C products (18% year-on-year), and the congestion index of major logistics channels. Production and sales data from the Enterprise Resource Planning (ERP) system: Access the ERP system of the e-commerce platform enterprise XYZ (using SAP S / 4HANA 2023 version) to collect 18 types of operational data for the past 6 months, including order volume of core 3C products (average 50,000 orders per month), inventory turnover rate (1.2 times / month), sales share of the best-selling SKU Top20 (68% in total), and supplier production cycle (average 15 days). Industry database trend data: By calling the cross-border e-commerce trend database API, we can obtain 10 types of trend data, including changes in consumer preferences for 3C products in the North American market (demand for wireless charging devices increased by 25%), cross-border logistics cost fluctuation index (quarterly fluctuation ±5%), and dynamic changes in customs clearance policies.

[0055] The data preprocessing stage employs a hybrid strategy of forward padding and mean completion: For the three missing supplier production cycle data in the ERP system, the forward filling method was used, and the supplier's previous production cycle record (14 days) was used. For the five missing monthly market size data in the industry data, the mean imputation method was used to calculate the average of the data from the three consecutive months (US$102 billion) to fill in the missing data. Outliers were removed using the 3σ criterion. Two extreme data points (+20% and -18%) that deviated from the mean by three times the standard deviation in the logistics cost volatility index were identified as outliers and removed to ensure data validity.

[0056] 1.2 Construction of Business Ecosystem Prediction Model and Generation of Solutions A Long Short-Term Memory (LSTM) network trend prediction model was built based on the TensorFlow 2.15 machine learning framework. The model parameters were set as follows: Input layer dimension: 40 dimensions (corresponding to 40 classes of preprocessed feature data); Hidden layers: 3 layers, the first layer has 128 neurons, the second layer has 64 neurons, and the third layer has 32 neurons; Activation functions: The hidden layer uses the ReLU function, and the output layer uses the Sigmoid function; Number of iterations: 200 rounds; Batch size: 32; Optimizer: Adam optimizer, learning rate 0.001; Loss function: Mean Squared Error (MSE).

[0057] The preprocessed commercial source data was divided into training and testing sets in a 7:3 ratio. The training set was used to optimize model parameters, and the testing set was used to verify model accuracy. The final model prediction accuracy reached 89.7%. Based on the model output, a "Cross-border E-commerce Ecosystem Supply Chain Collaborative Construction Plan" was generated, which clarified the core objectives of the project (reducing the order fulfillment cycle from 45 days to 25 days and reducing logistics costs by 12%), the types of ecosystem participants (suppliers, logistics providers, customs clearance agents, overseas warehouses, payment institutions), and key information such as collaborative process nodes (stock preparation, transportation, customs clearance, warehousing, and delivery).

[0058] 1.3 Participation in the screening of business entities A weighted scoring method was used to select participating businesses, with the following scoring indicators and weights: Scoring indicators Weight Scoring details (full score 100) Enterprise qualification 30% Complete business license and cross-border business qualification for 30 points, deduct 15 points for missing one item Historical business cooperation record 25% ≥10 times of cooperation with core platform in the past 2 years for 25 points, 5-9 times for 15 points, 1-4 times for 5 points, no cooperation record for 0 points Production capacity / service capacity 25% Supplier daily production capacity≥500 pieces for 25 points, logistics merchant daily processing cross-border parcels≥1000 pieces for 25 points, customs clearance agent daily customs clearance≥50 for 25 points Enterprise service capacity 20% Customer complaint rate≤1% for 20 points, 1%-3% for 10 points, >3% for 0 points The first threshold was set at 60 points. Through automatic system calculation and manual review, 15 qualified participating businesses were selected from 30 candidate businesses, including: 3 core suppliers (mainly engaged in 3C charging equipment), 4 cross-border logistics providers (covering China-US and China-Canada routes), 2 customs clearance agencies (with North American Customs AEO certification), 3 overseas warehousing companies (located in Los Angeles, New York, and Toronto), and 3 payment institutions (including Alipay Cross-border, PayPal Enterprise, and WorldFirst).

[0059] 1.4 Resource Requirement List Generation and Transmission Based on the above schemes and selection results, the ecosystem building intelligent agent has compiled a resource requirement list, which includes the following core contents: Supplier resources: Suppliers must be able to provide an average daily inventory of 500 3C charging devices, respond to replenishment requests within 7 days, and have a product qualification rate of ≥99.5%. Logistics resources: Must have the capacity to transport 1,000 cross-border parcels per day, with a transit time of ≤12 days for China-US routes and a damage rate of ≤0.5%; Customs clearance resources: Must have the capacity to process an average of 50 customs clearance shipments per day, with a clearance time of ≤3 days and a clearance success rate of ≥99%; Warehousing resources: 5,000 square meters of overseas warehouse storage area is required, with an inventory turnover rate of ≥1.2 times / month and an order outbound delivery time of ≤24 hours; Payment resources: Must support multi-currency settlement (USD, CAD), transfer fee ≤0.8%, and arrival time ≤2 business days.

[0060] The ecosystem building agent pushes the resource requirement list to the central collaborative agent through a RESTful API interface (using HTTPS encrypted transmission and JSON data format). The interface response time is ≤300ms. After the push is completed, a "Resource Requirement List Transmission Receipt" is generated, which records the transmission time, data integrity verification results (MD5 value matching), and other information, and starts the resource scheduling process.

[0061] Step 2: The central coordinating agent requests data from the human resource database and technical resource pool from the task scheduling agent, and the task scheduling agent returns the resource data. 2.1 Initiating a Data Request After receiving the resource request list, the central coordinating agent initiates a data query request to the task scheduling agent through an internal message queue (using RabbitMQ version 3.12). The request message contains the following core fields: Request ID: QS20240601001 (Format: QS + Date + Serial Number); Request type: Repository query; Required resource types: Human resources pool, technology resource pool; Search criteria: Suitable for cross-border e-commerce supply chain collaboration scenarios, supporting end-to-end collaboration for 3C products; Response time: 1000ms.

[0062] The request transmission uses the AES-256 encryption algorithm to ensure data security, and the message queue is equipped with a retry mechanism (maximum number of retries 3, retry interval 500ms) to avoid request failure due to network fluctuations.

[0063] 2.2 Human Resources Database Data Preparation The human resources database maintained by the task scheduling agent is stored in MySQL 8.0, deployed on an Alibaba Cloud ECS server (configuration: 8 cores, 16GB memory, 500GB SSD storage), supporting 1000 concurrent queries per second. The core data of the human resources database includes: Basic personnel information: Total number of employees in the business entity (1200 people), job types (8 categories including production, logistics, customs clearance, warehousing, and customer service), and years of service (average 5.2 years); Skill tag data: The Word2Vec word vector model is used to convert unstructured skill descriptions into numerical features. For example, "cross-border logistics route planning" is mapped to vector [0.82, 0.35, 0.61], and "interpretation of North American customs clearance policies" is mapped to vector [0.78, 0.42, 0.59]. Skill level data: Discrete skill levels (beginner, intermediate, advanced, and expert) are mapped to the [0,1] interval through Min-Max normalization, with beginner corresponding to 0.25, intermediate to 0.5, advanced to 0.75, and expert to 1.0; Available status data: Real-time synchronization of employees' current work status (idle, busy, on leave), with ≥60% of employees in an idle state, and support for responding to task assignments within 7 days.

[0064] Based on knowledge graphs, human skill dimensions are divided, and a skill tree structure is constructed: Primary dimensions: Supply chain operations (weight 0.4), technical support (weight 0.3), customer service (weight 0.3); Secondary dimension: Supply chain operations include inventory management (weight 0.3), logistics scheduling (weight 0.3), customs clearance coordination (weight 0.2), and warehouse management (weight 0.2). The third level of dimension: Inventory management includes inventory monitoring (weight 0.5), replenishment response (weight 0.5), and so on.

[0065] Based on the collaborative task requirements of cross-border e-commerce supply chain, the feature weights in the skill tree are dynamically adjusted. For example, the weight of the logistics scheduling dimension is adjusted from 0.3 to 0.4, and the weight of the customs clearance coordination dimension is adjusted from 0.2 to 0.3, to adapt to the core needs of cross-border transportation and customs clearance.

[0066] Combining three quantitative indicators—skill mastery (weight 0.4), task execution proficiency (weight 0.3), and cross-scenario adaptability (weight 0.3)—a weighted summation algorithm is used to calculate the skill score: Skill Score = Skill Mastery × 0.4 + Task Execution Proficiency × 0.3 + Cross-Scenario Adaptability × 0.3. For example, a logistics dispatcher with a skill mastery of 0.8, task execution proficiency of 0.7, and cross-scenario adaptability of 0.6 would have a skill score of 0.8 × 0.4 + 0.7 × 0.3 + 0.6 × 0.3 = 0.71, which corresponds to an advanced skill level.

[0067] 2.3 Data Preparation for the Technical Resource Pool The technology resource pool adopts a distributed storage architecture and covers the following core technology resources: Logistics Management System: Adopts WMS version 5.0, supports real-time tracking of logistics routes, transportation cost accounting, damage rate statistics, and API interface response time ≤200ms; Customs clearance declaration system: integrates the General Administration of Customs clearance declaration interface, supports automatic filling of declaration documents and policy compliance verification, and has a customs clearance data processing speed of ≥10 declarations / minute; Warehouse Management System: Supports real-time inventory monitoring, intelligent warehouse allocation, and rapid order fulfillment; inventory data update latency ≤500ms. Payment and settlement system: Supports multi-currency conversion, automatic fee calculation, and tracking of payment status; settlement success rate ≥99.9%. Data encryption tool: Employs the RSA-2048 encryption algorithm to ensure the security of cross-border data transmission, and supports data anonymization and access control.

[0068] The technical resource pool data includes fields such as resource name, version number, function description, interface parameters, availability status, and service response time. It uses Redis 7.0 to cache hot resource data, with a cache hit rate of ≥90%, improving data query efficiency.

[0069] 2.4 Resource Data Return The task scheduling agent completes data querying and organization from the human resources database and technical resource pool within 800ms, generating standardized response data (JSON format), which includes the following core content: Response ID: XS20240601001 (Format: XS + Date + Serial Number); Request ID: QS20240601001 (corresponds one-to-one with request ID); Human Resources Summary: There are 720 available employees, including 288 senior skilled employees, covering core positions such as logistics scheduling and customs clearance coordination; Technical resources summary: 12 available systems, 100% interface availability, supporting full-process collaboration in cross-border e-commerce; Data verification code: MD5 value (32-bit string).

[0070] The response data is transmitted to the central collaborative agent via HTTPS encryption. After receiving the data, the central collaborative agent verifies its integrity using MD5 checksum. If the verification is successful, the data reception time is recorded (accurate to milliseconds). If the verification fails, the task scheduling agent is triggered to re-push the data.

[0071] Step 3: The central collaborative agent generates a resource scheduling plan and pushes it to the task scheduling agent. 3.1 Triggering of Agent Collaboration The central collaborative intelligent agent integrates the resource requirement list and the returned resource data, and synchronizes basic data with the AI-enabled intelligent agent and the fund management intelligent agent through an internal bus (using the Spring Cloud Stream framework), triggering the start of the intelligent agent collaborative process. The synchronized data includes core information such as project scenario description, resource requirement specifications, and existing resource inventory, with a synchronization latency of ≤500ms.

[0072] 3.2 Task Difficulty Prediction and Human Skill Matching After receiving basic data, the AI-enabled intelligent agent uses a rule engine (Drools version 8.44) to match the "cross-border supply chain collaboration model" in the business ecosystem construction solution data with the industry tag "cross-border e-commerce supply chain" and the task attribute data "3C product full-process collaboration".

[0073] This model is optimized for cross-border e-commerce scenarios and has the following core functions: Task Difficulty Quantification: The data volume of the business task (50,000 monthly orders for 3C products, weight 0.3), number of process nodes (5 nodes including inventory preparation, transportation, customs clearance, warehousing, and delivery, weight 0.2), number of cross-agent collaborations (estimated 8 times, weight 0.1), urgency level (normal, weight 0.2), task delivery deadline and current remaining cycle (90-day cycle, 60 days remaining, weight 0.2) are quantified. The task difficulty coefficient is calculated by weighted summation: 0.3×0.8+0.2×0.6+0.1×0.4+0.2×0.3+0.2×0.7=0.58, which is judged as a medium difficulty task. Human resource skills matching: Output skill level adaptation data (logistics scheduling requires advanced or higher skills, customs clearance coordination requires intermediate or higher skills) and technical resource requirements list (logistics management system and customs clearance declaration system must be deployed, and warehouse management system should be deployed first).

[0074] The AI-enabled intelligent agent pushes task difficulty prediction data (difficulty coefficient 0.58, medium difficulty level) and human skill matching data to the task scheduling intelligent agent in ProtoBuf format (serialization efficiency is 30% higher than JSON). After the push is completed, a collaborative log is generated.

[0075] 3.3 Task Breakdown The task scheduling agent uses a business process model and labeling method (BPMN 2.0) to parse the process structure of the original task "Cross-border E-commerce 3C Product Supply Chain Full-Process Collaboration", transforming it into a directed acyclic graph (DAG) containing 12 task steps and 15 dependency edges. A code example of the DAG is described below: graph TD A [Order Received] --> B [Stock Preparation Requirement Generated] B --> C [Supplier prepares goods] C --> D [Stock Preparation Quality Inspection] D --> E [Cross-border Transportation Dispatch] E --> F [Logistics Tracking Monitoring] F --> G [Customs Clearance Declaration] G --> H [Customs Clearance and Review] H --> I [Overseas Warehouse Inbound] I --> J [Order Outbound] J --> K [Overseas Shipping] K --> L [Order Receipt] D --> M [Handling of Non-conforming Products] M --> C The Tarjan algorithm (a graph theory algorithm) is used to extract strongly connected components from the directed acyclic graph (DAG), identifying six indivisible atomic task units: supplier inventory preparation unit, cross-border transportation unit, customs clearance and declaration unit, overseas warehouse storage unit, overseas delivery unit, and quality inspection unit. These serve as the smallest granularity task units for decomposition. Each atomic task unit specifies its input parameters (e.g., inventory preparation unit input: quantity and specifications), output results (e.g., inventory preparation unit output: qualified products and inventory preparation report), and execution time limit (e.g., inventory preparation unit time limit: 7 days).

[0076] 3.4 Resource Scheduling Scheme Generation The central collaborative intelligent agent combines resource demand lists, human skill matching data, and task breakdown results to formulate a resource scheduling plan. The core content is as follows: Supplier resource allocation: The three core suppliers are allocated the replenishment tasks in a ratio of 2:2:1. Supplier A is responsible for replenishing an average of 200 units per day, Supplier B is responsible for replenishing an average of 200 units per day, and Supplier C is responsible for replenishing an average of 100 units per day. The replenishment response time is specified to be ≤7 days. Logistics resource allocation: The four logistics providers are assigned tasks according to the routes. Logistics providers 1 and 2 are responsible for the China-US route (each undertaking 500 pieces / day of transportation), and logistics providers 3 and 4 are responsible for the China-Canada route (each undertaking 500 pieces / day of transportation). The transportation time is ≤12 days. Customs clearance resource allocation: Two customs clearance agents will divide tasks by region. Customs clearance agent 1 will be responsible for customs clearance at US ports (average 30 shipments per day), and customs clearance agent 2 will be responsible for customs clearance at Canadian ports (average 20 shipments per day). The clearance time is ≤3 days. Warehousing resource allocation: 3 overseas warehouses cover different regions. The Los Angeles warehouse is responsible for delivery in the western United States (2000㎡ storage area), the New York warehouse is responsible for delivery in the eastern United States (2000㎡ storage area), and the Toronto warehouse is responsible for delivery in Canada (1000㎡ storage area). Outbound delivery time is ≤24 hours. Human resource allocation: Each atomic task unit is assigned a dedicated team, the logistics scheduling unit is assigned 10 senior skilled employees, and the customs clearance coordination unit is assigned 8 intermediate or above skilled employees; Technical resource allocation: Deploy logistics management systems and customs declaration systems for all participating businesses, additionally deploy inventory monitoring modules for core suppliers, and deploy intelligent warehousing modules for overseas warehouses.

[0077] The plan clarifies key information such as the resource allocation time (all resource deployments to be completed before June 10, 2024), allocation method (system integration using API interface, personnel allocation using project-specific configuration), and responsible entities (project leaders of each participating business entity). The plan is generated in PDF format and its integrity and immutability are ensured by digital signature (using the SM2 algorithm).

[0078] 3.5 Solution Push The central collaborative agent pushes resource scheduling plans to the task scheduling agent via the Enterprise Service Bus (ESB). The push uses a breakpoint resume mechanism (transmitting in 500KB chunks) to support resumption of transmission after interruption. Upon receiving the plan, the task scheduling agent parses and verifies it. Verification includes the rationality of resource allocation (e.g., whether the logistics provider's transportation capacity meets the task allocation requirements) and the feasibility of time nodes (e.g., whether deployment time is sufficient). If verification passes, a receipt confirmation is returned (including the plan's MD5 value and the receipt time). If verification fails, specific issues are reported (e.g., insufficient transportation capacity of a logistics provider), prompting the central collaborative agent to re-optimize the plan.

[0079] Step 4: The task scheduling agent pushes project progress nodes to the fund management agent, which then calculates the disbursement ratio and initiates fund disbursement to the third-party payment system. 4.1 Definition and Push of Progress Nodes Based on the directed acyclic graph after task decomposition, the task scheduling agent identified five core fund disbursement nodes using the Critical Path Method (CPM), as follows: Node number Fund disbursement node Node completion standard Weight Expected completion time 1 Preparation completion node All suppliers complete the first batch of inventory on demand, the qualified rate ≥ 99.5% 30% June 20, 2024 2 Cross-border transportation starting node All batches of goods complete loading / machine, logistics track can be tracked 20% June 30, 2024 3 Clear customs completion node All goods complete customs declaration, pass rate ≥ 99% 20% July 15, 2024 4 Overseas warehouse storage node All goods complete overseas warehouse storage, inventory data synchronization complete 15% July 20, 2024 5 The first batch of orders received node The first 5000 single orders complete terminal user acceptance 15% August 10, 2024 The task scheduling agent monitors the progress of each node through scheduled tasks (using the Quartz framework, executed once per hour). Once a node reaches its completion criteria, it collects progress data in real time (e.g., for a node completing inventory preparation, it collects the quantity of goods prepared, the pass rate, and the completion time), generates a "Progress Node Completion Report," and pushes it to the funds management agent via an HTTPS interface. The pushed data includes core information such as node sequence number, completion time, completion quality indicators, and corresponding subtask weights. The interface response time is ≤200ms.

[0080] 4.2 Total Capital Allocation and Capital Pool Configuration The total project funding is 50 million RMB. The fund management agent allocates the total funds to various fund disbursement nodes based on the weight of each sub-task: Inventory completion milestone: 50 million × 30% = 15 million yuan; Cross-border transportation start-up point: 50 million × 20% = 10 million yuan; Customs clearance completion milestone: 50 million × 20% = 10 million yuan; Overseas warehouse inbound transaction: 50 million × 15% = 7.5 million yuan; First batch of orders signed off: 50 million × 15% = 7.5 million yuan.

[0081] The intelligent fund management agent opens a dedicated fund pool account at a commercial bank. This account is a corporate current deposit account. The fund pool employs a segregated management model, setting up an independent sub-account for each disbursement node. Funds in these sub-accounts are used solely for fund disbursements at their corresponding nodes, preventing misappropriation. The fund pool account features real-time balance monitoring and supports SMS alerts and system push notifications for balance changes.

[0082] 4.3 Progress data collection and payment calculation The financial management agent collects sub-task progress data in real time through IoT devices (sensors deployed in supplier warehouses, logistics vehicles, and overseas warehouses) and system interfaces: Inventory preparation progress data: Collected through the supplier's ERP system interface, the inventory preparation quantity, completion rate, pass rate and other data are synchronized once per hour; Transportation progress data: GPS location data, transportation distance, and estimated arrival time are collected through the logistics management system and synchronized once every 30 minutes; Customs clearance progress data: By connecting to the customs interface through the customs declaration system, the real-time status of customs clearance review and processing progress can be obtained; Inbound progress data: The quantity of goods entering the warehouse and the allocation of storage locations are collected through the overseas warehouse WMS (Warehouse Management System) and synchronized once per hour; Delivery progress data: Order delivery status and delivery time are collected through the delivery system and synchronized in real time.

[0083] The amount to be disbursed is calculated based on a preset subtask progress-funding function. The formula is: Amount to be disbursed = Funds allocated to each node × (Actual progress / Standard progress of each node). For example, if the standard progress of the inventory preparation completion node is 100% (the first batch of inventory is fully completed), and the actual progress is 80%, then the amount to be disbursed = 15 million × 80% = 12 million.

[0084] When the balance of the sub-task fund pool is less than the amount to be disbursed (e.g., the balance of the fund pool for the preparation node is 12 million yuan, and the amount to be disbursed is 13 million yuan), the sub-tasks are ranked according to their impact on the overall target task (the impact is calculated using the analytic hierarchy process). The priority from high to low is as follows: preparation completion node (impact 0.35) > customs clearance completion node (0.25) > cross-border transportation start node (0.2) > overseas warehouse entry node (0.1) > first batch order receipt node (0.1). Priority is given to ensuring the partial disbursement of funds for nodes with high impact. This disbursement is 12 million yuan (the full amount of the fund pool), and the remaining 1 million yuan will be disbursed after the funds are replenished.

[0085] 4.4 Initiation of Fund Disbursement The funds management agent generates a "Fund Disbursement Instruction," which includes the following core information: Disbursement instruction ID: BF20240620001 (Format: BF + Date + Serial Number); Payment milestone: Inventory preparation completion milestone; Amount disbursed: 12 million yuan; Payee information: Corporate bank accounts of 3 suppliers (account name, account number, bank name); Allocation ratio: Based on the inventory preparation task allocation ratio (40% for Supplier A, 40% for Supplier B, and 20% for Supplier C). Postscript: Funding for inventory preparation in cross-border e-commerce projects.

[0086] The disbursement instruction is sent to a third-party payment system (such as the Industrial and Commercial Bank of China's corporate online banking system) via a direct bank-enterprise connection interface (using the ISO 20022 standard protocol). The interface employs double encryption (SSL + MAC, SSL SecureSockets Layer + MAC Message Authentication Code) to ensure transmission security. After the disbursement instruction is sent, the funds management agent starts a status query timer, checking the disbursement status every 30 seconds until a clear result is obtained.

[0087] Step 5: The third-party payment system returns a fund disbursement receipt to the fund management intelligent agent. 5.1 Fund Disbursement Processing After receiving a fund disbursement instruction, the third-party payment system processes it according to the following procedure: Instruction verification: Verify the digital signature of the disbursement instruction (SM2 algorithm), the validity of the payee's account, and whether the funds pool balance is sufficient; Fund transfer: 12 million yuan was transferred to the corporate accounts of three suppliers according to the disbursement ratio. Supplier A received 4.8 million yuan, Supplier B received 4.8 million yuan, and Supplier C received 2.4 million yuan. Receipt Generation: After the fund transfer is completed, a standardized fund disbursement receipt is generated, which includes the following core fields: Receipt ID: HZ20240620001 (Format: HZ + Date + Serial Number); Corresponding disbursement instruction ID: BF20240620001; Payment status: Successful; Actual disbursement amount: 12 million yuan; Receipt information: Receipt time (accurate to the second) and transaction number for each supplier; Transaction fee: 9600 yuan (calculated at 0.08%); Receipt generated at 14:35:22 on June 20, 2024.

[0088] The receipt is generated in XML format and returned to the fund management intelligent agent through the bank-enterprise direct connection interface. The transmission process uses an encrypted channel to ensure data security.

[0089] 5.2 Receipt Receipt Receipt and Verification After receiving the payment receipt, the funds management agent performs the following verification operations: Integrity verification: Check if all fields on the receipt are complete (12 required fields in total). If none are missing, the verification passes. Consistency verification: Compare the payment instruction ID and payment amount in the receipt with the information at the time of initiation. If they match, the verification passes. Signature verification: Verify the digital signature of the third-party payment system (using the official public key of ICBC). If the signature is successful, the source of the receipt is confirmed to be legitimate.

[0090] Upon successful verification, the fund management agent stores the receipt information in the fund disbursement log table (MySQL database) and updates the fund pool balance (the balance of the sub-account of the inventory preparation node is updated from 15 million yuan to 3 million yuan). If verification fails (e.g., missing fields, signature failure), a receipt retransmission request is triggered, with a maximum of 3 retransmissions and a 1-minute interval between retransmissions. If the verification still fails, an exception alarm is generated and pushed to the project administrator terminal.

[0091] Step 6: The fund management agent synchronizes the fund disbursement status to the central collaborative agent and pushes progress and fund data to the revenue distribution agent. 6.1 Fund disbursement status synchronized The funds management agent synchronizes the fund disbursement status to the central collaborative agent via an internal message bus (Kafka version 3.6). The synchronization message contains the following core information: Message ID: XT20240620001 (Format: XT + Date + Serial Number); Payment milestone: Inventory preparation completion milestone; Payment status: Successful; Actual disbursement amount: 12 million yuan; Remaining balance in the capital pool: 3 million yuan; Synchronization time: June 20, 2024, 14:36:05.

[0092] The message bus is configured with a message persistence mechanism to ensure that messages are not lost. After receiving the message, the central collaborative intelligent agent updates the project's global status table, marks the funding status of the inventory preparation node as "partially disbursed," and synchronizes it to the project management backend for managers to view in real time.

[0093] 6.2 Progress and Funding Data Push The fund management agent integrates the following data and pushes it to the profit distribution agent: Progress data: The actual progress of the inventory preparation node is 80%, the pass rate is 99.7%, and the completion time is June 20, 2024; Funding data: 15 million yuan was split into nodes, 12 million yuan was actually disbursed, 9,600 yuan was charged in handling fees, and 3 million yuan remained in the fund pool; Data on contributions from participating businesses: the quantity of goods in stock from the three suppliers (1600 units from Supplier A, 1600 units from Supplier B, and 800 units from Supplier C), and the quality of the goods in stock (99.8% pass rate from Supplier A, 99.7% from Supplier B, and 99.6% from Supplier C).

[0094] Data is transmitted in Parquet format (compression rate is 50% higher than JSON) and pushed through an HTTPS encrypted channel. After the push is completed, a data transmission confirmation is generated, recording information such as data size (1.2MB), transmission time (200ms), and verification result (CRC32 value matching).

[0095] 6.3 Optimization of Ecological Evolutionary Intelligent Agent Process The ecological evolutionary intelligent agent collects process data in the commercial system in real time, including the collaborative response time of each intelligent agent (the response time of the central collaborative intelligent agent is 250ms), task execution efficiency (the efficiency of the inventory preparation task is 5% higher than expected), and fund disbursement timeliness (the time from node completion to fund disbursement is 45 minutes).

[0096] Generate process optimization strategy data through data analysis: Ecosystem building intelligent agent optimization: Adjust the frequency of commercial source data collection, changing industry data collection from once a day to once every 12 hours to improve the real-time performance of trend prediction; Task scheduling agent optimization: The task decomposition algorithm is optimized, and the atomic units of the inventory preparation task are further refined into three sub-units: "raw material procurement", "production and processing" and "quality inspection", thereby improving the accuracy of task management.

[0097] The ecological evolution agent synchronizes optimization strategy data back to the ecological building agent and the task scheduling agent via a RESTful API. After receiving the data, the two agents complete the rule iteration update within 24 hours and generate an optimization log to record the optimization content, effective time and other information.

[0098] Step 7: After calculating the phased revenue, the revenue distribution agent triggers the distribution to the third-party payment system. 7.1 Contribution Measurement Model Accounting The revenue-sharing agent constructs a contribution index model based on four dimensions: capital investment, task quality, technical support, and ecosystem recommendation. The weights and calculation methods for each dimension are as follows: Contribution dimension Weight Accounting method Capital investment 30% According to the proportion of the preparation fund of the participating commercial body, the capital investment score = (the preparation fund of the enterprise / the total preparation fund) × 100 × 30% Task quality 40% According to the preparation qualified rate and completion time, the task quality score = (qualified rate × 0.6 + time efficiency standard rate × 0.4) × 100 × 40% Technical support 20% According to whether the intelligent preparation system is deployed, 20 points are given for deployment, and 0 points are given for non-deployment Ecological recommendation 10% According to the number of recommended cooperation parties meeting the requirements, 5 points are given for recommending 1, and a full score of 10 points is given A dimensional comprehensive quantification algorithm is used to map the data of each dimension index to the [0,100] interval for normalization. The contribution score of the participating business entity is calculated through a weighted summation model. Supplier A: Capital input score 24 points (30%×80%), task quality score 39.92 points (40%×(99.8%×0.6+100%×0.4)), technical support score 20 points, ecosystem recommendation score 5 points, total contribution score 88.92 points; Supplier B: Capital investment score 24 points, task quality score 39.88 points (40%×(99.7%×0.6+100%×0.4)), technical support score 20 points, ecosystem recommendation score 0 points, total contribution score 83.88 points; Supplier C: Funding input score 12 points, task quality score 39.84 points (40%×(99.6%×0.6+100%×0.4)), technical support score 20 points, ecosystem recommendation score 5 points, total contribution score 76.84 points.

[0099] 7.2 Calculation of Periodic Returns The project's phased revenue (stock preparation stage) is 5 million yuan, which will be distributed according to the contribution score. Profit distribution ratio = Contribution score of a certain enterprise / Sum of contribution scores of all enterprises; Supplier A's allocation ratio = 88.92 / (88.92+83.88+76.84) = 36.7%, allocation amount = 5 million × 36.7% = 1.835 million yuan; Supplier B's allocation ratio = 83.88 / 249.64 = 33.6%, allocation amount = 5 million × 33.6% = 1.68 million; Supplier C's allocation ratio = 76.84 / 249.64 = 30.7%, and the allocation amount = 5 million × 30.7% = 1.535 million yuan.

[0100] The revenue distribution smart agent predefines a structured smart contract template, which includes core terms such as revenue distribution ratio, payment time, and objection handling mechanism. The contribution measurement results and distribution rules are embedded into the smart contract to generate the "Phase-based Revenue Distribution Smart Contract". The contract is written in Solidity and deployed on the Ethereum test network (Sepolia). The contract address is: 0x789dF36Ae8B3d88556f990aB34d9909e96645f.

[0101] 7.3 Profit Distribution Trigger The revenue distribution agent calls the third-party payment system interface through a smart contract to trigger the revenue distribution process. The distribution instruction contains the following core information: Assignment instruction ID: SF20240625001 (format: SF + date + serial number); Smart contract address: 0x789dF36Ae8B3d88556f990aB34d9909e96645f; Allocation recipients: Corporate accounts of 3 suppliers; Allocation amounts: Supplier A: RMB 1,835,000, Supplier B: RMB 1,680,000, Supplier C: RMB 1,535,000; Payment method: Bank transfer to corporate account, with a processing time of ≤2 business days.

[0102] After receiving the instruction, the third-party payment system executes the fund transfer based on the immutability of smart contracts. Upon completion, it returns a revenue distribution receipt. The revenue distribution smart agent synchronizes the distribution results to all smart agents and participating businesses through system message pushes, email notifications, and SMS reminders to ensure that all participants are aware of the distribution results in a timely manner.

[0103] The system devices and module embodiments described in the above examples are merely illustrative. The units described as separate components may or may not be physically separate. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent-based business data processing method, running in a business system, comprising: The intelligent agent for building a business ecosystem collects and preprocesses business source data, builds a business ecosystem prediction model, and generates business ecosystem construction plan data based on the business source data, thereby selecting participating business entities accordingly. The central collaborative intelligent agent distributes the data of the business ecosystem construction plan and the information of participating business entities to the AI-enabled intelligent agent, the task scheduling intelligent agent, and the fund management intelligent agent, triggering the start of the intelligent agent collaborative process; AI-enabled intelligent agents use data from business ecosystem construction solutions to call up scenario-based model libraries to generate task difficulty prediction and human skill matching data, and output them to the task scheduling intelligent agent. The task scheduling agent combines task difficulty prediction and human skill matching data to construct a human skill profile of the participating business entity, and completes task decomposition, pushing sub-tasks to the matching terminals of the participating business entity. The fund management agent performs fund management at the progress nodes pushed by the task scheduling agent, dynamically matching the task progress with the required funds. After the task is completed, the revenue distribution agent integrates the financial data of the fund management agent and the task quality data of the task scheduling agent, calculates the revenue through the contribution measurement model, triggers the smart contract to execute the distribution, and synchronizes the distribution results to each agent and participating business entity.

2. The method according to claim 1, characterized in that, The intelligent agent for building the business ecosystem collects and preprocesses business source data to build a business ecosystem prediction model, including: Collect commercial source data including industry data from business data platforms, production and sales data from enterprise resource planning systems, and trend data from industry databases; The commercial source data is preprocessed by using a hybrid strategy of forward imputation and mean completion to handle missing data and remove outliers. A long short-term memory network trend prediction model is constructed based on the tensor flow machine learning framework, and the preprocessed commercial source data is input to generate an ecosystem construction scheme.

3. The method according to claim 1, characterized in that, The corresponding selected participating businesses include: Based on the participating business entities' qualifications and / or historical business cooperation records, as well as their production capacity and / or service capabilities, a weighted scoring method is used to calculate the scores of the participating business entities, and those with scores exceeding the first threshold are selected.

4. The method according to claim 1, characterized in that, The AI-enabled intelligent agent, based on data from a business ecosystem construction plan, utilizes a scenario-based model library to generate task difficulty prediction and human skill matching data, including: Based on industry tags and / or business task attribute data in the business ecosystem construction scheme data, a rule engine is used to match the corresponding business application scenario model in the business scenario-based artificial intelligence model library. Based on the business application scenario model, the data volume, number of process nodes, number of cross-agent collaborations, urgency, task delivery deadline and current remaining cycle of the business task are quantified to generate business task difficulty prediction data. Based on the business application scenario model, acquire skill level adaptation data and technical resource demand list data for participating business entities as human resource skill matching data.

5. The method according to claim 1, characterized in that, The task scheduling agent, combining task difficulty prediction and human skill matching data, constructs a human skill profile for participating businesses, including: It receives task difficulty coefficients and core technical requirements from task difficulty prediction data, as well as skill tags and skill adaptation suggestions from human skill matching data. It converts unstructured text data into numerical features through word vector mapping and maps discrete skill level data to a fixed interval through normalization processing. Based on knowledge graphs, human skill dimensions are divided, information is filtered from the numerical features and / or the data in the fixed intervals to construct a skill tree, and the weights of the features in the skill tree are dynamically adjusted according to task requirements; By combining at least one quantitative indicator from skill mastery, task execution proficiency, and cross-scenario adaptability, a skill score is calculated using a weighted summation algorithm and mapped to the human skill level of the participating business entity.

6. The method according to claim 5, characterized in that, Further includes: When receiving at least one of the following: task execution feedback data, skill improvement proof data, or task difficulty prediction update data, the human skill profile of the participating business entity is updated. The updated data is then updated with corresponding features and model training updates. Set an accuracy evaluation index for the human skill profile of participating businesses. When the accuracy evaluation index is lower than the threshold, adjust the feature extraction rules and quantification model parameters to achieve optimized processing of the human skill profile of participating businesses.

7. The method according to claim 1, characterized in that, The task breakdown includes: The original task's process structure is parsed using a business process model and a notation method, and then transformed into a directed acyclic graph containing nodes, task steps, nodes, and dependency edges. Strongly connected components in a directed acyclic graph are extracted using graph theory algorithms, and indivisible atomic task units are identified as the smallest granular task units for decomposition.

8. The method according to claim 1, characterized in that, The fund management at the execution progress nodes, which dynamically matches task progress with required funds, includes: Based on the dependencies in the directed acyclic graph of subtasks, the critical path method is used to identify fund disbursement nodes, and the total funds are distributed to each task node according to the weight of the subtasks. Collect subtask progress data in real time and generate subtask progress values; The amount payable for a subtask is calculated based on a preset subtask progress-funds function. When the balance of the sub-task fund pool is less than the amount to be disbursed, a portion of the funds will be disbursed based on the sub-task's impact on the overall target task.

9. The method according to claim 1, characterized in that, Further includes: The ecological evolution agent analyzes process data in the business system, generates process optimization strategy data, and back-synchronizes it to the ecological building agent and / or task scheduling agent to iterate the rules of the ecological building agent and / or task scheduling agent.

10. The method according to claim 1, characterized in that, The revenue calculation through the contribution metric model includes: A contribution index model is constructed based on the dimensions of capital investment, task quality, technical support, and ecosystem recommendation. A dimensional comprehensive quantification algorithm is adopted to normalize the mapping of indicator data of each dimension, and the contribution score of participating business entities is calculated through a weighted summation model. A predefined structured smart contract template is used to combine the contribution measurement results with the allocation rules to generate the revenue data of the business entity.

Images