An operating-scene-oriented agent cooperative response and task execution method and system, and a storage medium

Abstract

The present application relates to the technical field of multi-agent system, and more particularly to a business scenario-oriented agent collaborative response and task execution method and system and storage medium. The context label of the current page is collected, the user question is routed to the corresponding scene sub-agent according to the page type, the scene sub-agent matches the target Skill in the pre-registered Skill set, the execution sequence corresponding to the target Skill is generated based on the Skill dependency graph recording the front-end dependency relationship between the Skills, the business system interface is called in turn and the result is written into the session state, and the answer is assembled and returned to the user. The present application can solve or at least alleviate the problem of disconnection of response and omission of steps caused by lack of runtime arrangement of composite task execution of the existing AI customer service without awareness of page context, and provides a business scenario-oriented agent collaborative response and task execution method and system and storage medium.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system technology, and in particular to a method, system and storage medium for intelligent agent collaborative response and task execution in business scenarios. Background Technology

[0002] As enterprise-level AI applications evolve from question-and-answer assistants to executable intelligent agent platforms, AI customer service for merchant business scenarios has become a common access module in business processes such as live streaming, e-commerce, product details, order details, and mobile apps. Existing AI customer service systems often use a unified dialogue entry point and a unified response engine, failing to differentiate between the specific business pages users are on. They use the same processing path for different scenarios such as inquiring about products in a live stream, browsing and guiding products in an e-commerce store, asking about applicability on product details pages, and checking logistics on order details pages. This results in a lack of correspondence between response strategies and business scenarios, and complex tasks involving multiple business interfaces are prone to omissions or missing parameters during execution.

[0003] To address page-specific differences, some solutions overlay page type tags onto a unified dialogue engine, configuring different welcome messages, keyword templates, or FAQ subsets for different pages. Regarding complex task execution, some solutions use predefined fixed task templates, calling business system interfaces sequentially according to the template order. While these two solutions improve individual capabilities, they fail to resolve the interplay between them: page tags only apply to greetings and keywords, failing to reach the intent recognition and business execution stages. Existing technologies often fail to address "intent jumps" in user interactions when fixed task templates are used; furthermore, when user questions span multiple intents or intent combinations are outside the template's preset range, they cannot dynamically generate execution sequences and lack the ability to achieve "on-demand assembly" through dynamic reverse orchestration.

[0004] From a theoretical perspective, user questions in business scenarios have two essential characteristics. Firstly, there's page relevance: the set of questions users care about differs significantly across pages. For example, users in a live stream focus on products and activities, while those on order details pages focus on fulfillment and after-sales service. Without considering page context, intent recognition can easily lead to misidentification when the intent space is too large. Secondly, there's the dependency between skill processing units: a business operation often requires first performing a pre-query to obtain parameters before executing subsequent business actions; creating a work order requires an order number, which is output by an order query; explaining a refund policy requires the current order status, also output by an order query. Existing solutions separate dialogue from business execution, failing to simultaneously utilize page context to converge the intent space and failing to orchestrate the order of business interface calls according to skill dependencies at runtime. Summary of the Invention

[0005] To achieve the above-mentioned objectives, this invention provides a method, system, and storage medium for intelligent agent collaborative response and task execution in business scenarios. It aims to solve or at least alleviate the problem that existing AI customer service lacks awareness of the user's current page, resulting in a disconnect between response strategies and business scenarios; and to address the problem that complex tasks involving multiple business interfaces lack runtime orchestration, leading to execution failures or missing steps.

[0006] To address the shortcomings of the existing technologies, the core design concept of this invention is as follows: The front-end collects page context tags and uploads the page type and business parameters to the back-end; the back-end routes user queries to the corresponding scene sub-agents based on the page type, and each scene sub-agent pre-registers its own set of Skills, thus converging the search space for intent recognition; within the scene sub-agents, reverse backtracking and topological sorting are performed based on the Skill dependency graph that records the pre-dependencies between Skills, dynamically reversing the arrangement according to user requests and generating the execution sequence corresponding to the target Skill to achieve "on-demand assembly." Combined with a session state cross-page continuation and failure compensation mechanism, this solves the linkage problem between page context convergence and composite task arrangement.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent agent collaborative response and task execution in business scenarios, the method comprising: Collect the context tags of the current page, which include the page type and page business parameters; Based on the page type, the user's question is routed to the corresponding scene sub-agent. Each scene sub-agent pre-registers a set of skill processing units (Skills) encapsulated by multiple atomic business interfaces. The scene sub-agent performs intent recognition on the user's question and searches for a matching target Skill in the Skill set; Based on the target Skill and its prerequisite dependencies, an execution sequence corresponding to the target Skill is generated from a pre-configured Skill dependency graph that records the prerequisite dependencies between each Skill. The business system interfaces are called sequentially according to the execution sequence, and the results are written to the session state. The final execution result is assembled into a response and returned to the user.

[0008] To further realize the present invention, the following technical solutions may be preferred: Preferably, the page types include live streaming room, mall homepage, product details page, order details page, and App assistant section; the scene sub-intelligent agents include live streaming scene sub-intelligent agents, mall shopping guide sub-intelligent agents, product consultant sub-intelligent agents, after-sales service sub-intelligent agents, and comprehensive service sub-intelligent agents corresponding to the page types.

[0009] Preferably, the Skill dependency graph includes multiple Skill nodes and directed dependency edges between nodes. Each Skill node records input parameters, output results, and permission requirements. The directed dependency edges indicate that the preceding Skill points to the following Skill.

[0010] Preferably, generating the execution sequence includes: taking the target Skill as the endpoint node, backtracking all preceding Skill nodes in the Skill dependency graph to form a Skill subgraph, with the backtracking terminating at a Skill node that no longer has incoming edges; and performing a topological sort on the Skill subgraph to obtain the execution sequence.

[0011] Preferably, when calling the business system interface according to the execution sequence, if the input parameters required by the current Skill have an unexpired cached result in the session state, the business system interface call of the Skill is skipped, and the cached result is used as the output of the Skill; the unexpired result is determined by whether the remaining lifetime of the cached result is greater than zero.

[0012] Preferably, the method further includes a step of continuing the session state across pages: using the user ID and the session window ID as the primary key to save the historical session state; when the user jumps from the original page to the new page, the scene sub-agent corresponding to the new page reads the historical session state as context supplement when it obtains session control to respond to the new page request.

[0013] Preferably, the method further includes a Skill execution failure compensation step: each Skill pre-registers a corresponding compensation action in the Skill dependency graph; if any Skill fails to execute, the compensation actions bound to it are triggered sequentially in reverse order of the successfully executed Skills according to the execution path determined by the Skill dependency graph, so as to roll back the business state and return the reason for failure to the user.

[0014] Preferably, if the scene sub-agent does not match the target Skill in the Skill set, it enters the knowledge question answering rollback process: it searches for the answer in the knowledge base subset pre-configured by the scene sub-agent; if the confidence of the search result is lower than a set threshold, it triggers manual processing.

[0015] A collaborative response and task execution system for intelligent agents in business scenarios, the system comprising: a page context acquisition module for acquiring and uploading context tags of the current page; an intelligent agent routing module for routing user queries to corresponding scenario sub-intelligent agents based on the page type; multiple scenario sub-intelligent agents, each of which pre-registers a set of Skills and performs intent recognition and Skill matching on user queries; and a Skill dependency graph management module for maintaining Skill nodes and dependencies, and generating the execution corresponding to the target Skill from a pre-configured Skill dependency graph that records the prerequisite dependencies between Skills, based on the matched target Skill. The system includes: a line sequence; a Skill execution engine for calling business system interfaces according to the execution sequence; a session state management module for saving and continuing session state using user ID and session window ID as primary keys; and a data analysis dashboard module for aggregating session events generated by sub-intelligent agents in various scenarios and the Skill execution engine, calculating and outputting total inquiries, inquiries for each page dimension, AI automatic replies, AI resolution rate, conversion rate to human intervention, inquiries conversion rate on product details page, after-sales interception rate on order page, top 10 hot questions, top 10 unresolved questions, hot keywords for questions in the live broadcast room, number of leads left, number of appointments, and order conversion related data, providing visualized operational data support for merchant managers.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0017] The beneficial effects of this invention are: This invention drives agent routing through page context tags, converging user questions to corresponding scenario sub-agents for processing, thus reducing the intent recognition search space and improving intent recognition accuracy and AI automatic response rate. It generates execution sequences through Skill dependency graph backtracking and topological sorting, organizing business interface calls according to actual runtime needs, reducing Skill execution failure rate and parameter missing rate. It preserves historical session context when users browse across pages by extending session state across pages, reducing repeated questioning. Finally, it improves state consistency and response coverage during complex task execution through Skill failure compensation and fallback to knowledge-based question answering when intents do not match. Attached Figure Description

[0018] Figure 1 This is the system architecture diagram of the present invention.

[0019] Figure 2 This is the overall flowchart of the method of the present invention.

[0020] Figure 3 This is an example diagram of the Skill dependency graph structure of the present invention.

[0021] Figure 4 This is a schematic diagram of cross-page session continuation according to the present invention.

[0022] Figure 5 This is a flowchart of the Skill execution failure compensation process of the present invention. Detailed Implementation

[0023] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0025] This embodiment provides an intelligent agent collaborative response and task execution system for business scenarios. The system is deployed in a multi-tenant environment of a beauty brand SaaS service provider, and its architecture is as follows: Figure 1 As shown. The front-end SDK is embedded in the H5 pages of the live streaming room, online store, product details page, and order details page as JavaScript scripts, as well as the assistant-level section embedded in the mobile app. The back-end Agent service cluster is deployed on a private cloud and includes an intelligent agent routing module, 5 scene sub-intelligent agents, a Skill dependency graph management module, a Skill execution engine, a session state management module, a knowledge Q&A rollback module, and a data analysis dashboard module. The business system access layer connects to seven existing business systems—live streaming, online store, product, order, marketing, SCRM, and customer service ticketing—via an API gateway.

[0026] After the page loads, the front-end SDK reads the pre-embedded `data-nuo-page` property in the DOM to obtain the page type, and reads business parameters from URL parameters or the `window` variable. For example, the live stream page reads the live stream ID, the e-commerce homepage reads the current e-commerce ID and category ID, the product details page reads the product ID, and the order details page reads the order ID. These parameters, the page type, the user ID in the user's login state, and the current session window ID together constitute the context object.

[0027] After receiving the context object, the agent routing module performs a route lookup. The routing table uses page type as the key and the service address of the scene's sub-agent as the value. The five scene sub-agents operate independently, and each sub-agent loads its own skill list from the Skill registry upon startup. The live streaming scene sub-agent loads five skills: knowledge Q&A, product Q&A, event guidance, lead generation, and referral to human agent; the e-commerce guide sub-agent loads five skills: knowledge Q&A, product recommendation, event guidance, and lead generation; the product consultant sub-agent loads six skills: product Q&A, product recommendation, event guidance, lead generation, order inquiry, and referral to human agent; the after-sales service sub-agent loads four skills: order inquiry, after-sales consultation, work order creation, and referral to human agent; and the comprehensive service sub-agent loads all skills.

[0028] When multiple sub-agents in a scenario have overlapping skills, the system standardizes the dependency through a global skill dependency graph and combines it with the global lock or state synchronization mechanism of the session state management module (such as a distributed lock based on a Redis cluster) to ensure strong consistency of business logic and state flow when cross-agent collaboration and concurrent calls to the same overlapping skill, thus avoiding data read and write conflicts.

[0029] The Skill dependency graph management module internally stores a directed graph, the structure of which is as follows: Figure 3 As shown, nodes represent Skills, and edges represent prerequisites. Typical examples of dependency edges are: the Order Query Skill points to the Ticket Creation Skill, which requires the Order Number field, provided by the Order Query Skill's output; the Product Q&A Skill points to the Product Recommendation Skill, which requires the current product category as input. Users may also inquire about historical order status on the product details page to determine whether to repurchase; therefore, the Product Advisor sub-agent pre-registers the Order Query Skill as an auxiliary capability. The dependency graph is stored in the configuration library in JSON format, allowing operations personnel to extend it according to industry templates. When adding a Skill, only the node and incoming edges need to be appended to the dependency graph configuration; no modification to the agent code is required.

[0030] The Skill execution engine schedules Skills sequentially according to their execution sequence. Each Skill is encapsulated as an independent executable unit, taking a parameter dictionary as input and outputting a result dictionary. Fields in the result dictionary are then written to the session state. The session state uses a Redis cluster as its storage backend, with the primary key being the concatenation of the user ID and the session window ID, and a time-to-live (TTL) of 30 minutes.

[0031] The knowledge-based question-and-answer rollback module integrates a vector retrieval service, with each scenario's sub-agent pre-configured with a subset of the knowledge base. The knowledge base subset for the after-sales service sub-agent includes brand company introduction, after-sales rules, and contact information; the knowledge base subset for the product consultant sub-agent includes product usage instructions, product ingredient information, and product pairing suggestions.

[0032] The data analysis dashboard module receives session events generated by the sub-agents and Skill execution engine in various scenarios via asynchronous event streams. These events include: session start and end, user question reporting (including page type tags), Skill matching results, Skill execution results, triggering of human intervention, completion of lead generation registration, and triggering of knowledge Q&A rollback. The dashboard module aggregates and calculates these event streams, outputting the following four categories of statistical indicators: Traffic: Total number of inquiries, number of inquiries per page dimension (live stream / store homepage / product details page / order details page / App assistant); Quality: AI automatic reply volume, AI resolution rate (= number of sessions where AI completed the response but did not trigger human intervention / total number of sessions), human intervention rate; Conversion: product details page inquiry conversion rate, order page after-sales interception rate, number of leads, number of appointments; Content: Top 10 hot questions, Top 10 unresolved questions (clustering of questions that triggered knowledge Q&A rollback and had a confidence level below the threshold), and hot keywords in the live stream. The above metrics are written into the time-series database, with merchant tenant ID as the isolation dimension, and can be queried by daily / weekly / monthly dimensions. Merchant owners, operations managers, and customer service supervisors can access them through independent dashboard pages. Example 2

[0033] Based on the system architecture in Embodiment 1, this embodiment provides a method for intelligent agent collaborative response and task execution in business scenarios. The overall process of this method is as follows: Figure 2 As shown below, a complete user session will be used as an example for illustration.

[0034] When browsing a face mask on a beauty brand's product details page, a user clicked on the floating customer service icon on the page and sent a question: "Is this suitable for sensitive skin? Are there any discounts?"

[0035] The front-end SDK collects data on a page type of "product details page", determines the product ID, and then encapsulates the page type, product ID, user ID, and session window ID into a context object for uploading.

[0036] The agent routing module looks up the table and finds the product consultant sub-agent, forwarding the user's question and context object. Upon receiving the request, the product consultant sub-agent identifies the intent. This question involves two intents: one is a product suitability inquiry, and the other is a discount inquiry. The sub-agent matches the product Q&A skill and the activity guidance skill from its registered skill list.

[0037] The Skill dependency graph management module backtracks dependency nodes based on two target Skills. The Product Q&A Skill requires basic product information as input, which is obtained through "Product Information Query" as the preceding node. The Activity Guidance Skill requires on-sale activity information corresponding to the current product ID as input, with "Activity Information Query" as the preceding node. Backtracking terminates at a leaf node that no longer has incoming edges, resulting in the Skill subgraph: {Product Information Query, Activity Information Query, Product Q&A, Activity Guidance}. The topological sorting result is: Product Information Query → Product Q&A, Activity Information Query → Activity Guidance, two parallel paths.

[0038] The Skill execution engine calls the business systems sequentially based on the topology sorting results. For product information queries, Skill sends a request to the e-commerce system API, passing in the product ID, and returns fields such as product name, target audience, ingredients, and efficacy. For activity information queries, Skill sends a request to the marketing system API, passing in the product ID, and returns a list of activities the product is currently participating in and the discount rules. For product Q&A, Skill matches user concerns based on the target audience field in the product information, generating a response fragment such as "This product is suitable for people with sensitive skin, and the Centella Asiatica extract in the ingredients has soothing properties." For activity guidance, Skill generates information such as "This product participates in a new customer discount of 15 yuan, and a brand coupon of 30 yuan off for purchases over 200 yuan can also be used."

[0039] The two Skill response fragments are merged into a complete reply by the product consultant sub-agent's response assembler and returned to the user: "This face mask is suitable for sensitive skin, and the Centella Asiatica extract in the ingredients has soothing properties. Currently, the product is participating in a new customer discount of 15 yuan, and you can also claim a brand coupon for 30 yuan off purchases over 200 yuan." After the reply is sent, the product consultant sub-agent simultaneously reports this conversation event to the data analysis dashboard module: page type is "product details page," the Skills involved are "product Q&A" and "activity guidance," the conversation result is "AI automatic reply," and no human intervention was triggered. The dashboard module includes this event in the total number of inquiries for the day and updates the frequency count of the top 10 most popular questions. If the user subsequently clicks to place an order, the front-end SDK reports the order event, and the dashboard module associates it with this consultation conversation, adding it to the product details page consultation conversion rate numerator.

[0040] The user then clicks "Place Order" and is redirected to the order details page. The cross-page session continuation process is as follows: Figure 4As shown. The front-end SDK re-collects the page type "Order Details Page", determines the order ID, and uploads it along with the session window ID. The intelligent agent routing module forwards the request to the after-sales service sub-intelligent agent. The after-sales service sub-intelligent agent reads the historical sessions in the session state management module, using the user ID + session window ID as the primary key, and learns that the user previously inquired about applicability and discounts on the product details page, which can be continued in the context of the order page.

[0041] A user asks on the order details page, "When will my order arrive? How do I get a refund if there's a problem?" The after-sales service sub-agent identifies the intent as order logistics tracking and refund process consultation. Matching the order query skill and the after-sales consultation skill, dependency graph backtracking reveals that the order query skill requires the order number (already in the context), while the after-sales consultation skill requires the order status (output by the order query skill). Topological sorting yields: Order Query → After-Sales Consultation.

[0042] The Skill execution engine is invoked sequentially. For order inquiries, Skill returns the order status as "Paid, pending shipment, estimated shipment time is 10:00 AM the next day"; for after-sales inquiries, Skill returns a refund explanation based on the order status "pending shipment": "For orders pending shipment, you can click the 'Apply for Refund' button on the order details page. The funds will be returned to the original payment method, and the arrival time will be 1-3 business days." The after-sales service sub-agent combines the responses and returns them to the user.

[0043] During the above call process, if the output result of the order query skill already exists in the session state and its remaining time to life is greater than zero, the skill execution engine will skip the interface call of the order query skill and directly reuse the cached result in the session state as the output of the order query skill, thereby reducing the latency and business system pressure caused by repeated interface calls.

[0044] When the order query skill returns an order status of "delivery error," and the system business rules determine that this error status requires manual or after-sales processing, the after-sales service sub-AI will use the "work order creation skill" as the new endpoint target node, and re-initiate an independent dependency graph backtracking to ask the user if they need to create an after-sales work order. If the user confirms, and the work order creation skill fails to execute (e.g., the customer service work order system interface times out), the pre-registered compensation action will be rolled back. The skill execution failure compensation process is as follows: Figure 5As shown. The compensation actions are called in reverse order of the executed Skills: First, the compensation action registered with the ticket creation Skill is called to clear the ticket draft flag already written to the customer service ticket system; second, the compensation action registered with the after-sales consultation Skill is called to clear the consultation flag already written to the session status; finally, the user is returned with "Ticket creation has failed temporarily. Your consultation information has been saved, and a human customer service representative will contact you within 30 minutes," and this failure is recorded in the retry queue.

[0045] If a user's question falls under the category of a skill not pre-registered by the scenario sub-agent, such as asking "Where is your company located?" on an order details page, and the after-sales service sub-agent's intent recognition does not match any skill, a knowledge-based question-and-answer rollback is initiated. The rollback module searches a subset of the knowledge base pre-configured by the after-sales service sub-agent. If the search confidence score is higher than 0.7, an answer is returned; otherwise, a human agent referral skill is triggered, generating a conversation summary and pushing it to the human agent system. The confidence threshold of 0.7 is an empirical value obtained before project launch based on historical manually labeled data and adjusted during trial operation. Its value is determined by ensuring that the sum of the mis-referral rate and the missed referral rate at this threshold reaches a low level during the trial operation.

[0046] In a live streaming scenario, the agent routing module routes user questions to the live streaming scenario sub-agent. Frequently asked questions in live streaming scenarios include product suitability, activity rules, discount information, and order placement methods. When the live streaming scenario sub-agent searches the dependency graph, questions like "What is the price of this product?" are matched with the product Q&A skill and the product information query front-end, with a path similar to the product details page; questions like "What benefits are there in the live stream?" are matched with the activity guidance skill and the activity information query front-end. In a live streaming scenario, many users may ask similar questions at the same time. The skill execution engine reuses the result of the first execution for the same question under the same live stream ID within a short time window, avoiding impact on the business system interfaces.

[0047] Once the lead generation skill of the live streaming scenario sub-agent or the e-commerce guide sub-agent successfully collects and writes the phone number, WeChat ID, or appointment information into the SCRM system, it simultaneously reports the lead generation completion event to the data analysis dashboard module and includes it in the daily lead generation statistics. For those who further complete appointment confirmation after lead generation, the appointment skill will report the appointment count event separately after completion. The two metrics are displayed independently in the dashboard to help merchants evaluate the potential customer acquisition efficiency of each page scenario.

[0048] In the App Assistant section, the intelligent agent routing module routes user questions to the comprehensive service sub-intelligent agent. This sub-intelligent agent dynamically loads the authorized Skill set based on the tenant configuration, covering core Skills such as knowledge Q&A, product recommendation, order inquiry, after-sales consultation, activity guidance, and transfer to human agent by default. It can handle various questions covering products, orders, activities, and after-sales, and is suitable for scenarios where users actively initiate inquiries through the App's primary entry point.

[0049] Example 3

[0050] This embodiment provides a storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the method in Embodiment 2.

[0051] The technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0052] Example 4 Verification A 30-day comparative test was conducted in the real production environment of the beauty brand's business. The control group used a generic unified AI customer service system without this invention deployed, while the experimental group used the system of this invention. Both groups of businesses were of similar size, with an average monthly consultation volume ranging from 100,000 to 130,000. The user profile distribution, product category distribution, and promotional schedule distribution were also similar between the control and experimental groups.

[0053] The core metrics collected are as follows: AI automatic response rate: 84.6% in the experimental group and 71.2% in the control group, an increase of 13.4 percentage points in the experimental group; Conversion to human agent: 14.8% in the experimental group and 27.5% in the control group, a decrease of 12.7 percentage points in the experimental group; Session abandonment and abnormal interruption rate: 0.6% in the experimental group and 1.3% in the control group, a decrease of 0.7 percentage points in the experimental group; Intent recognition accuracy: 89.2% in the experimental group and 78.5% in the control group, an increase of 10.7 percentage points in the experimental group; Inquiry conversion rate on the product details page: 5.8% in the experimental group and 4.2% in the control group, an increase of 1.6 percentage points in the experimental group; Percentage of automatic generation of after-sales work orders on the order details page: 63.5% in the experimental group, no such capability deployed in the control group; Skill execution failure rate: 3.7% in the experimental group and 14.3% in the control group, a decrease of 10.6 percentage points in the experimental group; Cross-page session continuation success rate: 91.3% in the experimental group; Average response time: 2.14 seconds in the experimental group and 1.82 seconds in the control group, an increase of 0.32 seconds in the response time of the experimental group.

[0054] During the aforementioned 30-day test, the data analytics dashboard module performed statistical analysis on all conversations in the experimental group. By analyzing the top 10 most frequently asked questions, operations staff identified two high-frequency unresolved issues: "real-time inventory query" and "promotion stacking rules." Five days after the test ended, corresponding skills were added, increasing the AI ​​resolution rate from 0% to 68%. The top 10 unresolved issues also identified "cross-brand package deals" as a long-tail demand exceeding the scope of the skills and marked it as a direction for expansion. Based on the lead generation metric, the merchant operations manager assessed that the experimental group's daily lead generation in the livestream room increased by approximately 23% compared to the control group. Through consultation volume across various page dimensions, the customer service manager found that consultations on the product details page accounted for 41% of all consultations, leading to optimization of human customer service scheduling. These operational feedbacks validated the practical value of the data analytics dashboard module in skill iteration and human resource optimization.

[0055] The experimental group experienced a slight increase in response time, due to the additional overhead of dependency graph backtracking and execution sequence generation, averaging 0.32 seconds. However, in subsets involving complex tasks (such as order exception handling), the control group required an average of 3-5 rounds of dialogue to complete the task, while the experimental group completed it within 1-2 rounds, resulting in an overall reduction in session time of approximately 58%.

[0056] The decrease in Skill execution failure rate primarily stemmed from the validation of pre-defined parameters in the dependency graph. In the control group, approximately 10.6% of failures were due to missing parameters or unmet pre-defined conditions, such as the work order creation interface being called without an order number, or the refund policy description being returned as generic text without order verification. The experimental group eliminated these failures by backtracking through the dependency graph to pre-complete parameters. The remaining 3.7% of failures mainly originated from timeouts or business rule rejections within the business system interface itself, which are outside the scope of this invention.

[0057] The cross-page session continuation success rate was 91.3%. The main reason for not reaching 100% was that some users' session intervals exceeded the 30-minute lifespan threshold set by the system's Redis cache (approximately 6.1%). Additionally, a small number of users switched login devices across platforms, and unpredictable long-tail engineering anomalies such as occasional network packet loss and microservice communication interruptions contributed to the continuation. In scenarios where the same user continuously navigates within the same session window, the continuation success rate approached 100%.

[0058] The improvement in intent recognition accuracy comes from two aspects. First, the page context narrows the intent space from all intents to scene-related intents, reducing the classification space of the intent recognition model from dozens of categories to 5-10 categories for each scene. Second, the verification of the required parameters of the preceding nodes during the Skill dependency graph backtracking process can identify some false positive intents that are superficially matched but lack necessary parameters.

[0059] For questions that could not be answered successfully, this embodiment compiled a list of the top 10 unresolved issues, including real-time product inventory queries, specific activity combination rules, and cross-brand package deals. A common characteristic of these issues is that they fall outside the coverage of currently registered skills. These issues can be incorporated into the system's processing capabilities by adding corresponding skills and updating the dependency graph.

[0060] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent agent collaborative response and task execution in business scenarios, characterized in that, The method includes: Collect the context tags of the current page, which include the page type and page business parameters; Based on the page type, the user's question is routed to the corresponding scene sub-agent, and each scene sub-agent pre-registers a set of Skills; The scene sub-agent performs intent recognition on the user's question and searches for a matching target Skill in the Skill set; Based on the target Skill and its prerequisite dependencies, an execution sequence corresponding to the target Skill is generated from a pre-configured Skill dependency graph that records the prerequisite dependencies between each Skill. The business system interfaces are called sequentially according to the execution sequence, and the results are written to the session state. The final execution result is assembled into a response and returned to the user. The Skill dependency graph includes multiple Skill nodes and directed dependency edges between nodes. Each Skill node records input parameters, output results, and permission requirements. The directed dependency edges indicate that the preceding Skill points to the following Skill. Generating the execution sequence includes: taking the target Skill as the endpoint node, backtracking backward through all preceding Skill nodes in the Skill dependency graph to form a Skill subgraph, with the backtracking terminating at a Skill node that no longer has incoming edges; and performing a topological sort on the Skill subgraph to obtain the execution sequence. When calling the business system interface according to the execution sequence, if the input parameters required by the current Skill have an unexpired cached result in the session state, the business system interface call of the Skill is skipped, and the cached result is used as the output of the Skill; the unexpired result is determined by whether the remaining lifetime of the cached result is greater than zero.

2. The method according to claim 1, characterized in that, The page types include live streaming room, mall homepage, product details page, order details page, and App assistant section; the scene sub-intelligent agents include live streaming scene sub-intelligent agents, mall shopping guide sub-intelligent agents, product consultant sub-intelligent agents, after-sales service sub-intelligent agents, and comprehensive service sub-intelligent agents corresponding to the page types.

3. The method according to claim 1, characterized in that, The method also includes a step for continuing the session state across pages: the user ID and the session window ID are used together as the primary key to save the historical session state; when the user jumps from the original page to the new page, the scene sub-agent corresponding to the new page reads the historical session state as context supplement when it obtains session control to respond to the new page request.

4. The method according to claim 1, characterized in that, The method also includes a Skill execution failure compensation step: each Skill pre-registers a corresponding compensation action in the Skill dependency graph; if any Skill fails to execute, the compensation actions bound to it are triggered sequentially in reverse order of the successfully executed Skills according to the execution path determined by the Skill dependency graph, so as to roll back the business state and return the reason for failure to the user.

5. The method according to claim 1, characterized in that, If the scene sub-agent does not match the target Skill in the Skill set, it enters the knowledge question answering rollback process: the answer is retrieved in the knowledge base subset pre-configured by the scene sub-agent; If the confidence level of the search results is lower than the set threshold, manual processing will be triggered.

6. A smart agent collaborative response and task execution system for business scenarios, the system being used to implement the method according to any one of claims 1 to 5, characterized in that, The system includes: a page context acquisition module for acquiring and uploading context tags of the current page; an agent routing module for routing user queries to corresponding scene sub-agents based on page type; multiple scene sub-agents, each of which pre-registers a set of Skills and performs intent recognition and Skill matching on user queries; a Skill dependency graph management module for maintaining Skill nodes and dependencies, and generating an execution sequence corresponding to the target Skill from a pre-configured Skill dependency graph that records the prerequisite dependencies between Skills, based on the matched target Skill; a Skill execution engine for calling business system interfaces according to the execution sequence; a session state management module for saving and continuing session state using user ID and session window ID as primary keys; and a data analysis dashboard module for aggregating session events generated by various scene sub-agents and the Skill execution engine, providing visualized business data support for merchant managers.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 5.

Images