A multi-agent self-evolution system and method based on closed-loop evaluation feedback

Abstract

This invention discloses a multi-agent self-evolution system and method based on closed-loop evaluation feedback, relating to the fields of artificial intelligence, distributed system architecture, and knowledge engineering. The system adopts a layered and decoupled architecture, including a user interaction layer, a semantic core and basic support layer, an intent understanding layer, a task planning layer, a task execution layer, an evaluation and supervision layer, and a closed-loop evolution mechanism. Each module interacts through a standardized data protocol. A collaborative digital twin mechanism and a semantic core knowledge system are constructed. An independent evaluation and supervision agent cluster quantitatively evaluates the entire lifecycle of tasks, and the evaluation results are transformed into machine-readable structured instruction objects, driving the update of the underlying knowledge base and the parameter tuning of agent capabilities, thereby achieving a closed-loop system learning and evolution. This invention improves the success rate and collaborative efficiency of complex tasks, enabling the system to autonomously accumulate domain wisdom and adapt to dynamic environmental changes during long-term operation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, distributed system architecture and knowledge engineering, and in particular to a multi-agent self-evolution system and method based on closed-loop evaluation feedback. Background Technology

[0002] With the explosive growth of Large Language Model (LLM) technology, LLM-based agents have demonstrated powerful natural language understanding and reasoning capabilities. From early single-conversation assistants to today's autonomous agent systems capable of calling upon tools and decomposing tasks, the penetration rate of AI in enterprise operations continues to increase. However, despite significant progress in task automation made by existing multi-agent frameworks such as AutoGPT, LangChain, and MetaGPT, three major technical challenges remain when building truly intelligent operational collaboration systems:

[0003] First, the semantic gap and information silos: Modern enterprise IT architectures are complex, with inherent semantic barriers existing between business systems such as ERP, CRM, and PLM, as well as various heterogeneous databases (relational, graph, time-series). Inconsistent data models and a lack of metadata management make it difficult for AI agents to accurately understand the business meaning of data. For example, the "customer" in CRM may have completely different attribute definitions and business rules than the "customer" in the financial system. Existing multi-agent systems often lack a unified, machine-understandable "semantic core," leading to frequent misunderstandings and even "illusions" when agents collaborate across systems, preventing the formation of globally consistent business insights.

[0004] Second, knowledge fragmentation and static nature: Enterprises accumulate a wealth of core knowledge over long-term operations, including expert experience, implicit rules, and historical cases. This knowledge often exists in the form of unstructured documents (PDF, Word) or scattered database records, forming "knowledge silos." While current RAG (Retrieval Enhanced Generation) technology can solve some information retrieval problems, it lacks a mechanism for deep internalization of knowledge. Knowledge bases are typically static and cannot be automatically updated as business rules change or as agents learn lessons during operation. This leads to the problem of "delivery equals solidification," where the system's knowledge base gradually becomes outdated and unable to adapt to the dynamically changing business environment over time.

[0005] Third, the system lacks closed-loop evaluation and self-evolution capabilities: Current mainstream multi-agent frameworks primarily focus on task planning and execution, severely neglecting evaluation and evolution. For example, evaluation is often delayed and subjective, relying heavily on manual post-event reviews to assess task execution quality, lacking automated, data-driven quantitative evaluation standards. Alternatively, the feedback loop may be broken; even if errors are discovered during operation (such as improper tool parameter configuration or misjudgment of a rule), the system lacks standardized mechanisms to transform these negative feedbacks into corrective instructions. Unlike human teams, the system cannot summarize experiences through post-event reviews, automatically optimize collaboration processes, or update the knowledge base. This prevents the system from achieving compound growth in capabilities over the long term.

[0006] In conclusion, the industry urgently needs a new type of multi-agent system architecture that can break down data and knowledge silos and achieve an automated cycle from "execution" to "evaluation" and then to "evolution" through an endogenous closed-loop evaluation mechanism, thereby building an intelligent life form that can continuously learn and self-optimize. Summary of the Invention

[0007] The purpose of this invention is to overcome the defects existing in the above-mentioned background technology. This invention proposes a multi-agent self-evolution system and method based on closed-loop evaluation feedback. It takes semantic core as the cognitive foundation, collaborative digital twin as the operating carrier, and closed-loop evaluation and instruction-driven evolution as the core driving force to construct a hierarchical decoupled and autonomously evolving intelligent operation collaboration body.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A multi-agent self-evolutionary system based on closed-loop evaluation feedback, the system being as follows:

[0010] The system adopts a layered and decoupled architecture, which includes a user interaction layer, a semantic core and basic support layer, an intent understanding layer, a task planning layer, a task execution layer, an evaluation and supervision layer, and a closed-loop evolution mechanism. Each module interacts through a standardized data protocol.

[0011] The user interaction layer, as the sensory outpost of the system, is responsible for receiving multimodal input from users and displaying the real-time progress, intermediate results and final output of the task to users; this layer integrates an Agent operation monitoring dashboard, which allows users to view the agent's thinking process and resource consumption.

[0012] The semantic core and basic support layer serve as the system's long-term memory and cognitive foundation. Based on an ontology-based dynamic knowledge graph network, it stores and manages user feature bases, evaluation standard knowledge bases, domain-specific rules, task execution templates, agent collaboration modes, error correction strategies, and tool capability descriptions. It supports real-time entity linking and semantic reasoning, providing a unified business context for upper-layer components. The semantic core and basic support layer are also configured with maintenance agents, serving as specific computing power units for performing knowledge updates and parameter tuning.

[0013] The intent understanding layer is configured with a context-aware intent recognition agent cluster. The intent recognition agent cluster contains multiple professional agents with semantic parsing, slot extraction, entity linking and intent classification capabilities. It receives user natural language input and, combined with user profiles and historical interaction records in the semantic core and basic support layers, transforms unstructured requirements into structured task objects linked to semantic core entities.

[0014] The task planning layer is configured with a dynamic planning agent cluster. The dynamic planning agent cluster contains multiple planning agents with the capabilities of task decomposition, dependency analysis, resource scheduling and logic verification. Based on the structured task object, a collaborative digital twin model is instantiated to generate a DAG planning that includes multiple sub-task steps, data dependencies and execution roles. It has the ability to dynamically replan based on real-time feedback. The DAG is a directed acyclic graph.

[0015] The task execution layer is configured with a distributed heterogeneous execution intelligent agent cluster. The distributed heterogeneous execution intelligent agent cluster includes multiple specialized intelligent agents with specific atomic capabilities and general large model intelligent agents. According to the scheduling of the task planning layer, it calls external tools or internal algorithms to execute specific tasks and generates execution trajectory data of the entire life cycle of the task.

[0016] The evaluation and supervision layer is configured with an evaluation agent cluster independent of the execution link. The evaluation agent cluster includes multiple evaluation agents with the capabilities of rule compliance checking, intent alignment analysis, multi-dimensional quality scoring and instruction conversion. It asynchronously monitors the entire life cycle of the task, generates multi-dimensional evaluation reports based on preset standards and runtime data, and converts the identified optimization points into machine-readable structured instruction objects.

[0017] The closed-loop evolution mechanism serves as the system's self-updating engine. It receives and parses structured instructions, and drives the maintenance agent to perform operations such as adding, deleting, and modifying the content stored in the semantic core and basic support layer, as well as adjusting the agent's prompt word parameters or cooperation strategies, thereby achieving the system's endogenous evolution.

[0018] Furthermore, the semantic core and basic support layer consist of seven interconnected dynamic knowledge bases, specifically: a user feature knowledge base, a domain knowledge knowledge base, a task template knowledge base, an agent collaboration knowledge base, an error and fix knowledge base, a tool capability knowledge base, and an evaluation standard knowledge base, as detailed below:

[0019] The user feature knowledge base stores users' static attributes, dynamic preferences, historical interaction habits, and domain awareness levels.

[0020] The domain-specific knowledge base stores business concepts, logical rules, laws and regulations, and industry standards for specific vertical industries;

[0021] The task template knowledge base stores verified standardized operating procedures, task decomposition templates, and best practice paths;

[0022] The Agent Collaboration Knowledge Base records the collaboration efficiency, communication modes, and success rate statistics of different agent combinations under specific task types.

[0023] The error and repair knowledge base stores historical anomaly patterns, root cause analysis reports, and corresponding automatic repair strategies.

[0024] The tool capability knowledge base provides semantic definitions of the functions, parameter constraints, calling costs, and permissions of the system-integrated APIs, algorithm models, and RPA scripts.

[0025] The evaluation criteria knowledge base stores acceptance indicators, quality scoring dimensions, test cases, and gold standard samples for various tasks.

[0026] Furthermore, the dynamic knowledge graph network in the semantic core and basic support layer is constructed based on ontology and possesses the following characteristics:

[0027] The knowledge graph network maps entities such as users, tasks, tools, rules, and error patterns to graph nodes, and maps the relationships between entities to graph edges, forming a computable ontology network.

[0028] The knowledge graph supports semantic association reasoning between entities across knowledge bases, and can automatically identify the logical relationships between entities in different knowledge bases, providing logical verification support for task planning;

[0029] The node and edge attributes of the knowledge graph support dynamic updates. Update operations include adding entities, modifying the strength of associations, and marking the validity status of entities. All update operations are associated with corresponding instruction IDs to ensure that knowledge evolution is traceable.

[0030] Furthermore, the collaborative digital twin model is a dynamic digital mapping of the multi-agent collaborative process, including a dynamic state machine, a visual planning graph, and a human-machine collaboration interface, as detailed below:

[0031] The dynamic state machine maintains the current state of the task, the intermediate outputs of completed steps, the queue of tasks to be processed, and the load state of each participating agent in real time.

[0032] The visualized planning map presents the task execution path and dependencies in a graphical way, allowing users to intuitively view the system's decision-making logic;

[0033] The human-machine collaboration interface allows users to intervene in the twin model at any stage of task execution, including pausing the task, correcting the planned path, adding constraints, or modifying intermediate results. The system captures the intervention behavior and transmits it to the evaluation and supervision layer as a high-weight negative feedback signal.

[0034] Furthermore, the evaluation agent cluster in the evaluation supervision layer includes three specialized sub-evaluation agent groups: the system specification and design evaluation group, the interaction and alignment evaluation group, and the result verification and strategy evaluation group, as detailed below:

[0035] The system specification and design evaluation group is used in the planning phase before task execution to conduct a priori evaluation of the collaborative digital twin model, and to check the rationality of the planning logic, the risk of dead loops, and the expected resource consumption.

[0036] The interaction and alignment evaluation group is used to monitor the agent's communication logs in real time during task execution, and to evaluate the accuracy of information transmission, the smoothness of the collaboration process, and the intention alignment deviation.

[0037] The result verification and strategy evaluation group is used to perform a post-hoc evaluation of the final output after the task is completed. By comparing the user feature library preference settings or the gold sample in the evaluation standard knowledge base, the accuracy, completeness and user satisfaction of the results are calculated.

[0038] Furthermore, the dynamic replanning capability of the task planning layer satisfies the following execution logic:

[0039] The triggering conditions for dynamic replanning include user intervention signals, abnormal error reports from the executing agent, and planning deviation prompts provided in real time by the evaluation and supervision layer;

[0040] During the replanning process, the system only reconstructs the affected subtask nodes and dependencies in the collaborative digital twin model, retaining the results of successfully executed nodes to avoid redundant calculations.

[0041] The new DAG execution graph generated after replanning needs to be quickly verified by the system specifications and design evaluation group. The verification dimensions include logical consistency, incremental resource consumption, and matching degree with user intent. Execution can only be started after the verification is passed.

[0042] Furthermore, the structured instructions in the closed-loop evolution mechanism include instruction ID, instruction type, target object, change carrier, triggering reason, and confidence score, as detailed below:

[0043] The instruction ID is used to uniquely identify the corresponding structured instruction;

[0044] The instruction type indicates the nature of the operation, including updating knowledge, creating rules, discarding templates, or adjusting parameters;

[0045] The target object specifies the specific knowledge base or agent ID for the structured instruction to be applied.

[0046] The change carrier includes specific modifications, such as new confidence scores, revised rule text, or optimized prompt word templates;

[0047] The trigger reason record is the evaluation basis or user feedback summary of the trigger structured instruction, used for system interpretability traceability;

[0048] The confidence score represents the degree of certainty the supervisory layer has regarding the structured instruction, and is used to determine whether it should be executed automatically or requires manual approval.

[0049] A multi-agent self-evolution method based on closed-loop evaluation feedback includes the following steps:

[0050] Step S1, Multi-source data perception and intent modeling: The multi-agent self-evolution system receives multimodal input from users. The intent understanding layer combines user features and domain knowledge from the semantic core and basic support layer to eliminate ambiguity and construct structured task instance objects.

[0051] Step S2, Collaborative Twin Construction and Dynamic Planning: Based on the task instance object, the task planning layer retrieves the best task template, instantiates the collaborative digital twin model, generates an initial execution plan diagram, and displays it to the user for pre-review.

[0052] Step S3, Distributed Collaborative Execution and Trajectory Recording: The task execution layer schedules the corresponding intelligent agent cluster to execute sub-tasks according to the plan. During the execution process, the system records the thinking process, tool call parameters, intermediate outputs and communication logs of all intelligent agents in real time, forming an execution trajectory log for the entire life cycle.

[0053] Step S4, Dynamic Human-Machine Alignment and Replanning: During execution, if the user issues an intervention signal through the human-machine collaboration interface, the system immediately suspends the current execution. The intent understanding layer parses the user's new intent, and the task planning layer performs dynamic replanning based on the current twin state and the new intent to generate a corrected execution path.

[0054] Step S5, Asynchronous Closed-Loop Evaluation and Instruction Generation: After the task is completed, the evaluation and supervision layer asynchronously reads the execution trajectory data, performs multi-dimensional scoring against the evaluation standard knowledge base, identifies defects or highlights in the execution, and generates structured optimization instructions.

[0055] Step S6, Knowledge Internalization and System Evolution: The structured instructions generated in step S5 are parsed, and the maintenance agent in the semantic core and basic support layer receives and executes the instructions, updates the corresponding knowledge base entries in the semantic core, or adjusts the configuration parameters of the agent, and completes a self-evolution iteration of the system.

[0056] Furthermore, step S3 includes a secure and controllable invocation mechanism, specifically:

[0057] Before the corresponding intelligent agent attempts to call external tools or APIs, the system performs permission verification based on the tool capability knowledge base in the semantic core and basic support layer.

[0058] For operations marked as high-risk or irreversible, the system will forcibly trigger a human-machine collaborative confirmation process, and the operation can only be executed after obtaining explicit authorization from the user.

[0059] Furthermore, the execution trajectory log generated in step S3 has the following specifications:

[0060] The execution trajectory log includes the agent's thought chain, tool call parameters and return results, complete data of intermediate outputs, execution time and resource consumption data of each stage;

[0061] The execution trajectory log is stored in a standardized format, retains complete timestamps, ensures the temporal correspondence of data in each stage, and supports log backtracking and reproduction of the task execution process;

[0062] The execution trajectory log is asynchronously pushed to the evaluation and supervision layer. The pushed content is the full log data, and the log data is retained in the system for no less than the most recent 100 system evolution iteration cycles, providing data support for multiple rounds of evaluation and evolution analysis.

[0063] Compared with the prior art, the present invention, employing the above technical solution, has the following beneficial effects:

[0064] (1) The present invention proposes a multi-agent self-evolution system and method based on closed-loop evaluation feedback. The system adopts a layered decoupling architecture and standardized data protocol. Each module is independently encapsulated and has standardized interaction. This not only reduces the coupling between modules and facilitates individual upgrades and the addition of new functions, but also flexibly adapts to diverse business scenarios such as intelligent software development, medical compliance review, and government opinion analysis, which greatly improves the maintainability and horizontal scalability of the system.

[0065] (2) The present invention proposes a multi-agent self-evolution system and method based on closed-loop evaluation feedback. Through seven interconnected dynamic knowledge bases and ontological knowledge graphs, the system structurally associates scattered user features, domain rules and other knowledge, effectively breaking down semantic gaps and knowledge silos. At the same time, the knowledge graph supports dynamic updates and is traceable, ensuring that knowledge can be continuously optimized with business changes and user feedback, avoiding "delivery is fixed", and always adapting to the dynamic business environment.

[0066] (3) The present invention proposes a multi-agent self-evolution system and method based on closed-loop evaluation feedback. The collaborative digital twin model realizes the transparency (visualized decision logic) and interventionability (human-machine collaborative adjustment at any stage) of task execution. Combined with the dynamic replanning mechanism and the security mechanism of "authority verification + high-risk human-machine confirmation", it not only improves the fault tolerance and efficiency of task execution, but also ensures the security and controllability of the execution process, and ensures the accurate alignment of human and machine intentions.

[0067] (4) The present invention proposes a multi-agent self-evolution system and method based on closed-loop evaluation feedback. The evaluation and supervision layer accurately locates execution problems through three-dimensional quantitative evaluation before, during and after the event. The closed-loop evolution mechanism automatically converts the evaluation results into structured instructions, drives knowledge base updates and agent parameter optimization, and forms an automated closed loop of "evaluation-feedback-evolution". It does not require frequent manual modification of code or rules, realizes the compound growth of system capabilities, and significantly reduces long-term maintenance costs and manual dependence.

[0068] (5) The present invention proposes a multi-agent self-evolution system and method based on closed-loop evaluation feedback, which constructs a complete "perception-decision-execution-evaluation-evolution" link. It not only solves the core problems of semantic alignment and process smoothness in multi-agent collaboration in complex tasks, but also can quickly adapt to dynamic changes such as industry standard updates through self-evolution capabilities, greatly improving the success rate of task execution in complex business scenarios and expanding the application boundaries of the system. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the overall six-layer architecture and data flow of the system of the present invention;

[0070] Figure 2 This is a schematic diagram of the lifecycle state machine of the collaborative digital twin model of the present invention;

[0071] Figure 3 This is a flowchart of the evaluation and supervision layer of the present invention generating instruction objects and driving knowledge base updates;

[0072] Figure 4 This is a schematic diagram of the ontology structure of the dynamic knowledge graph according to an embodiment of the present invention. Detailed Implementation

[0073] 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.

[0074] A multi-agent self-evolutionary system based on closed-loop evaluation feedback, such as Figure 1 As shown, the system is as follows:

[0075] The system adopts a layered and decoupled architecture, which includes a user interaction layer, a semantic core and basic support layer, an intent understanding layer, a task planning layer, a task execution layer, an evaluation and supervision layer, and a closed-loop evolution mechanism. Each module interacts through a standardized data protocol.

[0076] The user interaction layer, as the sensory outpost of the system, is responsible for receiving multimodal input from users and displaying the real-time progress, intermediate results and final output of the task to users; this layer integrates an Agent operation monitoring dashboard, which allows users to view the agent's thinking process and resource consumption.

[0077] The semantic core and basic support layer serve as the system's long-term memory and cognitive foundation. Based on an ontology-based dynamic knowledge graph network, it stores and manages user feature bases, evaluation standard knowledge bases, domain-specific rules, task execution templates, agent collaboration modes, error correction strategies, and tool capability descriptions. It supports real-time entity linking and semantic reasoning, providing a unified business context for upper-layer components. The semantic core and basic support layer are also configured with maintenance agents, serving as specific computing power units for performing knowledge updates and parameter tuning.

[0078] The intent understanding layer is configured with a context-aware intent recognition agent cluster. The intent recognition agent cluster contains multiple professional agents with semantic parsing, slot extraction, entity linking and intent classification capabilities. It receives user natural language input and, combined with user profiles and historical interaction records in the semantic core and basic support layers, transforms unstructured requirements into structured task objects linked to semantic core entities.

[0079] The task planning layer is configured with a dynamic planning agent cluster. The dynamic planning agent cluster contains multiple planning agents with the capabilities of task decomposition, dependency analysis, resource scheduling and logic verification. Based on the structured task object, a collaborative digital twin model is instantiated to generate a DAG planning that includes multiple sub-task steps, data dependencies and execution roles. It has the ability to dynamically replan based on real-time feedback. The DAG is a directed acyclic graph.

[0080] The task execution layer is configured with a distributed heterogeneous execution intelligent agent cluster. The distributed heterogeneous execution intelligent agent cluster includes multiple specialized intelligent agents with specific atomic capabilities and general large model intelligent agents. According to the scheduling of the task planning layer, it calls external tools or internal algorithms to execute specific tasks and generates execution trajectory data of the entire life cycle of the task.

[0081] The evaluation and supervision layer is configured with an evaluation agent cluster independent of the execution link. The evaluation agent cluster includes multiple evaluation agents with the capabilities of rule compliance checking, intent alignment analysis, multi-dimensional quality scoring and instruction conversion. It asynchronously monitors the entire life cycle of the task, generates multi-dimensional evaluation reports based on preset standards and runtime data, and converts the identified optimization points into machine-readable structured instruction objects.

[0082] The closed-loop evolution mechanism serves as the system's self-updating engine. It receives and parses structured instructions, and drives the maintenance agent to perform operations such as adding, deleting, and modifying the content stored in the semantic core and basic support layer, as well as adjusting the agent's prompt word parameters or cooperation strategies, thereby achieving the system's endogenous evolution.

[0083] Furthermore, the semantic core and basic support layer consist of seven interconnected dynamic knowledge bases, specifically: a user feature knowledge base, a domain knowledge knowledge base, a task template knowledge base, an agent collaboration knowledge base, an error and fix knowledge base, a tool capability knowledge base, and an evaluation standard knowledge base, as detailed below:

[0084] The user feature knowledge base stores users' static attributes, dynamic preferences, historical interaction habits, and domain awareness levels.

[0085] The domain-specific knowledge base stores business concepts, logical rules, laws and regulations, and industry standards for specific vertical industries;

[0086] The task template knowledge base stores verified standardized operating procedures, task decomposition templates, and best practice paths;

[0087] The Agent Collaboration Knowledge Base records the collaboration efficiency, communication modes, and success rate statistics of different agent combinations under specific task types.

[0088] The error and repair knowledge base stores historical anomaly patterns, root cause analysis reports, and corresponding automatic repair strategies.

[0089] The tool capability knowledge base provides semantic definitions of the functions, parameter constraints, calling costs, and permissions of the system-integrated APIs, algorithm models, and RPA scripts.

[0090] The evaluation criteria knowledge base stores acceptance indicators, quality scoring dimensions, test cases, and gold standard samples for various tasks.

[0091] Furthermore, the dynamic knowledge graph network in the semantic core and basic support layer is constructed based on ontology and possesses the following characteristics:

[0092] The knowledge graph network maps entities such as users, tasks, tools, rules, and error patterns to graph nodes, and maps the relationships between entities to graph edges, forming a computable ontology network.

[0093] The knowledge graph supports semantic association reasoning between entities across knowledge bases, and can automatically identify the logical relationships between entities in different knowledge bases, providing logical verification support for task planning;

[0094] The node and edge attributes of the knowledge graph support dynamic updates. Update operations include adding entities, modifying the strength of associations, and marking the validity status of entities. All update operations are associated with corresponding instruction IDs to ensure that knowledge evolution is traceable.

[0095] Furthermore, the collaborative digital twin model is a dynamic digital mapping of the multi-agent collaborative process, including a dynamic state machine, a visual planning graph, and a human-machine collaboration interface, as detailed below:

[0096] The dynamic state machine maintains the current state of the task, the intermediate outputs of completed steps, the queue of tasks to be processed, and the load state of each participating agent in real time.

[0097] The visualized planning map presents the task execution path and dependencies in a graphical way, allowing users to intuitively view the system's decision-making logic;

[0098] The human-machine collaboration interface allows users to intervene in the twin model at any stage of task execution, including pausing the task, correcting the planned path, adding constraints, or modifying intermediate results. The system captures the intervention behavior and transmits it to the evaluation and supervision layer as a high-weight negative feedback signal.

[0099] Furthermore, the evaluation agent cluster in the evaluation supervision layer includes three specialized sub-evaluation agent groups: the system specification and design evaluation group, the interaction and alignment evaluation group, and the result verification and strategy evaluation group, as detailed below:

[0100] The system specification and design evaluation group is used in the planning phase before task execution to conduct a priori evaluation of the collaborative digital twin model, and to check the rationality of the planning logic, the risk of dead loops, and the expected resource consumption.

[0101] The interaction and alignment evaluation group is used to monitor the agent's communication logs in real time during task execution, and to evaluate the accuracy of information transmission, the smoothness of the collaboration process, and the intention alignment deviation.

[0102] The result verification and strategy evaluation group is used to perform a post-hoc evaluation of the final output after the task is completed. By comparing the user feature library preference settings or the gold sample in the evaluation standard knowledge base, the accuracy, completeness and user satisfaction of the results are calculated.

[0103] Furthermore, the dynamic replanning capability of the task planning layer satisfies the following execution logic:

[0104] The triggering conditions for dynamic replanning include user intervention signals, abnormal error reports from the executing agent, and planning deviation prompts provided in real time by the evaluation and supervision layer;

[0105] During the replanning process, the system only reconstructs the affected subtask nodes and dependencies in the collaborative digital twin model, retaining the results of successfully executed nodes to avoid redundant calculations.

[0106] The new DAG execution graph generated after replanning needs to be quickly verified by the system specifications and design evaluation group. The verification dimensions include logical consistency, incremental resource consumption, and matching degree with user intent. Execution can only be started after the verification is passed.

[0107] Furthermore, the structured instructions in the closed-loop evolution mechanism include instruction ID, instruction type, target object, change carrier, triggering reason, and confidence score, as detailed below:

[0108] The instruction ID is used to uniquely identify the corresponding structured instruction;

[0109] The instruction type indicates the nature of the operation, including updating knowledge, creating rules, discarding templates, or adjusting parameters;

[0110] The target object specifies the specific knowledge base or agent ID for the structured instruction to be applied.

[0111] The change carrier includes specific modifications, such as new confidence scores, revised rule text, or optimized prompt word templates;

[0112] The trigger reason record is the evaluation basis or user feedback summary of the trigger structured instruction, used for system interpretability traceability;

[0113] The confidence score represents the degree of certainty the supervisory layer has regarding the structured instruction, and is used to determine whether it should be executed automatically or requires manual approval.

[0114] A multi-agent self-evolution method based on closed-loop evaluation feedback includes the following steps:

[0115] Step S1, Multi-source data perception and intent modeling: The multi-agent self-evolution system receives multimodal input from users. The intent understanding layer combines user features and domain knowledge from the semantic core and basic support layer to eliminate ambiguity and construct structured task instance objects.

[0116] Step S2, Collaborative Twin Construction and Dynamic Programming: (e.g.) Figure 2 As shown, the task planning layer retrieves the best task template based on the task instance object, instantiates the collaborative digital twin model, generates an initial execution plan diagram, and displays it to the user for pre-review.

[0117] Step S3, Distributed Collaborative Execution and Trajectory Recording: The task execution layer schedules the corresponding intelligent agent cluster to execute sub-tasks according to the plan. During the execution process, the system records the thinking process, tool call parameters, intermediate outputs and communication logs of all intelligent agents in real time, forming an execution trajectory log for the entire life cycle.

[0118] Step S4, Dynamic Human-Machine Alignment and Replanning: During execution, if the user issues an intervention signal through the human-machine collaboration interface, the system immediately suspends the current execution. The intent understanding layer parses the user's new intent, and the task planning layer performs dynamic replanning based on the current twin state and the new intent to generate a corrected execution path.

[0119] Step S5, Asynchronous Closed-Loop Evaluation and Instruction Generation: After the task is completed, such as... Figure 3 As shown, the evaluation and supervision layer asynchronously reads the execution trajectory data, performs multi-dimensional scoring against the evaluation standard knowledge base, identifies defects or highlights in the execution, and generates structured optimization instructions;

[0120] Step S6, Knowledge Internalization and System Evolution: The structured instructions generated in step S5 are parsed, and the maintenance agent in the semantic core and basic support layer receives and executes the instructions, updates the corresponding knowledge base entries in the semantic core, or adjusts the configuration parameters of the agent, and completes a self-evolution iteration of the system.

[0121] Furthermore, step S3 includes a secure and controllable invocation mechanism, specifically:

[0122] Before the corresponding intelligent agent attempts to call external tools or APIs, the system performs permission verification based on the tool capability knowledge base in the semantic core and basic support layer.

[0123] For operations marked as high-risk or irreversible, the system will forcibly trigger a human-machine collaborative confirmation process, and the operation can only be executed after obtaining explicit authorization from the user.

[0124] Furthermore, the execution trajectory log generated in step S3 has the following specifications:

[0125] The execution trajectory log includes the agent's thought chain, tool call parameters and return results, complete data of intermediate outputs, execution time and resource consumption data of each stage;

[0126] The execution trajectory log is stored in a standardized format, retains complete timestamps, ensures the temporal correspondence of data in each stage, and supports log backtracking and reproduction of the task execution process;

[0127] The execution trajectory log is asynchronously pushed to the evaluation and supervision layer. The pushed content is the full log data, and the log data is retained in the system for no less than the most recent 100 system evolution iteration cycles, providing data support for multiple rounds of evaluation and evolution analysis.

[0128] To illustrate the working principle of this invention more intuitively, the implementation process of three typical scenarios is shown below with specific code snippets.

[0129] Example 1: The Self-Evolution of the Intelligent Medical Insurance Compliance Review System

[0130] Scenario: A hospital uses this system for automatic screening of medical insurance violations. In the initial system configuration, a certain rule (ID: 9527) imposes overly strict restrictions on the use of "antiplatelet drugs," and its dynamic knowledge graph ontology structure is as follows: Figure 4 As shown.

[0131] Step 1: Task Initiation and Intent Understanding

[0132] The head of the medical insurance department typed: "Please check the high-risk issues related to drug use throughout the hospital yesterday."

[0133] The intent understanding layer generates task instances, but in the **Step 2 (Siamese Construction)** stage, the user intervenes through the interface.

[0134] Step 2: Dynamic Human-Computer Alignment (Re-planning)

[0135] After reviewing the initial plan, the user clicks a button on the interface to request "highlight cardiology data." The system captures this intervention signal and triggers a replanning process.

[0136] Plain Text

[0137] {

[0138] "event_type": "USER_INTERVENTION",

[0139] "twin_id": "TWIN_MED_001",

[0140] "timestamp": "2023-10-26T10:05:00Z",

[0141] "intervention_content": {

[0142] "action": "ADD_CONSTRAINT",

[0143] "target_node": "Report_Generation_Step",

[0144] "parameter": {"highlight_dept": "Cardiology", "highlight_style":"Red_Bold"}

[0145] },

[0146] "system_reaction": {

[0147] "action": "REGENERATE_DAG", / / Trigger DAG graph reconstruction

[0148] "affected_nodes": ["NODE_3", "NODE_4"], / / Only reprogram the affected nodes

[0149] "state_transition": "PAUSED -> REPLANNING -> EXECUTING"

[0150] }

[0151] }

[0152] Step 3: Execution and Exception Feedback

[0153] The system executes the task, and the rule matching agent calls the knowledge base. After receiving the report, the user manually removed a large number of violation alerts generated by rule RULE_9527 regarding "aspirin" in the "cardiology" department.

[0154] Step 4: Evaluation and Knowledge Base Evolution

[0155] The evaluation layer detects this "high false alarm" phenomenon, generates evolutionary instructions, and drives the basic support layer to perform graph database update operations.

[0156] Plain Text

[0157] / / 1. Locate the target rule node

[0158] MATCH (r:Rule {rule_id: 'RULE_Medical_9527'})

[0159] / / 2. Create or merge specific medical context entities

[0160] MERGE (c:Context {name: 'Post-Surgery-Cardiology'})

[0161] / / 3. Establish an "exemption (HAS_EXCEPTION)" relationship and set an exclusive condition.

[0162] CREATE (r)-[:HAS_EXCEPTION {

[0163] Reason: 'Clinical Pathway Optimization', / / Reason: Clinical pathway optimization

[0164] confidence_decay: 0.2, / / Reduce the confidence level in this scenario

[0165] trigger_evidence: 'User_Feedback_Batch_20231026',

[0166] active: true

[0167] }]->(c)

[0168] / / 4. Update rule metadata and record version number

[0169] SET r.version = r.version + 0.1, r.last_updated = timestamp()

[0170] Result: After the map was updated, the system logically understood that "postoperative use of aspirin in cardiology" was a compliant exception, and false alarms could be avoided without hard coding.

[0171] Example 2: Optimization of the Collaboration Mode of the Electronic Component Selection Assistant (Yanfushi)

[0172] Scenario: The system needs to extract parameters from PDF specifications from different suppliers. One supplier (Supplier_X) uses a non-standard PDF format, causing the default OCR mode to frequently fail.

[0173] Steps 1 & 2: Anomaly Detection and Real-time Repair

[0174] During execution, the data extraction agent encountered an error. The system automatically searched the "Error and Fix Knowledge Base" and attempted to switch strategies.

[0175] Plain Text

[0176] {

[0177] "step_id": "STEP_EXTRACT_PARAMS",

[0178] "target_file": "Supplier_X_Spec_2024.pdf",

[0179] "attempt_1": {

[0180] "strategy": "FAST_TEXT_EXTRACTION", / / Default strategy: Fast text extraction "status": "FAILED",

[0181] "error_signature": "ERR_EMPTY_content_blocks", / / Error signature: content blocks are empty "duration": "0.5s"

[0182] },

[0183] "runtime_recovery": {

[0184] "action": "SWITCH_TOOL", / / Real-time switching tool

[0185] "new_tool": "OCR_ENHANCED_MODEL_V2", / / Switch to the enhanced version

[0186] OCR"reason_code": "MATCH_ERROR_PATTERN_09"

[0187] },

[0188] "attempt_2": {

[0189] "strategy": "OCR_ENHANCED",

[0190] "status": "SUCCESS",

[0191] "output_completeness": 0.99

[0192] }

[0193] }

[0194] Step 3: Evolution of Collaboration Mode (SOP Update)

[0195] The evaluation layer discovered that tasks for Supplier_X always went through a "failure-retry" process, resulting in wasted resources. Therefore, an instruction was generated to directly modify the task template knowledge base.

[0196] The following shows a comparison of the configurations before and after the task template evolution:

[0197] Plain Text

[0198] {

[0199] "template_id": "TMPL_SPEC_ANALYSIS",

[0200] "condition_match": "Supplier == 'Supplier_X'", / / Specialized template for this supplier "evolution_change": {

[0201] "before": {

[0202] "default_extractor": "Agent_FastText", / / Before evolution: Default uses fast extraction "retry_policy": "Linear_Backoff"

[0203] },

[0204] "after": {

[0205] "default_extractor": "Agent_OCR_Enhanced", / / After evolution: the enhanced OCR is used directly by default. "rationale": "Performance optimization based on 15 historical retrylogs."

[0206] "cost_implication": "Higher compute cost, but lower latency."

[0207] }

[0208] }

[0209] }

[0210] Result: When encountering tasks from this supplier next time, the system directly loads the evolved template, and the execution efficiency is increased by 30%.

[0211] Example 3: Revision of the Assignment Logic for Public Service Hotline (12345) Public Opinion Analysis

[0212] Scenario Background: A citizen complains about "illegal construction inside a certain community". Initially, the AI assigns it to the "Urban Management Bureau", but it is rejected and should be transferred to the "Housing Management Bureau".

[0213] Step 1: Assignment Failure and Manual Rectification

[0214] The initial system assignment is based on the keyword "illegal construction". The Urban Management Bureau rejects the assignment and marks the reason in the system: "The area within the community's red line belongs to the property management scope, transfer to the housing management department". The manual operator confirms this operation.

[0215] Step 2: Evolution of the Weight Matrix

[0216] The evaluation layer analyzes this closed-loop feedback, generates a weight adjustment instruction, and updates the intent classification weight matrix in the semantic core.

[0217] Plain Text

[0218] def apply_evolution_instruction(instruction):

[0219] """

[0220] Receive the instruction from the evaluation layer and adjust the contribution weight of specific keywords to the classification result

[0221] """

[0222] target_intent = instruction.payload['target_intent'] # "Department_Housing" (Housing Management Bureau)

[0223] negative_intent=instruction.payload['negative_intent'] # "Department_UrbanManage" (Urban Management Bureau) # Extract feature words

[0224] keywords = ["inside the community", "property management scope", "within the red line"]

[0225] for word in keywords:

[0226] # 1. Reduce the association weight of this word with the "Urban Management Bureau"

[0227] current_weight_neg = VectorDB.get_weight(word, negative_intent)

[0228] VectorDB.update_weight(word, negative_intent, current_weight_neg * 0.5)

[0229] # 2. Increase the weight of this term's association with "housing management bureau".

[0230] current_weight_pos = VectorDB.get_weight(word, target_intent)

[0231] VectorDB.update_weight(word, target_intent, current_weight_pos * 2.0 + 0.5)

[0232] # 3. Record the evolution log

[0233] SystemLog.info(f"Evolution applied: Keywords {keywords}shifted focus to {target_intent}")

[0234] # Triggering result:

[0235] # After the command is executed, the combined vector of "internal community" + "illegal construction" will be closer to the vector space of "housing management bureau".

[0236] Result: Through an error feedback, the system learned the subtle division of responsibilities for "illegal construction" under different geographical attributes (within the community vs. on the street), and achieved the internalization of domain knowledge.

[0237] 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent self-evolutionary system based on closed-loop evaluation feedback, characterized in that, The system is described in detail as follows: The system adopts a layered and decoupled architecture, which includes a user interaction layer, a semantic core and basic support layer, an intent understanding layer, a task planning layer, a task execution layer, an evaluation and supervision layer, and a closed-loop evolution mechanism. Each module interacts through a standardized data protocol. The user interaction layer, as the sensory outpost of the system, is responsible for receiving multimodal input from users and displaying the real-time progress, intermediate results and final output of the task to users; this layer integrates an Agent operation monitoring dashboard, which allows users to view the agent's thinking process and resource consumption. The semantic core and basic support layer serve as the system's long-term memory and cognitive foundation. Based on an ontology-based dynamic knowledge graph network, it stores and manages user features, evaluation standard knowledge bases, domain-specific rules, task execution templates, agent collaboration modes, error correction strategies, and tool capability descriptions. It supports real-time entity linking and semantic reasoning, providing a unified business context for upper-layer components. The semantic core and basic support layer are also configured with maintenance agents, serving as specific computing power units for performing knowledge updates and parameter tuning. The intent understanding layer is configured with a context-aware intent recognition agent cluster. The intent recognition agent cluster contains multiple professional agents with semantic parsing, slot extraction, entity linking and intent classification capabilities. It receives user natural language input and, combined with user profiles and historical interaction records in the semantic core and basic support layers, transforms unstructured requirements into structured task objects linked to semantic core entities. The task planning layer is configured with a dynamic planning agent cluster. The dynamic planning agent cluster contains multiple planning agents with the capabilities of task decomposition, dependency analysis, resource scheduling and logic verification. Based on the structured task object, a collaborative digital twin model is instantiated to generate a DAG planning that includes multiple sub-task steps, data dependencies and execution roles. It has the ability to dynamically replan based on real-time feedback. The DAG is a directed acyclic graph. The task execution layer is configured with a distributed heterogeneous execution intelligent agent cluster. The distributed heterogeneous execution intelligent agent cluster includes multiple specialized intelligent agents with specific atomic capabilities and general large model intelligent agents. According to the scheduling of the task planning layer, it calls external tools or internal algorithms to execute specific tasks and generates execution trajectory data of the entire life cycle of the task. The evaluation and supervision layer is configured with an evaluation agent cluster independent of the execution link. The evaluation agent cluster includes multiple evaluation agents with the capabilities of rule compliance checking, intent alignment analysis, multi-dimensional quality scoring and instruction conversion. It asynchronously monitors the entire life cycle of the task, generates multi-dimensional evaluation reports based on preset standards and runtime data, and converts the identified optimization points into machine-readable structured instruction objects. The closed-loop evolution mechanism serves as the system's self-updating engine. It receives and parses structured instructions, and drives the maintenance agent to perform operations such as adding, deleting, and modifying the content stored in the semantic core and basic support layer, as well as adjusting the agent's prompt word parameters or cooperation strategies, thereby achieving the system's endogenous evolution.

2. The multi-agent self-evolutionary system based on closed-loop evaluation feedback according to claim 1, characterized in that, The semantic core and basic support layer consist of seven interconnected dynamic knowledge bases: a user feature knowledge base, a domain knowledge knowledge base, a task template knowledge base, an agent collaboration knowledge base, a bug and fix knowledge base, a tool capability knowledge base, and an evaluation standard knowledge base, as detailed below: The user feature knowledge base stores users' static attributes, dynamic preferences, historical interaction habits, and domain awareness levels. The domain-specific knowledge base stores business concepts, logical rules, laws and regulations, and industry standards for specific vertical industries; The task template knowledge base stores verified standardized operating procedures, task decomposition templates, and best practice paths; The Agent Collaboration Knowledge Base records the collaboration efficiency, communication modes, and success rate statistics of different agent combinations under specific task types. The error and repair knowledge base stores historical anomaly patterns, root cause analysis reports, and corresponding automatic repair strategies. The tool capability knowledge base provides semantic definitions of the functions, parameter constraints, calling costs, and permissions of the system-integrated APIs, algorithm models, and RPA scripts. The evaluation criteria knowledge base stores acceptance indicators, quality scoring dimensions, test cases, and gold standard samples for various tasks.

3. A multi-agent self-evolutionary system based on closed-loop evaluation feedback as described in claim 2, characterized in that, The dynamic knowledge graph network in the semantic core and basic support layer is constructed based on ontology and has the following characteristics: The knowledge graph network maps entities such as users, tasks, tools, rules, and error patterns to graph nodes, and maps the relationships between entities to graph edges, forming a computable ontology network. The knowledge graph supports semantic association reasoning between entities across knowledge bases, and can automatically identify the logical relationships between entities in different knowledge bases, providing logical verification support for task planning; The node and edge attributes of the knowledge graph support dynamic updates. Update operations include adding entities, modifying the strength of associations, and marking the validity status of entities. All update operations are associated with corresponding instruction IDs to ensure that knowledge evolution is traceable.

4. A multi-agent self-evolutionary system based on closed-loop evaluation feedback as described in claim 3, characterized in that, The collaborative digital twin model is a dynamic digital mapping of the multi-agent collaborative process, including a dynamic state machine, a visual planning graph, and a human-machine collaboration interface, as detailed below: The dynamic state machine maintains the current state of the task, the intermediate outputs of completed steps, the queue of tasks to be processed, and the load state of each participating agent in real time. The visualized planning map presents the task execution path and dependencies in a graphical way, allowing users to intuitively view the system's decision-making logic; The human-machine collaboration interface allows users to intervene in the twin model at any stage of task execution, including pausing the task, correcting the planned path, adding constraints, or modifying intermediate results. The system captures the intervention behavior and transmits it to the evaluation and supervision layer as a high-weight negative feedback signal.

5. A multi-agent self-evolutionary system based on closed-loop evaluation feedback according to claim 4, characterized in that, The evaluation agent cluster in the evaluation supervision layer includes three specialized sub-evaluation agent groups: the system specification and design evaluation group, the interaction and alignment evaluation group, and the result verification and strategy evaluation group, as detailed below: The system specification and design evaluation group is used in the planning phase before task execution to conduct a priori evaluation of the collaborative digital twin model, and to check the rationality of the planning logic, the risk of dead loops, and the expected resource consumption. The interaction and alignment evaluation group is used to monitor the agent's communication logs in real time during task execution, and to evaluate the accuracy of information transmission, the smoothness of the collaboration process, and the intention alignment deviation. The result verification and strategy evaluation group is used to perform a post-hoc evaluation of the final output after the task is completed. By comparing the user feature library preference settings or the gold sample in the evaluation standard knowledge base, the accuracy, completeness and user satisfaction of the results are calculated.

6. A multi-agent self-evolutionary system based on closed-loop evaluation feedback according to claim 5, characterized in that, The dynamic replanning capability of the task planning layer satisfies the following execution logic: The triggering conditions for dynamic replanning include user intervention signals, abnormal error reports from the executing agent, and planning deviation prompts provided in real time by the evaluation and supervision layer; During the replanning process, the system only reconstructs the affected subtask nodes and dependencies in the collaborative digital twin model, retaining the results of successfully executed nodes to avoid redundant calculations. The new DAG execution graph generated after replanning needs to be quickly verified by the system specifications and design evaluation group. The verification dimensions include logical consistency, incremental resource consumption, and matching degree with user intent. Execution can only be started after the verification is passed.

7. A multi-agent self-evolutionary system based on closed-loop evaluation feedback according to claim 6, characterized in that, The structured instructions in the closed-loop evolution mechanism include instruction ID, instruction type, target object, change carrier, triggering reason, and confidence score, as detailed below: The instruction ID is used to uniquely identify the corresponding structured instruction; The instruction type indicates the nature of the operation, including updating knowledge, creating rules, discarding templates, or adjusting parameters; The target object specifies the specific knowledge base or agent ID for the structured instruction to be applied. The change carrier includes specific modifications, such as new confidence scores, revised rule text, or optimized prompt word templates; The trigger reason record is the evaluation basis or user feedback summary of the trigger structured instruction, used for system interpretability traceability; The confidence score represents the degree of certainty the supervisory layer has regarding the structured instruction, and is used to determine whether it should be executed automatically or requires manual approval.

8. A multi-agent self-evolution method based on closed-loop evaluation feedback, characterized in that, The method is applied to a multi-agent self-evolutionary system based on closed-loop evaluation feedback as described in any one of claims 1-7, and includes the following steps: Step S1, Multi-source data perception and intent modeling: The multi-agent self-evolution system receives multimodal input from users. The intent understanding layer combines user features and domain knowledge from the semantic core and basic support layer to eliminate ambiguity and construct structured task instance objects. Step S2, Collaborative Twin Construction and Dynamic Planning: Based on the task instance object, the task planning layer retrieves the best task template, instantiates the collaborative digital twin model, generates an initial execution plan diagram, and displays it to the user for pre-review. Step S3, Distributed Collaborative Execution and Trajectory Recording: The task execution layer schedules the corresponding intelligent agent cluster to execute sub-tasks according to the plan. During the execution process, the system records the thinking process, tool call parameters, intermediate outputs and communication logs of all intelligent agents in real time, forming an execution trajectory log for the entire life cycle. Step S4, Dynamic Human-Machine Alignment and Replanning: During execution, if the user issues an intervention signal through the human-machine collaboration interface, the system immediately suspends the current execution. The intent understanding layer parses the user's new intent, and the task planning layer performs dynamic replanning based on the current twin state and the new intent to generate a corrected execution path. Step S5, Asynchronous Closed-Loop Evaluation and Instruction Generation: After the task is completed, the evaluation and supervision layer asynchronously reads the execution trajectory data, performs multi-dimensional scoring against the evaluation standard knowledge base, identifies defects or highlights in the execution, and generates structured optimization instructions. Step S6, Knowledge Internalization and System Evolution: The structured instructions generated in step S5 are parsed, and the maintenance agent in the semantic core and basic support layer receives and executes the instructions, updates the corresponding knowledge base entries in the semantic core, or adjusts the configuration parameters of the agent, and completes a self-evolution iteration of the system.

9. A multi-agent self-evolution method based on closed-loop evaluation feedback according to claim 8, characterized in that, Step S3 includes a secure and controllable invocation mechanism, specifically as follows: Before the corresponding intelligent agent attempts to call external tools or APIs, the system performs permission verification based on the tool capability knowledge base in the semantic core and basic support layer. For operations marked as high-risk or irreversible, the system will forcibly trigger a human-machine collaborative confirmation process, and the operation can only be executed after obtaining explicit authorization from the user.

10. A multi-agent self-evolution method based on closed-loop evaluation feedback according to claim 8, characterized in that, The execution trajectory log generated in step S3 has the following specifications: The execution trajectory log includes the agent's thought chain, tool call parameters and return results, complete data of intermediate outputs, execution time and resource consumption data of each stage; The execution trajectory log is stored in a standardized format, retains complete timestamps, ensures the temporal correspondence of data in each stage, and supports log backtracking and reproduction of the task execution process; The execution trajectory log is asynchronously pushed to the evaluation and supervision layer. The pushed content is the full log data, and the log data is retained in the system for no less than the most recent 100 system evolution iteration cycles, providing data support for multiple rounds of evaluation and evolution analysis.

Images