A method and system for persistent evolution of agents based on multi-layer capability assets

Abstract

The application relates to a method and system for persistent evolution of an intelligent agent based on multi-layer capability assets, and the method comprises the following steps: S1, collecting a task execution track; S2, performing experience attribution analysis on the execution track; S3, abstractly generating a candidate capability asset from the attribution result; S4, performing multi-dimensional access evaluation on the candidate capability asset; S5, storing the capability asset in a capability asset warehouse and updating a capability atlas; and S6, versioned deployment, gray release, effect monitoring and dynamic rollback. The application not only has significant improvement in key indicators such as the first resolution rate, task processing speed and user satisfaction, but also can quickly locate the most relevant capability asset set for a new task.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for the continuous evolution of intelligent agents based on multi-layered capability assets. Background Technology

[0002] With the rapid development of large language model technology, intelligent agent systems based on large language models are gradually acquiring core capabilities such as task understanding, plan generation, tool invocation, external environment interaction, and result feedback. Existing intelligent agent systems typically include a cue word management module, a short-term memory module, a long-term memory module, a tool invocation module, and a task planning module, enabling them to complete complex multi-step tasks to a certain extent. However, existing solutions still have several shortcomings in terms of the continuous evolution and accumulation of intelligent agent capabilities.

[0003] In the area of ​​memory enhancement, existing technologies enable agents to recall historical information, user preferences, and domain knowledge in subsequent tasks through short-term memory, long-term memory, and personalized knowledge bases. However, the objects they store and retrieve are limited to dialogue text and historical records, and they cannot perform structured abstraction of higher-level task strategies, tool call chains, and workflows. In the area of ​​self-reflective memory, existing solutions summarize historical task communications and the reasons for task success or failure through reflective agents and write the summary into long-term memory. However, the reflection results are stored in natural language text, lacking executability and cross-task transferability. In the area of ​​prompt word optimization, existing systems dynamically optimize prompt words based on user feedback, historical interactions, and task results to improve the matching degree between output content and user needs in specific scenarios. However, the optimized prompt words lack version management and rollback mechanisms and cannot be shared and reused among other agents. In the area of ​​task decomposition and dynamic programming, existing solutions improve the ability to handle complex tasks through task decomposition, sub-task prompt word generation, thought space networks, and dynamic backtracking. However, the optimization process of planning strategies lacks structured evaluation, making it difficult to distinguish between strategy improvement and random fluctuations. In terms of program memory or tool memory, existing technologies treat code snippets, function blocks, external APIs, or tool call logic as program memory for intelligent agents, enabling them to quickly call existing execution capabilities in subsequent tasks. However, their focus is limited to the tool layer and fails to cover higher-level reusable capabilities such as task strategies, workflows, and evaluation rules.

[0004] In summary, the existing technical solutions have the following shortcomings: First, the evolution target is singular. Existing solutions typically focus only on one aspect of memory, cue words, tools, or task breakdown, lacking a mechanism for unified abstraction and hierarchical management of agent capabilities, resulting in the inability of improvements in different aspects to work synergistically.

[0005] Second, there is a lack of a multi-layered capability asset system. During operation, intelligent agents generate not only memories, but also a variety of reusable capabilities such as task strategies, prompt word templates, tool call chains, workflow fragments, sub-agent roles, evaluation rules, security policies, and model adaptation parameters. Existing solutions lack a unified asset representation and hierarchical management of these capabilities.

[0006] Third, there is a lack of capability asset access mechanisms. Existing solutions typically write generated experiences, prompts, or tools directly into memory or tool libraries, lacking multi-dimensional verification of their effectiveness, generalization, security, cost, and stability. This may lead to low-quality or harmful content contaminating the intelligent agent system.

[0007] Fourth, there is a lack of version control and rollback mechanisms. During the continuous optimization of intelligent agents, problems such as erroneous experience pollution, degraded prompts, tool execution failures, and unstable workflows may occur. Existing solutions typically lack complete version tree management, canary releases, A / B testing, failure rollback, and change auditing mechanisms.

[0008] Fifth, there is a lack of cross-agent co-evolutionary mechanisms. Multiple agents accumulate different experiences in different tasks, but existing solutions lack a co-evolutionary mechanism that involves desensitization, verification, and hierarchical sharing, which can easily lead to isolated experiences or the spread of erroneous experiences.

[0009] Sixth, there is a lack of an abstraction mechanism from task trajectories to transferable capabilities. Existing solutions often only store dialogue history or tool code, failing to extract failure modes, success strategies, general SOPs, evaluation rules, and reusable workflows from task execution trajectories.

[0010] Therefore, it is necessary to propose a continuous evolution method for intelligent agents based on multi-layer capability assets, which enables intelligent agents to automatically extract multi-layer capability assets from historical tasks and apply effective capabilities to future tasks in a stable, secure, and controllable manner through access verification, version governance, and canary deployment mechanisms. Summary of the Invention

[0011] In view of the shortcomings of the prior art, the present invention provides a method and system for the continuous evolution of intelligent agents based on multi-layer capability assets, which not only significantly improves key indicators such as first-time resolution rate, task processing speed and user satisfaction, but also can quickly locate the most relevant set of capability assets when a new task is encountered.

[0012] To achieve the above and other related objectives, the present invention provides the following technical solution: A method for the continuous evolution of agents based on multi-layered capability assets, the method comprising: S1. Collect the task execution trajectory; S2. Perform empirical attribution analysis on the execution trajectory; S3. Abstract candidate capability assets from attribution results; S4. Conduct a multi-dimensional admission assessment of candidate capability assets; S5. Store in the ability asset warehouse and update the ability map; S6. Versioned deployment, canary release, effect monitoring and dynamic rollback.

[0013] Further, in step S4, each candidate capability asset is sequentially subjected to five evaluation checkpoints: validity verification, generalization verification, security audit, cost assessment, and stability test. If any checkpoint is not passed, the asset is rejected and returned to step S2, with an evaluation report attached for subsequent optimization reference; if all checkpoints are passed, the process proceeds to step S5.

[0014] Furthermore, the validity verification involves the system applying the candidate asset to a set of standard test tasks and calculating its performance improvement relative to the baseline. Only when the improvement exceeds a preset threshold can it proceed to generalization verification. The generalization verification involves the system testing the performance of the asset in diverse task scenarios to verify its adaptability in different domains, complexities, and user groups, preventing the asset from overfitting to specific scenarios.

[0015] Furthermore, the security audit involves the system performing multi-layered security checks on the asset content, including harmful content detection, data leakage risk assessment, permission overreach detection, and adversarial attack testing. The cost assessment involves the system comprehensively analyzing the impact of the asset on storage overhead, inference latency, token consumption, and API call fees, and weighing these against the benefits to ensure that the marginal benefits of introducing the asset exceed the marginal costs. The stability test involves the system evaluating the performance variance of the asset in multiple repeated executions to detect whether there are uncertain fluctuations, degradation trends, or failures under boundary conditions.

[0016] Furthermore, in step S1, the acquisition of the task execution trajectory refers to the system automatically collecting complete execution trajectory data when the agent performs the user task, including task definition, agent reasoning process, tool calls and their parameters and return results at each step, intermediate decisions and planning adjustments, final output results, user evaluation feedback, and system-level performance indicators.

[0017] Furthermore, in step S2, the empirical attribution analysis of the execution trajectory involves the system performing structured processing on the collected trajectory data, attributing the task results to specific decision-making steps, prompt word design, tool selection, and execution order, and identifying key factors that have a decisive impact on the task results through comparative analysis of successful and failed trajectories.

[0018] Further, in step S3, the abstraction of candidate capability assets from the attribution results involves the system generating capability assets from different abstraction levels based on the attribution analysis results. At the memory layer, key facts, user preferences, and domain knowledge are extracted as memory assets. At the prompt word layer, optimized task instruction templates, role settings, and constraint rules are generated. At the strategy layer, task decomposition strategies, planning strategies, and reflection strategies are encoded. At the tool layer, independent API call patterns, code functions, and external service calls are encapsulated. At the workflow layer, multi-step orchestration templates, conditional branch logic, and sub-process definitions are generated. At the sub-agent layer, role configurations, skill combinations, and collaboration protocols are defined. At the evaluation layer, evaluation criteria and scoring rules are refined. At the security layer, content filtering rules and access control strategies are generated. At the model adaptation layer, optimal parameter configurations and inference settings are recorded.

[0019] Furthermore, in step S5, storing assets in the capability asset warehouse and updating the capability graph means that assets that have passed the access assessment are stored in the capability asset warehouse, assigned a unique asset identifier and an initial version number, and at the same time, the corresponding nodes and the relationship edges between nodes are created or updated in the capability graph.

[0020] Furthermore, in step S6, the versioned deployment, canary release, effect monitoring, and dynamic rollback are the steps to gradually push the new version's capability assets to the production environment according to a predetermined strategy, from the test environment to the canary environment and then to the full environment. Effect monitoring is performed at each stage, and rollback is automatically triggered when an anomaly is detected. The canary release mechanism pushes the new version's capability assets to the production environment in stages: the first stage involves full-function verification in an isolated test environment; the second stage releases the version to a small proportion of traffic and monitors key performance indicators; the third stage gradually increases the traffic proportion, continuously monitoring and comparing the effects of A / B groups; the fourth stage releases it to full traffic and maintains an observation period. At any stage, if the monitoring system detects an abnormal drop in key indicators, the system automatically triggers the rollback process, reverting the asset to the previous stable version and generating an anomaly report. The rollback operation is also recorded in the version tree, forming a rollback version node for subsequent analysis and improvement.

[0021] To achieve the above and other related objectives, this invention also provides a system for the continuous evolution of intelligent agents based on multi-layered capability assets, used to implement the aforementioned method for the continuous evolution of intelligent agents based on multi-layered capability assets. The system includes a user task interface, an intelligent agent execution module, a task trajectory acquisition module, an experience attribution module, a capability asset generation module, an admission assessment module, a capability asset repository, a capability graph, a dynamic invocation module, and a gray-scale rollback module. The user submits tasks to the intelligent agent execution module through the user task interface. After receiving the tasks, the intelligent agent understands, reasones, and plans, and completes the task execution through tool invocation and external interaction. The task trajectory acquisition module records all trajectory data in real time during the execution of the intelligent agent, including prompt input, inference chain, tool call sequence, intermediate results, final output, abnormal information, explicit and implicit user feedback, as well as task success rate, response time, and cost evaluation indicators. The experience attribution module receives raw trajectory data provided by the task trajectory acquisition module and transforms it into structured evolutionary signals. Specifically, this includes: quantifying user feedback and evaluation results into scoring vectors; marking task success or failure as positive or negative samples; identifying key decision points and performance bottlenecks through comparative analysis; and mapping tool call sequences and intermediate results to corresponding nodes and edge relationships in the capability asset space. The capability asset generation module, based on the structured evolutionary signal output by the experience attribution module, abstracts and generates candidate capability assets from multiple dimensions, including extracting reusable prompt word templates, task decomposition strategies, and tool call chains from success trajectories, extracting constraint rules and safety boundaries from failure trajectories, and extracting independent API call patterns and combination patterns from tool call sequences. The admission assessment module evaluates candidate capability assets from multiple dimensions, including effectiveness, generalization, security, cost, and stability. Capability assets that pass the admission assessment are stored in the capability asset repository, and the nodes and edge relationships in the capability graph are updated synchronously. The gray rollback module manages the versioning of the capability assets stored in the repository, supports gray release, A / B testing, online monitoring, and automatic rollback in case of anomalies, and ensures that the update process of capability assets is safe and controllable. The dynamic invocation module is connected to the intelligent agent execution module and is used to execute commands issued by the intelligent agent execution module and provide feedback on the results of task execution completed through external interaction.

[0022] The present invention has the following positive effects: 1. This invention transforms the experience gained during the operation of intelligent agents into structured and standardized capability assets, avoiding repetitive learning and trial-and-error. Intelligent agent systems deploying this invention show significant improvements in key indicators such as first-time resolution rate, task processing speed, and user satisfaction when handling similar tasks, compared to baseline systems that do not adopt this solution. Regarding capability quality assurance, the five-stage admission evaluation mechanism ensures the quality and reliability of the capability assets included in the database. The probability of approved assets exhibiting positive effects in subsequent use is significantly higher than in solutions that directly include assets without an evaluation mechanism.

[0023] 2. Regarding system stability and security, this invention utilizes version trees and gray-scale rollback mechanisms to keep the production accident rate caused by capability updates at an extremely low level, significantly shortening the average recovery time. In terms of multi-agent collaboration, the hierarchical sharing mechanism significantly shortens the capability accumulation cycle for newly deployed agents. By reusing existing capability assets, the initial task success rate of new agents is far higher than that of agents trained from scratch. Regarding the capability graph, graph traversal-based reasoning queries enable agents to quickly locate the most relevant set of capability assets when facing new tasks, achieving millisecond-level asset retrieval time, and the relevance score of recommended assets is superior to traditional keyword-matching-based retrieval schemes.

[0024] 3. This invention improves the long-term adaptability of intelligent agents to complex tasks, reduces the cost of repeated trial and error, improves the efficiency of tool invocation and workflow orchestration, reduces the risk of erroneous experience contaminating the system, prevents the degradation of capabilities during the continuous optimization process of intelligent agents, improves the interpretability and traceability of the evolution process of intelligent agents, supports capability reuse and co-evolution among multiple intelligent agents, improves the security and stability in high-risk tasks, and supports the continuous operation and capability accumulation of enterprise-level intelligent agent systems. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a flowchart of the capability asset access assessment process of the present invention; Figure 3 This is a diagram of the multi-layered capability asset structure of the present invention; Figure 4 This is a schematic diagram of multi-agent cooperative evolution according to the present invention; Figure 5 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0026] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0027] Example: Figure 1 As shown, a method for the continuous evolution of an agent based on multi-layered capability assets is described, the method comprising: S1. Collect the task execution trajectory; S2. Perform empirical attribution analysis on the execution trajectory; S3. Abstract candidate capability assets from attribution results; S4. Conduct a multi-dimensional admission assessment of candidate capability assets; S5. Store in the ability asset warehouse and update the ability map; S6. Versioned deployment, canary release, effect monitoring and dynamic rollback.

[0028] In this embodiment, in step S4, each candidate capability asset is sequentially subjected to five evaluation checkpoints: validity verification, generalization verification, security audit, cost assessment, and stability test. If any checkpoint is not passed, the asset is rejected and returned to step S2, with an evaluation report attached for subsequent optimization reference; if all checkpoints are passed, the process proceeds to step S5.

[0029] In this embodiment, as Figure 2 As shown, the validity verification involves the system applying the candidate asset to a set of standard test tasks and calculating its performance improvement relative to the baseline. Only when the improvement exceeds a preset threshold can it proceed to generalization verification. The generalization verification involves the system testing the performance of the asset in diverse task scenarios to verify its adaptability in different domains, complexities, and user groups, preventing the asset from overfitting to a specific scenario.

[0030] In this embodiment, the security audit involves the system performing multi-layered security checks on the asset content, including harmful content detection, data leakage risk assessment, permission overreach detection, and adversarial attack testing. The cost assessment involves the system comprehensively analyzing the impact of the asset on storage overhead, inference latency, token consumption, and API call fees, and weighing these against the benefits to ensure that the marginal benefits of introducing the asset are greater than the marginal costs. The stability test involves the system evaluating the performance variance of the asset in multiple repeated executions to detect whether there are uncertain fluctuations, degradation trends, or failures under boundary conditions.

[0031] In this embodiment, as Figure 3 As shown, this invention defines a ten-layer capability asset system, progressing from concrete to abstract. The first layer is Memory Assets, including short-term memory, long-term memory, dialogue history, and user preferences—the most basic form of experience storage. The second layer is Prompt Assets, including task instruction templates, role settings, and constraint rules—refined and reusable prompt designs. The third layer is TaskStrategy Assets, including task decomposition strategies, planning strategies, and reflection strategies—an abstraction of the agent's thinking and decision-making methods. The fourth layer is Tool Assets, including API call patterns, code functions, and external service encapsulation—executable tool usage knowledge.

[0032] The fifth layer consists of Workflow Assets, including multi-step orchestration templates, conditional branches, and sub-process definitions, representing the combined orchestration of multiple tools and strategies. The sixth layer comprises Sub-Agent Assets, including role definitions, skill configurations, and collaboration protocols, encapsulating the complete capabilities of an agent. The seventh layer consists of Evaluation Rule Assets, including evaluation criteria, scoring rules, and quality gates, serving as the basis for agent self-evaluation. The eighth layer consists of Safety Policy Assets, including content filtering rules, access control policies, and exception handling protocols. The ninth layer consists of Model Adaptation Assets, including model parameter configurations, context window settings, and inference optimization parameters. The tenth layer consists of Capability Graph Assets, which record the relationships between task types, user groups, failure modes, experience sources, capability assets, tools, workflows, sub-agents, evaluation metrics, and versions in a graph structure, serving as the metadata layer for the entire capability asset system.

[0033] In this embodiment, in step S1, the acquisition of the task execution trajectory is the automatic collection of complete execution trajectory data by the system when the agent performs the user task. This includes the task definition, the agent's reasoning process, the tool calls and their parameters and return results at each step, intermediate decisions and planning adjustments, the final output results, user evaluation feedback, and system-level performance indicators.

[0034] In this embodiment, as Figure 4 As shown, this invention supports collaborative evolution among multiple agents through a shared capability asset repository. Each agent generates candidate capability assets while independently executing tasks within its own domain, and these assets are stored in the shared repository after admission evaluation. Other agents can retrieve and call verified capability assets from the repository according to their own task requirements, avoiding redundant invention and trial-and-error. Shared assets are adapted for cross-agent use to suit the task characteristics and configuration parameters of the target agent.

[0035] To ensure the security and controllability of co-evolution, the system implements a tiered sharing strategy. Tier 1 is individual-level, where assets are only usable by the agent that generated them. Tier 2 is team-level, where assets are shared among multiple agents within the same organization or project group. Tier 3 is global-level, where assets are open to all agents. The sharing level of an asset is determined jointly by the admission assessment results and manual approval. The system tracks and records each cross-agent asset call, collects its effect data on the target agent, and feeds it back into the asset's evaluation metrics. When the same asset achieves consistently positive results on multiple agents, its credibility and recommendation priority are automatically increased. When an asset exhibits performance degradation in a new scenario, the system automatically notifies all agents using the asset and conducts an overall evaluation.

[0036] The capability graph is the core data structure of this invention, constructed using a directed attribute graph model. Node types in the capability graph include task type nodes, user group nodes, failure mode nodes, experience source nodes, capability asset nodes, tool nodes, workflow nodes, sub-agent nodes, evaluation metric nodes, and version nodes. Edge relationship types include "belongs to," "references," "dependencies," "associations," "sequences," and "evolutions."

[0037] The capability graph is constructed as follows: When a new capability asset is admitted and stored in the repository, the system automatically creates a corresponding asset node in the capability graph and establishes edge relationships with other nodes based on the asset's metadata and content characteristics. When an agent executes a task, the system updates the attribute values ​​of task nodes, tool nodes, and evaluation metric nodes in the graph in real time. The capability graph supports graph traversal-based reasoning queries. For example, when an agent encounters a new task, the system can use graph search to discover the set of capability assets most relevant to that task type, as well as the strategy paths used in handling similar tasks historically. Graph update operations ensure consistency through a transaction mechanism and support both incremental updates and full reconstruction modes.

[0038] In this embodiment, in step S2, the empirical attribution analysis of the execution trajectory involves the system performing structured processing on the collected trajectory data, attributing the task results to specific decision-making steps, prompt word design, tool selection, and execution order, and identifying key factors that have a decisive impact on the task results through comparative analysis of successful and failed trajectories.

[0039] In this embodiment, in step S3, the abstraction of candidate capability assets from the attribution results involves the system generating capability assets from different abstraction levels based on the attribution analysis results. At the memory layer, key facts, user preferences, and domain knowledge are extracted as memory assets. At the prompt word layer, optimized task instruction templates, role settings, and constraint rules are generated. At the strategy layer, task decomposition strategies, planning strategies, and reflection strategies are encoded. At the tool layer, independent API call patterns, code functions, and external service calls are encapsulated. At the workflow layer, multi-step orchestration templates, conditional branch logic, and sub-process definitions are generated. At the sub-agent layer, role configurations, skill combinations, and collaboration protocols are defined. At the evaluation layer, evaluation criteria and scoring rules are refined. At the security layer, content filtering rules and access control strategies are generated. At the model adaptation layer, optimal parameter configurations and inference settings are recorded.

[0040] In this embodiment, in step S5, storing the capability asset warehouse and updating the capability graph means that the assets that have passed the access assessment are stored in the capability asset warehouse, assigned a unique asset identifier and an initial version number, and at the same time, the corresponding nodes and the relationship edges between the nodes are created or updated in the capability graph.

[0041] In this embodiment, in step S6, the versioned deployment, canary release, effect monitoring, and dynamic rollback are the steps to gradually push the new version's capability assets to the production environment according to a predetermined strategy, from the test environment to the canary environment and then to the full environment. Effect monitoring is performed at each stage, and a rollback is automatically triggered when an anomaly is detected. The canary release mechanism pushes the new version's capability assets to the production environment in stages: the first stage involves full-function verification in an isolated test environment; the second stage releases the version to a small proportion of traffic and monitors key performance indicators; the third stage gradually increases the traffic proportion, continuously monitoring and comparing the effects of A / B groups; and the fourth stage releases it to full traffic and maintains an observation period. At any stage, if the monitoring system detects an abnormal drop in key indicators, the system automatically triggers the rollback process, reverting the asset to the previous stable version and generating an anomaly report. The rollback operation is also recorded in the version tree, forming a rollback version node for subsequent analysis and improvement.

[0042] In this embodiment, as Figure 5 As shown, this invention provides a system for the continuous evolution of intelligent agents based on multi-layered capability assets, used to implement the aforementioned method for the continuous evolution of intelligent agents based on multi-layered capability assets. The system includes a user task interface, an intelligent agent execution module, a task trajectory acquisition module, an experience attribution module, a capability asset generation module, an admission assessment module, a capability asset repository, a capability graph, a dynamic invocation module, and a gray-scale rollback module. The user submits tasks to the intelligent agent execution module through the user task interface. After receiving the tasks, the intelligent agent understands, reasones, and plans, and completes the task execution through tool invocation and external interaction. The task trajectory acquisition module records all trajectory data in real time during the execution of the intelligent agent, including prompt input, inference chain, tool call sequence, intermediate results, final output, abnormal information, explicit and implicit user feedback, and evaluation indicators such as task success rate, response time, and cost. The experience attribution module receives raw trajectory data provided by the task trajectory acquisition module and transforms it into structured evolutionary signals. Specifically, this includes: quantifying user feedback and evaluation results into scoring vectors; marking task success or failure as positive or negative samples; identifying key decision points and performance bottlenecks through comparative analysis; and mapping tool call sequences and intermediate results to corresponding nodes and edge relationships in the capability asset space. The capability asset generation module, based on the structured evolutionary signal output by the experience attribution module, abstracts and generates candidate capability assets from multiple dimensions, including extracting reusable prompt word templates, task decomposition strategies, and tool call chains from success trajectories, extracting constraint rules and safety boundaries from failure trajectories, and extracting independent API call patterns and combination patterns from tool call sequences. The admission assessment module evaluates candidate capability assets from multiple dimensions, including effectiveness, generalization, security, cost, and stability. Capability assets that pass the admission assessment are stored in the capability asset repository, and the nodes and edge relationships in the capability graph are updated synchronously. The gray rollback module manages the versioning of the capability assets stored in the repository, supports gray release, A / B testing, online monitoring, and automatic rollback in case of anomalies, and ensures that the update process of capability assets is safe and controllable.

[0043] In this embodiment, the implementation environment and configuration rely on a large language model intelligent agent platform. Specifically, the following technology stack can be used: The intelligent agent kernel is based on a large language model with a Transformer architecture, running on a GPU-accelerated computing cluster. The task trajectory acquisition module is embedded into the core execution flow of the intelligent agent via AOP, achieving full trajectory recording without intruding on business logic. The experience attribution module uses difference analysis and decision tree attribution methods to attribute task results to specific execution steps and decision points. The capability graph storage is based on a graph database, supporting efficient relation traversal and subgraph matching. The capability asset repository uses a combination of object storage and a metadata database; asset entities are stored in a distributed file system, while metadata and version information are stored in a relational database.

[0044] The intelligent agent for autonomous driving data closed-loop continues to evolve. In a specific embodiment, this invention is applied to an autonomous driving data closed-loop scenario. During the development of autonomous driving, massive amounts of roadside data are generated daily, requiring a series of complex operations such as data collection task configuration, data annotation rule generation, scene mining, model training task orchestration, and evaluation report generation. A group of intelligent agents is deployed to assume the roles of data collection and configuration agent, annotation rule agent, scene mining agent, model training agent, and evaluation and analysis agent, respectively. These agents collaborate to complete each stage of the data closed-loop process. The dynamic invocation module is connected to the intelligent agent execution module and is used to execute commands issued by the intelligent agent execution module and provide feedback on the results of task execution completed through external interaction.

[0045] Taking a scene mining agent as an example, its core task is to automatically identify and extract valuable driving scenarios from massive amounts of road data, such as emergency braking scenarios, unusual vehicle entry scenarios, low-light nighttime scenarios, and severe weather scenarios. When the scene mining agent performs a scene mining task, the system collects its complete execution trajectory: the agent first reads the data acquisition configuration to obtain sensor type and data format information, then performs data annotation according to the annotation rules provided by the annotation rule generation module, then runs the scene mining algorithm to perform pattern recognition on the annotated data, filters out target scene fragments that meet preset conditions, and finally outputs the mining results to the model training task orchestration module. The task trajectory acquisition module records the input and output parameters of each step, algorithm call records and return results, as well as the number of scenes mined and the changes in evaluation metrics after using the scene data in subsequent model training.

[0046] The attribution module analysis revealed that the current scene mining agent has a low recall rate when identifying "atypical intersections." The attribution results indicate two main issues: first, the scene mining strategy lacks multi-dimensional feature descriptions of the intersection's topology, preventing the algorithm from effectively distinguishing between regular and atypical intersections; second, the labeling rules do not cover subcategories such as "unlabeled intersections," "roundabout exits," and "multi-way intersections." Based on these attribution results, the capability asset generation module abstracts the following candidate capability assets from successfully mined trajectories: at the tool layer, it generates an "Intersection Topology Feature Extraction Tool Asset," encapsulating multi-dimensional feature calculation functions and calling interfaces based on road topology; at the strategy layer, it generates an "Atypical Intersection Scene Mining Strategy Asset," containing a multi-stage screening strategy: a three-level progressive mining strategy involving coarse screening through geometric features, fine screening through topological structure, and final confirmation through semantic verification; and at the evaluation rule layer, it generates a "Scene Mining Recall Rate Evaluation Rule Asset," defining a tiered recall rate calculation method and threshold based on scene type.

[0047] These three candidate capability assets were entered into the capability asset repository after undergoing a five-stage admission assessment. During the assessment process, effectiveness verification confirmed a significant improvement in recall rate at atypical intersections through historical data playback; generalization verification confirmed the adaptability of the strategy to cross-regional scenarios on datasets from multiple different cities; security audit confirmed that the tool assets did not pose a data access violation risk; cost assessment showed that the cost of additional tool calls for each scenario mining task was relatively low; and stability testing showed that the recall rate remained consistent and fluctuated little in multiple repeated executions.

[0048] Assets that pass the access test are assigned version numbers and stored in the database. The system creates corresponding nodes in the capability graph: intersection topology feature extraction tool node, atypical intersection scene mining strategy node, and scene mining recall evaluation rule node, and establishes edge relationships with existing "scene mining" task type nodes, "annotation rule" nodes, and "training data" nodes. Other agents retrieve these new assets through the capability graph. The annotation rule agent supplements the annotation rule categories such as "unlabeled intersection," "roundabout exit," and "multi-way intersection" according to the new scene types. The model training agent adjusts the sampling weights of the training data according to the new mining strategy, increasing the proportion of atypical intersection scene training samples.

[0049] Regarding version management, the version evolution of the "Intersection Topology Feature Extraction Tool Asset" is as follows: Version 1.0 implemented basic topology feature calculation functions; version 1.1 added a real-time visual debugging interface based on user feedback from the scene mining agent, and after minor version upgrades were verified in a gray-scale environment and confirmed to be error-free, it was fully released; version 2.0 restructured the feature calculation architecture to support topology analysis of LiDAR point cloud data. During the gray-scale testing phase, a compatibility issue with the interface of the existing annotation rule agent was discovered, and the system automatically rolled back to version 1.1. After fixing the interface, it was re-released as version 2.1. The entire version evolution process is fully recorded through a version tree, and any historical version can be rolled back with one click. As shown in Table 1, Table 1

[0050] The difference between this invention and existing memory enhancement solutions is that existing solutions mainly solve the problem of storing and recalling historical information, while this invention not only stores information, but also transforms task trajectories into multi-layered capability assets such as prompts, tools, workflows, evaluation rules, and security policies, achieving a qualitative leap from information storage to capability accumulation.

[0051] The difference between this invention and existing prompt word optimization solutions is that existing solutions mainly optimize prompt words based on user feedback, while this invention treats prompt words as only one of the multi-layered capability assets, and introduces capability asset access, version governance, and gray-scale rollback mechanisms to ensure the security and controllability of prompt word optimization.

[0052] The difference between this invention and existing program memory schemes is that existing program memory schemes mainly focus on the generation, retrieval, and execution of code, functions, and tools. This invention includes not only tool assets, but also task strategy assets, workflow assets, sub-agent assets, evaluation rule assets, security strategy assets, and capability graph assets. It also implements unified access control and version governance for all assets, covering the complete capability chain from execution to evaluation.

[0053] The difference between this invention and existing task decomposition and dynamic programming schemes is that existing schemes mainly focus on how to decompose and execute a single task, while this invention focuses on the continuous evolution process of the agent across tasks, time, and versions, enabling the agent to accumulate transferable capability assets from historical tasks and continuously improve the processing level of future tasks.

[0054] The difference between this invention and existing long-term memory reflection schemes is that existing schemes mainly write the reflection results into long-term memory, while this invention further transforms the reflection results into verifiable, deployable, and rollbackable capability assets, and supports multi-agent collaborative sharing, thus extending the value of reflection from the individual level to the group level.

[0055] In summary, this invention not only significantly improves key indicators such as first-time resolution rate, task processing speed, and user satisfaction, but also enables the rapid identification of the most relevant set of capability assets for new tasks.

[0056] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for the continuous evolution of intelligent agents based on multi-layered capability assets, characterized in that, The method includes: S1. Collect the task execution trajectory; S2. Perform empirical attribution analysis on the execution trajectory; S3. Abstract candidate capability assets from attribution results; S4. Conduct a multi-dimensional admission assessment of candidate capability assets; S5. Store in the ability asset warehouse and update the ability map; S6. Versioned deployment, canary release, effect monitoring and dynamic rollback.

2. The method for continuous evolution of intelligent agents based on multi-layered capability assets according to claim 1, characterized in that, In step S4, each candidate capability asset is sequentially subjected to five evaluation checkpoints: validity verification, generalization verification, security audit, cost assessment, and stability test. If any checkpoint is not passed, the asset is rejected and returned to step S2, with an evaluation report attached for subsequent optimization reference. If all passes, proceed to step S5.

3. The method for continuous evolution of agents based on multi-layered capability assets according to claim 2, characterized in that: The validity verification involves the system applying the candidate asset to a set of standard test tasks and calculating its performance improvement relative to the baseline. Only when the improvement exceeds a preset threshold can it proceed to generalization verification. The generalization verification involves the system testing the asset's performance in diverse task scenarios to verify its adaptability in different domains, complexities, and user groups, preventing the asset from overfitting to specific scenarios.

4. The method for continuous evolution of agents based on multi-layered capability assets according to claim 2, characterized in that: The security audit involves the system performing multi-layered security checks on the asset content, including harmful content detection, data leakage risk assessment, permission overreach detection, and adversarial attack testing. The cost assessment involves the system comprehensively analyzing the impact of the asset on storage overhead, inference latency, token consumption, and API call fees, and weighing these against the benefits to ensure that the marginal benefits of introducing the asset exceed the marginal costs. The stability test involves the system evaluating the performance variance of the asset in multiple repeated executions to detect whether there are uncertain fluctuations, degradation trends, or failures under boundary conditions.

5. The method for continuous evolution of agents based on multi-layered capability assets according to claim 1, characterized in that, In step S1, the acquisition of the task execution trajectory refers to the automatic collection of complete execution trajectory data by the system when the agent performs the user task. This data includes the task definition, the agent's reasoning process, the tool calls and their parameters and return results at each step, intermediate decisions and planning adjustments, the final output results, user evaluation feedback, and system-level performance indicators.

6. The method for continuous evolution of intelligent agents based on multi-layered capability assets according to claim 1, characterized in that, In step S2, the empirical attribution analysis of the execution trajectory involves the system structuring the collected trajectory data, attributing the task results to specific decision-making steps, prompt word design, tool selection, and execution order, and identifying key factors that have a decisive impact on the task results through comparative analysis of successful and failed trajectories.

7. The method for continuous evolution of agents based on multi-layered capability assets according to claim 1, characterized in that, In step S3, the abstract generation of candidate capability assets from the attribution results is the process by which the system generates capability assets from different abstraction levels based on the results of attribution analysis. At the memory level, key facts, user preferences, and domain knowledge are extracted as memory assets. At the prompt word layer, optimized task instruction templates, role settings, and constraint rules are generated; at the strategy layer, task decomposition strategies, planning strategies, and reflection strategies are encoded; at the tool layer, independent API call patterns, code functions, and external service calls are encapsulated; at the workflow layer, multi-step orchestration templates, conditional branching logic, and sub-process definitions are generated. At the sub-agent layer, role configuration, skill combinations, and collaboration protocols are defined; at the evaluation layer, evaluation criteria and scoring rules are extracted; at the security layer, content filtering rules and access control policies are generated. In the model adaptation layer, the optimal parameter configuration and inference settings are recorded.

8. The method for continuous evolution of intelligent agents based on multi-layered capability assets according to claim 1, characterized in that, In step S5, storing assets in the capability asset warehouse and updating the capability graph means that assets that have passed the access assessment are stored in the capability asset warehouse, assigned a unique asset identifier and an initial version number, and at the same time, the corresponding nodes and the relationship edges between nodes are created or updated in the capability graph.

9. The method for continuous evolution of intelligent agents based on multi-layered capability assets according to claim 1, characterized in that, In step S6, the versioned deployment, canary release, effect monitoring, and dynamic rollback are the steps to gradually push the new version's capability assets to the production environment according to a predetermined strategy, from the test environment to the canary environment and then to the full environment. Effect monitoring is carried out at each stage, and rollback is automatically triggered when an anomaly is detected. The canary release mechanism pushes the new version's capability assets to the production environment in stages. The first stage is to conduct full-function verification in an isolated test environment; the second stage is to release the version to a small proportion of traffic and monitor key performance indicators; the third stage is to gradually increase the proportion of traffic, continuously monitor and compare the effects of A / B groups. The fourth phase involves releasing the product across all traffic and maintaining an observation period. During any phase, if the monitoring system detects an abnormal drop in key metrics, the system will automatically trigger a rollback process, reverting the asset to the previous stable version and generating an anomaly report. The rollback operation is also recorded in the version tree, forming a rollback version node for subsequent analysis and improvement.

10. A system for the continuous evolution of intelligent agents based on multi-layered capability assets, characterized in that, A method for implementing the continuous evolution of an agent based on multi-layered capability assets as described in any one of claims 1-9, the system comprising a user task interface, an agent execution module, a task trajectory acquisition module, an experience attribution module, a capability asset generation module, an admission assessment module, a capability asset repository, a capability graph, a dynamic invocation module, and a gray-scale rollback module. The user submits a task to the agent execution module through the user task interface. After receiving the task, the agent understands, reasones, and plans, and completes the task execution through tool invocation and external interaction. The task trajectory acquisition module records all trajectory data in real time during the execution of the intelligent agent, including prompt input, inference chain, tool call sequence, intermediate results, final output, abnormal information, explicit and implicit user feedback, as well as task success rate, response time, and cost evaluation indicators. The experience attribution module receives raw trajectory data provided by the task trajectory acquisition module and transforms it into structured evolutionary signals. Specifically, this includes: quantifying user feedback and evaluation results into scoring vectors; marking task success or failure as positive or negative samples; identifying key decision points and performance bottlenecks through comparative analysis; and mapping tool call sequences and intermediate results to corresponding nodes and edge relationships in the capability asset space. The capability asset generation module, based on the structured evolutionary signal output by the experience attribution module, abstracts and generates candidate capability assets from multiple dimensions, including extracting reusable prompt word templates, task decomposition strategies, and tool call chains from success trajectories, extracting constraint rules and safety boundaries from failure trajectories, and extracting independent API call patterns and combination patterns from tool call sequences. The admission assessment module evaluates candidate capability assets from multiple dimensions, including effectiveness, generalization, security, cost, and stability. Capability assets that pass the admission assessment are stored in the capability asset repository, and the nodes and edge relationships in the capability graph are updated synchronously. The gray rollback module manages the versioning of the capability assets stored in the repository, supports gray release, A / B testing, online monitoring, and automatic rollback in case of anomalies, and ensures that the update process of capability assets is safe and controllable. The dynamic invocation module is connected to the intelligent agent execution module and is used to execute commands issued by the intelligent agent execution module and provide feedback on the results of task execution completed through external interaction.

Images