Skill automatic evolution method and agent system

Abstract

This application provides a method for automatic skill evolution and an intelligent agent system. The method includes: if the intelligent agent responds to a target task, evaluating the execution process of the intelligent agent's response to the target task, wherein the execution process of the target task involves the invocation of a skill; in response to a first evaluation result satisfying a skill defect condition, repairing a first skill from a first skill library through a first self-evolutionary link, wherein the first skill corresponds to a skill; in response to a second evaluation result satisfying a skill missing condition, creating a second skill through a second self-evolutionary link and storing it in the first skill library; wherein the second skill corresponds to a skill.

Description

Technical Field

[0001] This application relates to the field of computer technology, and to, but is not limited to, a method for automatic skill evolution and an intelligent agent system. Background Technology

[0002] Currently, the skills used by intelligent agents are obtained from a cloud-based skill library. When the skills can no longer meet the functional requirements, they will continue to acquire skills with corresponding functions from the cloud to expand their capabilities.

[0003] Newly acquired skills may overlap significantly with existing local skills in terms of functionality, which may not only waste storage resources but also lead to policy conflicts and inconsistent results during execution due to the coexistence of multiple similar skills. Summary of the Invention

[0004] Based on the problems existing in related technologies, this application provides a method for automatic skill evolution and an intelligent agent system.

[0005] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for automatic skill evolution, the method comprising: If the agent responds to the target task, evaluate the agent's execution process of responding to the target task, which involves the invocation of a skill; In response to the first assessment result meeting the skill defect condition, the first skill from the first skill library is repaired through the first self-evolutionary link, and the first skill corresponds to the skill. In response to the second assessment result meeting the skill deficiency condition, a second skill is created through the second self-evolutionary link and stored in the first skill library; the second skill corresponds to the skill.

[0006] In some embodiments, the skill deficiency condition includes: a first score greater than a first threshold and the first score being higher than a second score by a first difference to characterize a skill deficiency; the skill lack condition includes: a second score greater than a second threshold and the second score being higher than a first score by a first difference to characterize a skill lack.

[0007] In some embodiments, the method further includes: if the agent responds to the target task, synchronously collecting the execution trajectory of the agent's execution process in response to the target task.

[0008] In some embodiments, evaluating the agent's response to the target task includes: evaluating the agent's response to the target task based on the execution trajectory, and obtaining a first score and a second score, wherein the first score represents a score for skill deficiencies and the second score represents a score for skill deficiencies.

[0009] In some embodiments, repairing a first skill from a first skill library via a first self-evolutionary link includes: determining the cause of the defect based on a first portion of the execution trajectory data, repairing the first skill as a third skill, and storing the third skill in the first skill library to replace the first skill; creating a second skill via a second self-evolutionary link includes: creating a second skill based on a second portion of the execution trajectory data.

[0010] In some embodiments, creating a second skill through a second self-evolutionary link includes: generating descriptive information of the second skill based on a second portion of the execution trajectory data; matching the descriptive information with skills in a second skill library based on similarity; obtaining a fourth skill whose similarity meets a similarity threshold; creating a second skill if the fourth skill does not meet the reuse condition; and using the fourth skill as the second skill if the fourth skill meets the reuse condition. The second skill library is a cloud-based skill library, and the first skill library is a local skill library.

[0011] In some embodiments, the first self-evolutionary link further includes: performing an automated execution test on the third skill before storing the third skill in the first skill library; the second self-evolutionary link further includes: performing an automated execution test on the second skill before storing the second skill in the first skill library.

[0012] In some embodiments, evaluating the execution process of an agent responding to a target task based on the execution trajectory to obtain a first score and a second score includes: evaluating at least one of skill hit rate, user feedback information, and execution result quality during the execution process of the agent responding to the target task based on the execution trajectory; and determining the first score and the second score based on at least one of skill hit rate, user feedback information, and execution result quality, respectively.

[0013] Secondly, embodiments of this application further provide an intelligent agent system, comprising: a first automatic evolutionary link module with the ability to repair skills; a second automatic evolutionary link module with the ability to create skills; an evaluation module for evaluating the execution process of the intelligent agent responding to a target task if the intelligent agent responds to the target task, wherein the execution process of the target task involves the invocation of a skill; and a control module for repairing a first skill from a first skill library through the first automatic evolutionary link module in response to a first evaluation result satisfying a skill defect condition, wherein the first skill corresponds to a skill; and for creating a second skill through the second automatic evolutionary link module and storing it in the first skill library in response to a second evaluation result satisfying a skill deficiency condition, wherein the second skill corresponds to a skill.

[0014] In some embodiments, the intelligent agent system further includes a communication device; the communication device is used to communicate with a second skill library to obtain skills from the second skill library; the second skill library is a cloud-based skill library.

[0015] Thirdly, embodiments of this application provide a terminal device, including: a storage device for storing an intelligent agent system; and a processor for executing instructions of the intelligent agent system to: if the intelligent agent responds to a target task, evaluate the execution process of the intelligent agent responding to the target task, wherein the execution process of the target task involves the invocation of a skill; in response to a first evaluation result satisfying a skill defect condition, repair a first skill from a first skill library through a first self-evolutionary link, wherein the first skill corresponds to a skill; in response to a second evaluation result satisfying a skill missing condition, create a second skill through a second self-evolutionary link and store it in the first skill library; wherein the second skill corresponds to a skill. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application; Figure 2 This is an optional flowchart illustrating the automatic skill evolution method provided in this application embodiment; Figure 3 This is a flowchart illustrating the skill self-evolution method provided in the embodiments of this application. Detailed Implementation

[0017] To more clearly illustrate the purpose, technical solutions, and advantages of the embodiments of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the following description of the embodiments is intended to explain and illustrate the overall concept of the embodiments of this application, and should not be construed as a limitation of the embodiments of this application. In the specification and drawings, the same or similar reference numerals refer to the same or similar parts or components. For clarity, the drawings are not necessarily drawn to scale, and some well-known parts and structures may be omitted from the drawings.

[0018] In some embodiments, unless otherwise defined, the technical or scientific terms used in the embodiments of this application shall have the ordinary meaning understood by one of ordinary skill in the art to which the embodiments of this application pertain. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The word "a" or "an" does not exclude multiple components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," "right," "top," or "bottom" are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes. When an element such as a layer, film, region, or substrate is referred to as being "above" or "below" another element, the element may be "directly" located "above" or "below" the other element, or there may be intermediate elements present.

[0019] To address the problems existing in related technologies, this application proposes an automatic skill evolution method. If an agent responds to a target task, the execution process of the agent's response to the target task is evaluated, and the execution process of the target task involves the invocation of a skill. If the first evaluation result satisfies the skill defect condition, the first skill from the first skill library is repaired through the first self-evolution link, and the first skill corresponds to the skill. If the second evaluation result satisfies the skill missing condition, the second skill is created through the second self-evolution link and stored in the first skill library, and the second skill corresponds to the skill.

[0020] This application's embodiments introduce an evaluation mechanism for the agent's execution process, which can accurately distinguish between two different skill problems: skill defects and skill deficiencies. It also employs different self-evolutionary paths, enabling the agent system to supplement its capabilities more flexibly and efficiently according to actual task requirements. This achieves intelligent and efficient evolution of the skill library, reduces resource waste and potential conflicts caused by directly acquiring new skills from the cloud when skill problems arise, reduces dependence on cloud-based skill libraries, and improves the accuracy of the agent's task response.

[0021] The automatic skill evolution method provided in the embodiments of this application can be executed by an intelligent agent system, which is stored in the storage device of the terminal device.

[0022] Figure 1 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application. Figure 1The terminal device 10 shown may include at least one processor 110, a storage device 150, at least one network interface 120, and a user interface 130. The various components of the intelligent agent system are coupled together via a bus system 140. It is understood that the bus system 140 is used to implement communication between these components. In addition to a data bus, the bus system 140 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 1 The general labeled all buses as Bus System 140.

[0023] The processor 110 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0024] Here, the processor can execute executable instructions of the intelligent agent system to implement the automatic skill evolution method provided in the embodiments of this application.

[0025] User interface 130 includes one or more output devices 131 that enable the presentation of media content, and one or more input devices 132.

[0026] Storage device 150 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage devices, hard disk drives, optical disk drives, etc. Storage device 150 may optionally include one or more storage devices physically located remote from processor 110. Storage device 150 may include volatile memory or non-volatile memory, or both. Non-volatile memory may be read-only memory (ROM), and volatile memory may be random access memory (RAM). The storage device 150 described in this application embodiment is intended to include any suitable type of memory. In some embodiments, storage device 150 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as exemplarily described below.

[0027] Operating system 151 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 152 is used to reach other computing devices via one or more (wired or wireless) network interfaces 120, exemplary network interfaces 120 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc. The input processing module 153 is used to detect one or more inputs or interactions from one or more input devices 132.

[0028] In some embodiments, the automatic skill evolution method provided in this application can be implemented through an intelligent agent system. Figure 1 An agent system 154 stored in a storage device 150 is shown. The system includes: a first self-evolutionary link module 1541 with the ability to repair skills; a second self-evolutionary link module 1542 with the ability to create skills; an evaluation module 1543 for evaluating the agent's execution process in response to a target task, wherein the execution process of the target task involves the invocation of a skill; and a control module 1544 for repairing a first skill from a first skill library via the first self-evolutionary link module in response to a first evaluation result satisfying a skill deficiency condition, wherein the first skill corresponds to a skill; and for creating a second skill via the second self-evolutionary link module and storing it in the first skill library in response to a second evaluation result satisfying a skill missing condition, wherein the second skill corresponds to a skill.

[0029] The intelligent agent system 154 can be software in the form of programs and plug-ins. These modules can be logical and therefore can be arbitrarily combined or further broken down according to the functions they implement. The functions of each module will be described below.

[0030] In other embodiments, the intelligent agent system 155 provided in this application embodiment can also be implemented in hardware. As an example, the intelligent agent system 155 provided in this application embodiment can be a processor in the form of a hardware decoding processor, which is programmed to execute the skill automatic evolution method provided in this application embodiment. For example, the processor in the form of a hardware decoding processor can adopt one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0031] In some embodiments, the intelligent agent system 155 may further include a communication device for communicating with a second skill library to obtain skills from the second skill library; the second skill library is a cloud-based skill library.

[0032] Here, the communication device is a hardware or software combination that enables the intelligent agent system to exchange data with external networks, providing the intelligent agent system with the ability to establish connections and transmit data between the intelligent agent system and the cloud-based skill library.

[0033] In some embodiments, the communication device may be a network interface card, such as an Ethernet card or a wireless local area network (Wi-Fi) module, which connects to the Internet via wired or wireless means; or, the communication device may be a cellular communication module, such as a 4G or 5G module, which transmits data via a mobile communication network.

[0034] Communicating with the second skill library refers to the process by which an intelligent agent system establishes a data connection with the second skill library through a communication device and interacts with it, enabling the intelligent agent system to access and obtain the skill information stored in the second skill library. For example, the intelligent agent system can send skill requests to the second skill library and receive data through standard network protocols such as HTTP / HTTPS or TCP / IP; alternatively, the intelligent agent system can also interact with data through specific application programming interfaces (APIs), such as calling software development kits (SDKs) provided by cloud service providers. Obtaining skills from the second skill library refers to the process by which the intelligent agent system retrieves and receives skill definitions, metadata, or executable code from the second skill library.

[0035] A cloud-based skill library is used to provide external skills to intelligent agent systems for reuse or integration. Intelligent agent systems can retrieve and download matching skills from the second skill library through a query interface, based on criteria such as skill name, function description, or tags; alternatively, they can automatically receive relevant skill information when the second skill library is updated or new skills are added, via a subscription mechanism.

[0036] The second skill library, a cloud-based skill library, can refer to a skill library deployed on a cloud computing platform and providing services to the outside world through the network. The cloud-based skill library can be deployed in a distributed database or object storage service on a public cloud service to store a large amount of skill data; or, the cloud-based skill library can also be maintained by a dedicated skill platform provider and provide skill services to the outside world through API interfaces.

[0037] The technical solution of this application will now be described in detail with reference to the accompanying drawings.

[0038] Figure 2This is an optional flowchart illustrating the automatic skill evolution method provided in this application embodiment, such as... Figure 2 As shown, the executing entity of the automatic skill evolution method is an intelligent agent system. The automatic skill evolution method provided in this application embodiment can be implemented through steps S201 to S203: S201, If ​​the agent responds to the target task, evaluate the agent's execution process of responding to the target task, the execution process of the target task involves the invocation of a skill.

[0039] In this embodiment, an intelligent agent can refer to a software or hardware entity capable of perceiving the environment, having autonomous decision-making and learning abilities, performing tasks, and possessing the ability to evolve skills, such as an intelligent assistant. In an intelligent agent system, the intelligent agent has the ability to perform various tasks. The target task can refer to a specific job or instruction that the intelligent agent is required to complete, such as a user's request to the intelligent agent to "provide an interactive mind map."

[0040] A skill can be the capability required by an intelligent agent to perform a target task. Each skill encapsulates the logic and data for completing a specific task or operation. The intelligent agent responds to the target task by invoking one or more skills. Skill invocation refers to the process by which the intelligent agent runs the corresponding skill according to the task requirements during the execution of the target task.

[0041] In some embodiments, starting from the moment the agent responds to the target task, the agent invokes local skills to begin execution. The execution process of the agent responding to the target task can be evaluated. The execution process can include data such as the APIs invoked, intermediate results generated, time consumed, and error codes encountered, to obtain an evaluation result.

[0042] Here, evaluating the execution process determines the skill's performance in executing the target task. For example, the skill might be invoked but fail during execution, or the skill might be required, but the agent uses an alternative path to complete the task. Based on the evaluation of the execution process, it can be determined whether the skill is defective or the agent lacks the skill.

[0043] In some embodiments, the evaluation of the execution process may include at least: evaluating whether all necessary intermediate products have been generated, whether the logical connections between steps are smooth, whether there are any logical gaps, whether the execution process complies with preset security rules or business constraints, and whether there are any redundant steps.

[0044] S202, in response to the first evaluation result satisfying the skill defect condition, the first skill from the first skill library is repaired through the first self-evolution link, and the first skill corresponds to the skill.

[0045] In this embodiment, the first evaluation result can be a judgment criterion obtained by the agent after analyzing its own execution of the target task. It characterizes whether the currently used skill meets the task requirements and what type of problem exists. The skill defect condition characterizes that when the agent executes the target task, although the skill is invoked, the skill fails to effectively or correctly complete the corresponding function, resulting in poor task execution or failure. For example, the target task is to reserve a meeting room at 9:00 PM. The agent finds the skill "meeting room reservation" in the first skill library and successfully invokes it. However, during execution, the skill fails to correctly identify the available time of the meeting room, resulting in reservation failure. At this time, the agent evaluates the execution process of the target task. The evaluation result shows that although the skill was invoked, the task was not successfully completed or the completion quality was poor. Therefore, the first evaluation result is obtained, which satisfies the skill defect condition, indicating that the "meeting room reservation" skill has a defect.

[0046] The first self-evolutionary link can refer to an automated process for handling skill deficiencies. When a skill deficiency is detected, the link is activated to repair and optimize existing skills.

[0047] In some embodiments, when the evaluation results find that a skill has defects, such as a skill logic error or improper parameter configuration, the first self-evolutionary link will be activated. This link can modify the skill code or configuration to repair the skill. In this way, the existing but problematic first skills in the first skill library can be optimized to better meet the task requirements.

[0048] Here, the first skill can be the skill actually invoked by the agent when performing the target task. The first skill library can be the agent's local skill library, which is a set of skills maintained by the agent system itself, and can be stored on the device where the agent resides or on a local server.

[0049] S203, in response to the second evaluation result satisfying the skill deficiency condition, a second skill is created through the second self-evolutionary link and stored in the first skill library; the second skill corresponds to the skill.

[0050] Here, the second evaluation result is similar to the first evaluation result, representing the judgment basis obtained by the agent after analyzing its own execution of the target task. The skill deficiency condition is used to characterize that when the agent executes the target task, there is no skill in its local skill library that can respond to the current task requirements. For example, if the target task is to query the weather for the next week, the agent searches for a skill in the first skill library that can respond to the "query the weather" task, but finds no matching skill in the first skill library. At this time, the agent evaluates the execution process of the target task, and the evaluation result shows that no usable skill was found to respond to this task, thus obtaining the second evaluation result. This second evaluation result satisfies the skill deficiency condition, indicating that the "query the weather" skill is missing from the first skill library.

[0051] The second self-evolutionary link can refer to an automated process for handling skill gaps. When a skill gap is detected, the link is activated to create new skills to meet task requirements.

[0052] When the evaluation results find that the local skill library lacks the skills required to complete the target task, the second self-evolutionary link will be activated. This link can automatically create skills and add the created skills to the first skill library. In this way, the capabilities of the local skill library can be automatically supplemented, enabling the agent to handle task types that were previously unresponsive.

[0053] This application's embodiments introduce an evaluation mechanism for the agent's execution process, which can accurately distinguish between two different skill problems: skill defects and skill deficiencies. It also employs different self-evolutionary paths, enabling the agent system to supplement its capabilities more flexibly and efficiently according to actual task requirements. This achieves intelligent and efficient evolution of the skill library, reduces resource waste and potential conflicts caused by directly acquiring new skills from the cloud when skill problems arise, reduces dependence on cloud-based skill libraries, and improves the accuracy of the agent's task response.

[0054] In some embodiments, when an intelligent agent performs a target task, it may invoke multiple skills, each of which can undergo skill self-evolution using the skill automatic evolution method provided in this application embodiment. This application embodiment uses one skill as an example to explain the skill automatic evolution method provided in this application embodiment, and does not limit the scenario to only invoking one skill when the intelligent agent performs a target task.

[0055] In some embodiments, the skill deficiency condition includes: a first score greater than a first threshold and a first difference between the first score and a second score to characterize a skill deficiency.

[0056] Here, the skill defect condition is used to determine whether the skill invoked by the agent when performing the target task is insufficient in performance, accuracy or applicability, although it exists. Its function is to serve as the basis for triggering the first self-evolutionary link repair skill.

[0057] Skill deficiency criteria can be based on preset rules. When a specific indicator in the evaluation result (such as success rate, accuracy rate, user satisfaction, etc.) is lower than a certain standard and differs significantly from the indicator of skill deficiency, it is judged as a skill deficiency. Alternatively, machine learning models can be used to identify poor performance in the skill execution process by training a large amount of historical data and classify it as a skill deficiency.

[0058] The first score is a quantitative indicator representing the probability or severity of defects in the skills invoked by the agent during the execution of the target task. It is compared with a first threshold and a second score to determine whether the skill meets the condition for a skill defect. The first threshold is a preset critical value used to determine whether the first score has reached the threshold for triggering a skill defect, thus reducing the frequent triggering of repair operations due to minor defects. The first threshold can be set based on experience, determined according to the importance and fault tolerance of different skills; or it can be obtained through historical data analysis, statistically analyzing the distribution of first scores of skills under normal operating conditions, and using the boundaries of outliers or low-performance intervals as the first threshold.

[0059] Here, the first difference can also be set empirically to quantify the difference between the first and second scores, in order to distinguish between skill deficiencies and skill gaps. When the first score is significantly higher than the second score (reaching the first difference), it tends to be judged as a deficiency; conversely, when the second score is significantly higher than the first score, it tends to be judged as a gap.

[0060] In some embodiments, the first score can be calculated based on the success rate of skill execution, the number of errors during execution, and the difference between the output and the preset result. For example, the more errors during execution, the higher the first score. Alternatively, it can be generated by combining user feedback on the skill execution result (such as satisfaction rating and complaint rate). The lower the user satisfaction, the higher the first score.

[0061] In some embodiments, the skill deficiency condition includes: a second score greater than a second threshold, and the second score being higher than the first score by a first difference to characterize skill deficiency.

[0062] Here, the second score is a quantitative indicator that represents the possibility or severity of the lack of required skills in the process of the agent performing the target task. It is compared with the second threshold and the first score to determine whether the skills meet the skill deficiency conditions.

[0063] The second score can be calculated based on indicators such as skill hit rate, task completion rate, and the matching degree between user requests and existing skills. For example, a low skill hit rate results in a high second score. Alternatively, it can be calculated by analyzing the coverage of user intent and the first skill library. When user intent cannot be effectively covered by the first skill library, the second score will be increased accordingly.

[0064] The skill deficiency condition is used to determine when an agent lacks a skill in its first skill set to effectively respond to or complete a target task. This serves as the basis for triggering the second self-evolutionary link (skill creation). This condition can be based on semantic matching or keyword analysis; if the user's intent cannot be effectively matched with any skill in the local skill set, it is determined to be a skill deficiency. Alternatively, a deep learning model can be used to predict the required skill type by analyzing the context and historical behavior of the user's request, and then compare it with the existing skill set; if no match is found, it is determined to be a skill deficiency.

[0065] This application embodiment further provides a user confirmation mechanism. When the difference between the first score and the second score is less than the first difference, i.e., insufficient to clearly distinguish between a skill deficiency or a skill lack, the user can intervene to make a manual judgment and decision. This improves the accuracy of the judgment, avoids the system making incorrect decisions automatically, and provides users with the flexibility to make their own choices. The user confirmation mechanism may display the current evaluation results (first score, second score, and their difference) to the user and provide options such as "fix skill," "create skill," or "do not process for now" for the user to choose from.

[0066] This application embodiment distinguishes between two different situations—skill defects and skill absence—through quantitative scoring rules, ensuring the accuracy of skill evolution direction and providing accurate judgment basis for subsequent selection of corresponding self-evolution links. This ensures that the automatic skill evolution process can be correctly triggered, reduces the error repair of non-existent skills or the error creation of defective skills, thereby reducing unnecessary waste of processing and storage resources.

[0067] In some embodiments, the automatic skill evolution method may further include step S1: S1, If ​​the agent responds to the target task, synchronously collect the execution trajectory of the agent's response to the target task.

[0068] In some embodiments, synchronous acquisition refers to the real-time and parallel collection of data during the execution process while the agent is responding to the target task, providing a complete and accurate data foundation for subsequent evaluation and skill evolution. In this way, synchronous acquisition ensures that all key operations, state changes, and interaction information are recorded in real time and completely throughout the entire process of the agent processing the target task, forming an execution trajectory reflecting the execution status.

[0069] The execution trajectory can be a detailed record of the agent's execution process, which may include the user's original request, contextual information (such as historical dialogues and user profiles), user expected output (the output result the user hopes for), candidate skill matching results, actual skill invocation, tool invocation process, execution status, error logs, final output, explicit user feedback, and implicit user feedback.

[0070] This application embodiment solves the problems of inaccurate evaluation and poor reliability of skill evolution results caused by missing or incomplete data in related technologies by synchronously collecting the execution trajectory when the agent responds to the target task. Based on the execution trajectory, it is possible to more accurately determine whether the existing skills have defects or whether new skills need to be created.

[0071] In some embodiments, the process of evaluating the agent's response to the target task in step S201 can be implemented through step S2011: S2011, based on the execution trajectory, evaluate the execution process of the agent's response to the target task and obtain a first score and a second score. The first score represents the score of skill deficiency and the second score represents the score of skill lack.

[0072] In some embodiments, obtaining the first score and the second score can refer to quantitatively scoring the degree of skill deficiency and skill lack exhibited by the agent during execution. The first score represents the score for skill deficiency, which can reflect errors, inaccuracies, or other problems that occur during skill execution, or situations where the skill fails to fully meet user needs or task requirements. The second score represents the score for skill lack, used to measure whether the agent fails to invoke the required skills or fails to cover all task requirements when performing the target task. This score can reflect situations where the agent fails to invoke any skills or invokes irrelevant skills during task execution, or situations where, although skills are invoked, they fail to cover all task requirements, resulting in functional gaps.

[0073] This application uses the execution trajectory as the direct basis for evaluation and outputs two different types of quantitative scores, which reduces the ambiguity and inaccuracy in classifying skill problems, ensures the accuracy of subsequent skill evolution direction judgment, and significantly improves the efficiency and reliability of automatic skill evolution of the agent.

[0074] In some embodiments, the repair of the first skill from the first skill library via the first self-evolutionary link in step S202 can be achieved through step S2021: S2021, based on the first part of the execution trajectory data, determine the cause of the defect, fix the first skill as the third skill, and store the third skill in the first skill library to replace the first skill.

[0075] In this embodiment of the application, the first part of the data may refer to the part of the execution trajectory used to analyze the defects of the first skill, which may include error logs generated by the first skill during execution or negative feedback records of the user on the execution result of the first skill.

[0076] Defect causes can refer to analyzing the first part of the data to identify the reasons why the first skill failed to effectively complete the task or produced adverse results. This can be achieved through a pre-set rule engine; for example, if the font in the output is not the desired font, it is determined as a font matching error. Alternatively, a machine learning model can be used to automatically identify defect causes by training on historical defect data. Defect causes can include description defects, routing defects (i.e., failure to accurately assign tasks to the most suitable skill), parameter defects (errors in the process of extracting or validating parameters), dependency defects, execution defects (i.e., malfunctions occurred during the actual execution of the triggering tool or code), compatibility defects (i.e., incompatibility issues when the agent interacts with different systems, platforms, hardware, or data formats), product quality defects, and security defects, etc.

[0077] Here, after determining the cause of the defect, the first skill is modified and optimized to become the improved third skill. For example, if the defect is caused by a font mismatch, the fix might involve modifying the copper in the skill template; if the defect is a device performance issue, the fix might include adjusting the resource allocation strategy.

[0078] After obtaining the third skill, you can save the repaired third skill to your local skill library and replace the original first skill with the repaired third skill. This can be done by overwriting the file to ensure that subsequent tasks use the improved third skill.

[0079] In some embodiments, the creation of the second skill via the second self-evolutionary link in step S203 can be achieved through step S2031: S2031, Create a second skill based on the second part of the execution trajectory data.

[0080] The second part of the data can refer to the data in the execution trajectory used to identify skill deficiencies and guide the creation of new skills. For example, this part of the data can include user instructions that the agent failed to respond to, the agent used alternative paths to complete the task, or the user explicitly stated "I do not have this ability" or "I need to install the relevant skill". The second part of the data can also include the skill description and execution logic of the missing skill.

[0081] Creating a second skill can be based on data such as skill descriptions and execution logic. For example, if the second set of data indicates that a user needs a "schedule reminder" function that is not available in the first skill library, a second skill with schedule reminder functionality can be automatically generated based on the user's intent, skill description, and execution logic. This can be done by selecting from a predefined template and filling in the parameters, or by using natural language processing technology to transform the user's needs into executable skill logic.

[0082] This application's embodiments utilize execution trajectory data generated by the intelligent agent during the execution of a target task to accurately pinpoint the causes of deficiencies in existing skills and address them accordingly, thereby ensuring that the repaired skills truly adapt to task requirements. Simultaneously, for skill gaps, new skills can be created based on capability gaps exposed in actual tasks, making the generation of new skills more practical and targeted. This avoids creating skills divorced from actual needs, improves the effectiveness of the intelligent agent's skill evolution, and ensures that the intelligent agent can continuously optimize its skill capabilities to better complete various tasks.

[0083] In some embodiments, the creation of the second skill via the second self-evolutionary link in step S203 can also be achieved through steps S2032 to S2036: S2032, Generate description information of the second skill based on the second part of the execution trajectory data.

[0084] The second part of the execution trajectory data can include at least user intent, contextual information, user input, system feedback, and subsequent user actions related to the missing skill, providing a precise basis for the subsequent creation or reuse of skills.

[0085] The description information can be a structured or unstructured text description of the required secondary skill functions, inputs, outputs, and invocation methods. For example, the description information may include the skill name, applicable scenarios, task type tags (such as document processing, data analysis, content generation, etc.), a list of dependent tools, input and output format specifications, core execution process steps, constraints, capability boundaries, expected results, and quality requirements.

[0086] It can be achieved by using natural language processing (NLP) technology to perform semantic analysis, keyword extraction, and intent recognition on the second part of the execution trajectory data, thereby automatically generating a text description of the second skill; it can also be achieved by using a preset template and combining key information in the trajectory data to fill in structured descriptive information; or it can be achieved by using a Few-Shot information extraction model to extract structured capability requirement features from the execution trajectory.

[0087] S2033, Matching similarity between descriptive information and skills in the second skill library.

[0088] The second skill library can be a cloud-based skill library, which can be a collection of skill resources deployed on a cloud server. It includes at least several skills that can be downloaded and invoked by the agent. These skills can be skills uploaded by users for different tasks, as well as all currently open-source skills and basic skills.

[0089] Similarity matching can refer to comparing generated second skill descriptions with existing skill descriptions in a second skill database to quantify their functional or semantic similarity. Matching methods can employ keyword-based matching algorithms; vector space model-based algorithms, such as cosine similarity, which converts descriptions into vector representations for comparison; or deep learning models, such as BERT, for semantic similarity calculation.

[0090] S2034, obtain the fourth skill where the similarity meets the similarity threshold.

[0091] Here, the similarity threshold can be a value set based on actual scenarios and experience, used to determine whether the similarity between two skill descriptions is high enough to be considered functionally similar. The fourth skill refers to an existing skill in the second skill library whose description, after similarity matching, has a similarity to the description of the second skill to be created that is greater than or equal to the preset similarity threshold.

[0092] During the matching process, fourth skills that meet the preset similarity threshold can be selected as candidates to ensure that the identified cloud skills have a sufficiently high functional relevance to the needs of locally missing skills.

[0093] S2035, if the fourth skill does not meet the reuse conditions, create a second skill.

[0094] In some embodiments, reuse criteria are used to determine whether an existing fourth skill can be directly used as a missing second skill. Reuse criteria may include whether the functionality of the fourth skill fully covers the requirements of the second skill, whether the interface of the fourth skill is compatible with the local system, whether the performance indicators of the fourth skill (such as response time and success rate) meet the requirements, and whether the fourth skill is authorized for local use.

[0095] If the fourth skill does not meet these reuse conditions, it means that there is no directly reusable skill in the cloud skill library. In this case, a new second skill can be created from scratch based on the previously generated description information and the second part of the data.

[0096] S2036, if the fourth skill meets the reuse conditions, the fourth skill is used as the second skill; where the second skill library is the cloud skill library and the first skill library is the local skill library.

[0097] In some embodiments, if the fourth skill meets the reuse criteria, using the fourth skill as the second skill can mean not creating a new skill, but directly importing the fourth skill into the first skill library and using it as the currently missing second skill. Here, the second skill library is a cloud-based skill library, a remote service platform that centrally stores and manages a large number of general and mature skills. The first skill library is a local skill library, a set of skills maintained by the intelligent agent system itself, typically stored on the device where the intelligent agent resides or on a local server.

[0098] This application embodiment, by performing similarity matching with a cloud-based skill library and setting a similarity threshold, can efficiently identify existing cloud-based skills with similar functions, making full use of the rich skill resources in the cloud, reducing unnecessary redundant development, and saving costs and time. By introducing reuse conditions to rigorously evaluate similar skills and deciding whether to reuse cloud-based skills or create new skills based on the evaluation results, the decision-making is automated, reducing erroneous decisions caused by subjective judgment, and realizing the intelligent and efficient evolution of the agent's skill library.

[0099] In some embodiments, the repaired third skill and the newly created second skill are stored in the local skill library without verification. This may result in the repaired skill still not meeting the requirements of the target task, or the newly created skill itself having runtime errors or functional issues. Direct storage may cause the agent to fail to execute or produce unsatisfactory results when calling the skill later, affecting the stability and reliability of the agent's task execution.

[0100] Therefore, the first self-evolutionary link also includes: automatically executing tests on the third skill before storing the third skill in the first skill library.

[0101] Here, the automatic execution of the third skill ensures that its functionality and performance meet expectations before being stored in the first skill library, after being repaired by the first self-evolutionary link. The target task can be executed again based on the third skill to determine whether the output meets expectations and whether any anomalies occur during execution. This comprehensive evaluation of the third skill's stability and performance verifies whether the repaired third skill truly addresses the original defects and can operate normally and meet the requirements of the target task.

[0102] Only when the third skill passes the automatic execution test and proves that the function is complete and stable, will it be stored in the first skill library to replace the original first skill.

[0103] In some embodiments, the second self-evolutionary link further includes: performing an automated test on the second skill before storing the second skill in the first skill library.

[0104] Automated execution testing of the second skill is used to verify the functional correctness and stability of the newly created second skill through the second self-evolutionary link before it is stored in the first skill library. The actual output can be compared with the expected output for the target task to be solved by the second skill to determine whether the newly created second skill can execute normally, whether its function meets expectations, and whether it can effectively handle the target task.

[0105] The second skill will only be stored in the first skill library once it passes the automated execution test.

[0106] This mechanism of mandatory testing before skill storage is closely integrated with the overall process of agent evaluation and execution, and skill repair or creation through self-evolutionary links, forming a closed-loop quality assurance system.

[0107] The embodiments of this application ensure that both repaired and newly created skills are fully verified before entering the local skill library, thereby improving the quality of skills in the local skill library after the agent completes automatic skill evolution. This enhances the stability and reliability of the agent in subsequent task execution and effectively reduces the probability of errors in subsequent task execution.

[0108] In some embodiments, step S2011 can be implemented by steps S11 to S12: S11, based on the execution trajectory, evaluate at least one of the following during the execution process of the agent's response to the target task: skill hit rate, user feedback information, and execution result quality.

[0109] In some embodiments, skill hit rate refers to the degree of matching between the skills invoked by the agent and the task requirements when performing a target task. Skill hit rate can be obtained by analyzing skill invocation logs recorded in the execution trajectory. It can include features representing skill deficiencies such as no candidate skill hits, the actual execution path containing obvious degradation, multiple failures in similar historical tasks without stable skill hits, and features representing skill defects such as the presence of candidate skills with high coverage. Semantic analysis techniques can also be used to assess the semantic relevance between the invoked skills and the target task intent to determine whether the skills accurately respond to the task requirements.

[0110] User feedback refers to a user's subjective evaluation of the agent's performance of a target task. This information can be obtained through various means. The agent can proactively initiate a satisfaction survey after the task is completed, collecting user ratings or written comments on the task's execution. Alternatively, it can monitor the user's natural language expressions during interaction with the agent in real time, identifying positive or negative emotions, as well as correction instructions or supplementary explanations requested by the user, using this as the basis for user feedback. For example, explicit statements from users such as "I don't have this ability" or "I need to install related skills" represent a lack of skill, while multiple corrections to the same skill output by the user, or the agent's output failing acceptance, represent skill deficiencies.

[0111] Execution result quality refers to the objective effect and performance of an agent in completing a target task, and different indicators can be set according to the task type. For example, characteristics representing skill deficiencies include inconsistencies between the output type and the user's goal, and characteristics representing skill defects include anomalies after invocation, unmet dependencies, incorrect parameter mapping, abnormal output format, cross-environment compatibility failures, and historical records showing an increased failure rate of the skill on similar tasks. Execution result quality can be measured through preset rules or automated testing.

[0112] Here, skill hit rate reflects the matching degree between skill invocation and task intent, as well as the success rate of execution; user feedback reflects the user's subjective satisfaction with task execution; and the quality of execution results objectively measures the effectiveness of task completion. By analyzing information from these three dimensions, we can analyze the performance of skills in practical applications from multiple perspectives.

[0113] In this embodiment of the application, after the agent responds to the target task, the execution process is collected in real time, and the skill hit rate, user feedback information and execution result quality during the execution process are evaluated. Based on at least one of these, a first score representing skill defects and a second score representing skill deficiencies are calculated.

[0114] S12, determine the first score and the second score based on at least one of the following: skill hit rate, user feedback information, and execution result quality.

[0115] In some embodiments, different weights can be assigned to features in the three dimensions of skill hit rate, user feedback information, and execution result quality. Then, a first score is obtained by weighted summation of features representing skill defects (such as multiple corrections of the same skill output by the user or failure of the product to pass acceptance) and a second score is obtained by weighted summation of features representing skill deficiencies (such as the existence of a skill that was not called but the user had a clear need or the user reported that the task was not completed).

[0116] In some embodiments, a base score (e.g., between 0 and 10) can be assigned to each feature to represent the severity of the feature's occurrence. Then, based on the importance of different features, weights are assigned to each feature, with a total weight of 1. The first score can be a weighted sum of each feature indicating a deficiency in the current target task as identified through the execution process analysis, along with its corresponding weight.

[0117] In some embodiments, a machine learning model can be used, taking at least one of the skill hit rate, user feedback, and execution result quality as input features, and directly outputting a first and second score through a pre-trained classification model (e.g., a Few-Shot classifier). For example, a classification model can be trained to predict whether a skill has a defect or is missing, and then a regression model can be trained to output specific score values, i.e., the first and second scores. The training data for the model can be the agent's historical task execution data and manually labeled skill defects / missing information.

[0118] The agent determines whether to activate the first self-evolutionary link to repair the skill or activate the second self-evolutionary link to create a new skill based on whether the first and second scores meet the preset skill defect conditions or skill absence conditions.

[0119] This application embodiment evaluates the intelligent agent's response to the target task based on the execution trajectory of the intelligent agent, from multiple dimensions such as skill hit status, user feedback information and execution result quality. This reduces the limitations of relying on a single or fuzzy indicator for evaluation, and can effectively avoid skill status judgment errors caused by scoring bias. This ensures the accuracy and reliability of the subsequent skill evolution process (whether it is repairing existing skills or creating new skills), and significantly improves the adaptive capability and overall performance of the intelligent agent's skill library.

[0120] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0121] To address the problems existing in related technologies, this application provides a dual-channel skill evolution method, which enables the intelligent agent system to determine whether the current skill problem is a skill deficiency or a skill defect based on the actual execution trajectory, and to perform new skill creation or existing skill repair respectively, thereby improving the efficiency of skill evolution and reducing the risk of duplicate skill generation and error repair.

[0122] This application embodiment adds a skill problem triage layer before skill evolution. This triage layer calculates a skill deficiency score (i.e., a second score) and a skill defect score (i.e., a first score) based on the task trajectory. The skill deficiency score indicates whether there are effective skills in the system that can meet the current user's goals; the skill defect score indicates that there are relevant skills in the system, but the skills have problems in terms of invocation, dependency, parameters, execution, product quality, or compatibility.

[0123] Based on the relative relationship between the two scores and threshold conditions, different evolutionary channels are entered. Among them, the missing skill channel (i.e., the second self-evolutionary link) generates new skill requirements, skill names, capability boundaries, input and output specifications, implementation suggestions, and verification tasks; the defective skill channel (i.e., the first self-evolutionary link) locates existing skills, determines the defect type, generates repair patches, updates dependencies, or modifies the calling strategy; the uncertain channel, when the two scores are close, first conducts small-scale probing verification or requests manual confirmation to avoid misoperation.

[0124] The intelligent agent provided in this application embodiment may include a trajectory acquisition module, a skill profiling module, a task intent parsing module, a skill coverage calculation module, an execution anomaly analysis module, a dual-score judgment module, a regression verification module, and a publishing and monitoring module.

[0125] The system comprises several modules: **Trajectory Acquisition Module:** Collects the complete trajectory of a user task execution, including user requests, context files, task objectives, candidate skills, actual skill calls, tool call sequences, execution logs, error messages, output artifacts, and user feedback. **Skill Profile Module:** Maintains the capability description, applicable scenarios, input / output formats, dependencies, call parameters, historical success rate, failure type, and version information for each skill in the skill library. **Task Intent Parsing Module:** Extracts the target task, expected artifact type (e.g., text, JSON, code, charts, or clickable cards), input material type (e.g., user-provided files), constraints (e.g., user requirements), and acceptance criteria (e.g., what output constitutes fulfilling the user's requirements) from the user request and context. **Skill Coverage Calculation Module:** Calculates the matching degree between the user task and existing skills, obtaining a candidate skill set and coverage score. **Execution Anomaly Analysis Module:** Analyzes anomalies in the skill call chain, including missed skills, missing parameters, missing dependencies, runtime errors, format errors, artifacts not meeting expectations, repeated user corrections, and degraded execution. The dual-scoring judgment module calculates skill deficiency scores and skill defect scores separately, and determines the evolution path based on the scoring relationship. The evolution path includes a missing skill creation path and a defective skill repair path. The missing skill creation path generates new skill proposals, skill specifications, implementation plans, and verification tasks when a skill is determined to be missing. The defective skill repair path identifies the target skill and defect type when a skill is determined to be defective, generates a repair plan, and updates existing skills. The regression verification module converts original failure trajectories and similar task trajectories into regression verification tasks to verify whether newly created or repaired skills truly solve the problem. The release and monitoring module performs canary releases, effect monitoring, and automatic rollback of verified skill versions.

[0126] The dual-channel skill evolution method provided in this application embodiment can be implemented through steps one to twelve: Step 1: Collect the task trajectory.

[0127] When a user submits a task to the intelligent agent, the system records the complete execution trajectory of this task. The trajectory includes at least the user's original request, contextual materials (such as historical dialogues and user profiles), the user's expected output (the output result the user hopes for), candidate skill matching results, actual skill invocation, tool invocation process, execution status, error logs, final output, explicit user feedback, and implicit user feedback.

[0128] Step two: Analyze the user's target task.

[0129] The system parses the task objective based on the user's request and the contextual information. For example, if the user requests "generate an editable mind map," the target output type is "visual mind map," and the output attributes are "editable, exportable, and structured hierarchical." If the system ultimately only generates a plain text list or a non-editable image, it can be identified as a downgraded execution method.

[0130] Step 3: Calculate skill coverage.

[0131] The system matches the user's target task with the ability profiles of each skill in the skill library to obtain a set of candidate skills. Matching factors include skill name similarity, ability description similarity, input / output format matching, dependency condition satisfaction, and historical task similarity. If no candidate skill with coverage exceeding the first threshold is found, the skill missing score is increased.

[0132] Step 4: Analyze the skill activation status.

[0133] If a candidate skill exists but is not invoked by the agent, the system analyzes whether the reason is an unclear skill description, insufficient triggering conditions, or an incorrect routing strategy.

[0134] If a candidate skill is invoked but fails to execute, the reasons for the failure are further extracted, including incorrect parameters, incorrect dependencies, incorrect permissions, runtime errors, incorrect output formats, and substandard output quality.

[0135] Step 5: Identify and implement downgraded execution.

[0136] The system determines whether the agent has used an alternative path to complete the task. For example, the user needs specialized mind mapping skills, but the system uses a general drawing script instead; the user needs a spreadsheet with charts, but the system only generates a plain table; the user needs an editable file, but the system outputs a screenshot or plain text.

[0137] If there is a significant gap between the actual product and the expected product, and no suitable skill is matched, then the skill deficiency score is increased.

[0138] If relevant skills exist but the product quality does not meet requirements, the skill defect score will be increased.

[0139] Step six: Calculate the skill deficit score.

[0140] The skill deficiency score is determined by the following characteristics: no candidate skills are hit, the actual execution path contains obvious downgrades, the product type is inconsistent with the user's goal, the user explicitly states "I do not have this ability" or "I need to install the relevant skill", and similar historical tasks have failed multiple times without any stable skill hits.

[0141] Step 7: Calculate the skill deficiency score.

[0142] Skill defect scores are determined by the following characteristics: the existence of candidate skills with high coverage, the skill being invoked, an exception occurring after invocation, unmet dependencies, incorrect parameter mapping, abnormal output format, failure to achieve cross-environment compatibility, failure of the deliverable to pass acceptance, multiple corrections made by the user to the same skill output, and historical records showing an increased failure rate of the skill on similar tasks.

[0143] Step 8: Perform dual-channel determination.

[0144] When the skill deficiency score is greater than the second threshold and the skill deficiency score is higher than the skill defect score by a preset difference (i.e., the first difference), the system determines that the skill is missing and enters the missing skill creation channel.

[0145] When the skill defect score is greater than the first threshold and the skill defect score is higher than the skill lack score by a preset difference, the system determines it as a skill defect and enters the defective skill repair channel.

[0146] When both scores are low or the difference is insufficient, the system enters an uncertain channel, performing additional probing, requesting user confirmation, or temporarily suspending skill evolution.

[0147] Step nine: Enter the missing skills creation channel.

[0148] The system generates new skill proposals. Each proposal includes the skill name, applicable scenarios, input format, output format, dependent resources, invocation conditions, capability boundaries, implementation suggestions, testing tasks, and expected results. The system also generates the first batch of verification tasks based on the initial failure logs to ensure that the new skills cover real user needs.

[0149] Step 10: Enter the skill defect repair channel.

[0150] The system identifies target skills and target defect types. Defect types can include description defects, routing defects (failure to accurately assign tasks to the most suitable skills), parameter defects (errors in the process of extracting or validating parameters), dependency defects, execution defects (failures during the actual execution of triggering tools or code), compatibility defects (mismatches that occur when the agent interacts with different systems, platforms, hardware, or data formats), product quality defects, and security defects.

[0151] The system generates repair solutions based on the defect type, such as updating skill descriptions, adjusting trigger conditions, supplementing parameter conversions, fixing dependency installations, modifying execution scripts, adding output validation, or restricting high-risk calls.

[0152] Step 11: Perform regression validation.

[0153] Regardless of whether the system enters the creation or repair channel, it converts the original failure trajectory into a regression verification task. Verification includes whether the task was successfully completed, whether the artifact type meets the user's objectives, whether the output file can be opened, whether key fields are complete, whether execution is stable, and whether any new failures were introduced. Only when the verification results reach the release threshold will the system allow the skill version to enter a canary release phase.

[0154] Step 12: Canary release and continuous monitoring.

[0155] The system will release newly created or repaired skills to select users or specific task scenarios, and continuously monitor success rate, user satisfaction, call failure rate, rollback rate, and number of repeated fixes. If these metrics deteriorate, the system will automatically roll back to the old version or proceed to the next round of repairs.

[0156] This application also proposes a multi-agent collaborative AI skill self-evolution system and method based on trace distillation. Through the cognitive architecture of multi-agent collaboration, the original trajectory data of the agent (i.e., intelligent agent) performing tasks is transformed into reusable and evolvable standardized skill assets, and a full life cycle automated closed loop of "collection-diagnosis-proposal-construction-evaluation-iteration" is constructed to achieve continuous exponential evolution of skill capabilities.

[0157] This method can include trajectory semantic preprocessing, which uses a trajectory acquisition and preprocessing module to perform structured parsing, noise filtering, and session segmentation on the agent's execution trajectory, generating a standardized Task View data structure. A multi-agent collaborative evolutionary architecture is designed with four specialized agents: a diagnostic agent, a proposal agent, a construction agent, and an evaluation agent. These agents collaborate to manage the entire lifecycle of skills through a standardized communication protocol. A dual-path skill generation paradigm integrates external skill market reuse with internally automatically generated paths. A four-dimensional quality gating mechanism constructs a four-dimensional evaluation system encompassing static compliance, functional correctness, performance efficiency, and environmental compatibility, ensuring the production-grade reliability of deployed skills. A distributed collaborative evolutionary protocol is designed based on federated learning, enabling cross-node skill evolution knowledge sharing while protecting user privacy, achieving a 3.7-fold increase in overall system capability evolution speed.

[0158] In the multi-agent collaborative evolutionary architecture, four types of agents—diagnostic, proposal, construction, and evaluation—collaborate through a standardized communication protocol. The diagnostic agent clusters the trajectory vectors of tasks to determine the root cause types of failures; the proposal agent proposes skill reuse or construction for newly added skills, providing preset skill proposals; the construction agent automatically completes a skill package when no reusable skills are available; and the evaluation agent evaluates newly created or repaired skills, placing them in the skill market after passing the evaluation. In this way, multi-agent collaboration forms a closed loop of intelligent evolution, enabling each agent to continuously learn and evolve through collaboration, ultimately achieving continuous autonomous skill evolution to meet the needs of more scenarios.

[0159] Figure 3 This is a flowchart illustrating the skill self-evolution method provided in this application embodiment. The skill self-evolution method can be implemented through steps S301 to S312: S301, Quality Inspection of Reused or Newly Created Skills.

[0160] If the quality inspection passes, proceed to step S302; if the quality inspection fails, proceed to step S312 to create a new skill.

[0161] S302, conducts skills assessment on skills that have passed quality inspection.

[0162] Skill execution steps S303 are passed by the skills assessment department, and skill execution steps S307 are passed by the skills assessment department.

[0163] In some embodiments, steps S301 and S302 can be implemented by an evaluation agent, which obtains the final skill that meets comprehensive conditions through multi-dimensional evaluation. The multi-dimensional evaluation may include a static inspection stage, a functional testing stage, a performance evaluation stage, and a compatibility testing stage, with quality gating set for each stage.

[0164] The testing process is divided into several phases: Static Inspection (15% weight): This phase checks format compliance, parameter completeness, dependency completeness, security vulnerabilities, and code style; a score below 80 points results in immediate elimination. Functional Testing (40% weight): This phase runs test cases in an isolated sandbox environment to evaluate trigger rate, completion rate, exception handling capabilities, and output correctness; a functional pass rate of ≥95% is required. Performance Evaluation (25% weight): This phase evaluates execution latency (≤30s), token consumption (≥30% reduction compared to the no-skill solution), and resource utilization; a performance score below 70 points results in elimination. Compatibility Testing (20% weight): This phase tests compatibility under 3 different model versions, 2 different operating environments, and 3 different tool versions; a compatibility score of ≥80 points is required to pass.

[0165] S303: Skills that have passed the skill evaluation are placed on the skill platform.

[0166] Once the evaluation is passed, a complete evaluation report will be generated, including quantitative indicators, advantages and disadvantages analysis, and deployment suggestions. Only skills with a comprehensive score of ≥85 points can be included in the skill repository or platform.

[0167] S304, the intelligent agent system performs tasks based on a skill platform.

[0168] S305, capture failed / low-scoring tasks.

[0169] The task execution trajectory of the agent can be encoded through trajectory acquisition and semantic embedding modules to obtain a normalized trajectory vector. Failed / low-scoring tasks can be captured based on the dialogue quality assessment agent.

[0170] First, end-to-end trajectory acquisition can be implemented based on the OpenTelemetry standard to collect end-to-end data points. The data dimensions collected include: session metadata, execution sequence, and feedback data.

[0171] Session metadata should include at least: model version, timestamp, environment configuration, and user characteristic tags; execution sequence should include at least: user request, thought process, tool call parameters, tool return results, error stack, and final output; feedback data should include at least: user rating, error correction information, and task completion status tags.

[0172] Secondly, the multimodal semantic embedding adopts the improved BGE-M3 model as the trajectory encoder, which integrates multimodal information such as text content, tool call type, temporal relationship, and error type into a unified 768-dimensional vector representation, namely the trajectory vector Vtrace, as shown in formula (1); where the error type is obtained from the trajectory source data, and the large model can assist in the labeling and encoding: (1); Where Ttext represents the text trajectory, Ttool represents the tool call trajectory, and Ttime represents the time sequence; the 3D encoders are the text trajectory encoder Enc and Ttool respectively. text The tool calls the trajectory encoder Enc. tool Time-sequence encoder Enc time Furthermore, α+β+γ=1, and the optimal weights are obtained through comparative learning pre-training.

[0173] Finally, privacy computation processing is performed, using differential privacy technology to desensitize sensitive information in the trajectory and to inject noise intensity.

[0174] S306, Root Cause Diagnosis Agent performs root cause classification.

[0175] In some embodiments, skill gap / redundancy execution step S307, tool execution error logic error is to repair skills.

[0176] The root cause diagnosis agent clusters the trajectory vector Vtrace to determine the type of failure root cause.

[0177] First, failure scenario clustering: After receiving feedback from users about low-scoring tasks, the HDBSCAN density clustering algorithm can be used to cluster the trajectory embedding vectors of low-scoring tasks. The minimum cluster size is set to 5, and the neighborhood distance threshold is 0.35. This can automatically identify high-frequency failure scenarios, and the clustering purity reaches 91.2%.

[0178] Then, root cause classification is performed. For each cluster of trajectory vectors, a Few-Shot classifier can be used to automatically classify the root causes of failure. First, failure scenarios are clustered, and then different types of failure scenarios are further classified into root cause categories: capability gap category (lack of corresponding skills, accounting for 63%); technical error category (tool parameter errors, network anomalies, etc., accounting for 27%); and resource insufficiency category (token exceeding limits, timeouts, etc., accounting for 10%). The classification accuracy reaches 95.7%, and only capability gap category scenarios are sent to the subsequent skill generation process.

[0179] Finally, capability gap feature extraction is performed. The Few-Shot information extraction model can be used to extract structured capability requirement features (i.e., description information of the second skill) from the failure trajectory, including: task type labels (such as document processing, data analysis, content generation, etc.), list of dependent tools, input and output format specifications, core execution process steps, constraints and quality requirements.

[0180] S307, the proposal agent provides a preset skill proposal.

[0181] The multi-agent skill proposal module proposes skill reuse or construction for newly added skills, and provides preset skill proposals. These proposals are completed by a professional proposal agent and generated using a structured thinking framework.

[0182] First, existing skills are retrieved. Based on vector similarity matching, existing skills are retrieved in the skill market. This can be done by using skill names and descriptions and using a relevance ranking algorithm (such as bm25) to retrieve skills in the skill market. The cosine similarity threshold is set to 0.75. If the value is higher than the threshold, a proposal to edit or optimize existing skills is proposed, i.e., step 309 is executed; otherwise, a proposal to add new skills is proposed, i.e., step 308 is executed.

[0183] Then, a multi-solution brainstorming session is conducted to automatically generate multiple (up to 3) different implementation schemes. These schemes are scored from three dimensions: complexity, generalization, and reusability. The scheme with the highest overall score that exceeds the score threshold is selected to reduce the dependence of the proposal effect on the proposal agent's capabilities. The overall score can be achieved using formula (2): (2); Here, Sgenerality represents the scope of skill reuse; Scomplexity represents the complexity of the internal logic of the skill; and Svalue represents the reuse value of the skill.

[0184] Finally, a structured proposal is output, generating a standardized JSON format proposal that includes at least several structured fields such as operation type, skill name, ability description, triggering conditions, input and output parameters, implementation suggestions, and expected benefits. Historical data shows that the proposal pass rate reaches 89%.

[0185] S308, new skills added.

[0186] If new skills need to be added, proceed to step S310. First, perform similarity matching in the public skill market and internal skill repository, prioritize the reuse of mature skills, and perform lightweight adaptation of the reused skills through the adapter pattern, achieving a success rate of 78%.

[0187] When no reusable skills are available, the build agent automatically generates a complete skill package. The skill package includes at least a skill description file, execution logic code, a test case set, and a configuration file. The skill description file (SKILL.md) follows a unified skill metadata (i.e., skill file) specification and includes at least trigger conditions, capability descriptions, parameter definitions, dependency declarations, and usage examples. The execution logic code is a Python / Bash script automatically generated based on the skill generation template, integrating common capabilities such as automatic retries, error handling, and logging. The test case set consists of at least 10 automatically generated test cases covering normal, boundary, and abnormal scenarios for subsequent evaluation. The configuration file includes at least permission declarations, resource limits, and environment variables.

[0188] S309, Optimized Skills.

[0189] Skills can be optimized based on their defect types, and the optimized skill executes step S301.

[0190] S310, matched in the skills market.

[0191] S311 determines whether the matched skill can be reused.

[0192] If the step can be reused, proceed to step S301; if it cannot be reused, proceed to step S312.

[0193] S312, Skill Building Agent creates skills.

[0194] For skills that cannot be reused, the skill building agent creates the skills.

[0195] This application's embodiments can also employ the federated learning concept to design a cross-node collaborative evolution protocol: Each node completes the skill evolution process locally, uploading only the skill's metadata and 768-dimensional vector embedding to the central node, without uploading the original trajectory data to protect user privacy; the central node completes global deduplication, quality assessment, and distribution of skills, and high-quality skills are synchronized to all nodes; a contribution point mechanism is adopted to incentivize nodes to share high-quality skills, forming a positive evolutionary ecosystem.

[0196] High-quality skills (overall score ≥ 80) are automatically synchronized to the skill store of all nodes, while medium-quality skills (60-79) are recommended to nodes that are suitable for the scenario. The score is given by the evaluation agent.

[0197] The central node continuously tracks actual usage data for skills, such as downloads, activation rates, user ratings, error rates, and feedback / improvement suggestions during use.

[0198] For skills that are widely used and receive positive feedback, additional contribution points will be periodically awarded to the original contributing nodes; skills with low usage rates and poor ratings will be automatically removed to free up storage resources.

[0199] It should be noted that the description of the intelligent agent system embodiments in this application is similar to the description of the method embodiments described above, and has similar beneficial effects as the method embodiments; therefore, it will not be repeated. For technical details not disclosed in the embodiments of this intelligent agent system, please refer to the description of the method embodiments in this application for understanding.

[0200] It should be noted that, in the embodiments of this application, if the above-mentioned automatic skill evolution method is implemented in the form of software functional modules and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a terminal to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0201] This application provides a storage medium storing executable instructions of an intelligent agent system. When the executable instructions of the intelligent agent system are executed by a processor, the processor will execute the automatic skill evolution method provided in this application.

[0202] In some embodiments, the storage medium may be a computer-readable storage medium, such as a ferromagnetic random access memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic surface memory, optical disc, or a compact disk-read-only memory (CD-ROM); or it may be a device that includes one or any combination of the above-mentioned memories.

[0203] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0204] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts within a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files storing one or more modules, subroutines, or code sections). As an example, executable instructions may be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0205] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application. It should be understood that "an embodiment" or "one embodiment" mentioned throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in one embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence number of the above-described processes does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments of this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments.

[0206] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not performed.

[0207] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automatic skill evolution, the method comprising: If the agent responds to a target task, the execution process of the agent's response to the target task is evaluated, the execution process of the target task involving the invocation of a skill; In response to the first assessment result meeting the skill defect condition, the first skill from the first skill library is repaired through the first self-evolutionary link, and the first skill corresponds to the skill. In response to the second assessment result meeting the skill deficiency condition, a second skill is created through the second self-evolutionary link and stored in the first skill library; The second skill corresponds to the aforementioned skill.

2. The automatic skill evolution method according to claim 1, wherein the skill defect conditions include: The first score is greater than the first threshold, and the first score is higher than the second score to reach the first difference value, which is used to characterize skill deficiencies; The skill deficiency conditions include: the second score is greater than the second threshold, and the second score is higher than the first score to reach the first difference value, which is used to characterize the skill deficiency.

3. The automatic skill evolution method according to claim 2, further comprising: If the agent responds to the target task, the execution trajectory of the agent's response to the target task is collected synchronously.

4. The automatic skill evolution method according to claim 3, wherein the process of evaluating the agent's response to the target task includes: The execution process of the agent in response to the target task is evaluated based on the execution trajectory to obtain a first score and a second score. The first score represents a score for skill deficiencies, and the second score represents a score for skill deficiencies.

5. The automatic skill evolution method according to claim 4, wherein repairing the first skill from the first skill library through the first self-evolutionary link includes: Based on the first part of the execution trajectory data, the cause of the defect is determined, the first skill is repaired into a third skill, and the third skill is stored in the first skill library to replace the first skill; The creation of a second skill through a second self-evolutionary link includes: The second skill is created based on the second portion of the execution trajectory data.

6. The automatic skill evolution method according to claim 5, wherein creating the second skill through a second self-evolutionary link comprises: The description information of the second skill is generated based on the second part of the execution trajectory data; Based on the description information, a similarity match is performed with the skills in the second skill library; To acquire a fourth skill with a similarity threshold; If the fourth skill does not meet the reuse conditions, create the second skill; If the fourth skill meets the reuse condition, the fourth skill shall be used as the second skill. The second skill library is a cloud-based skill library, while the first skill library is a local skill library.

7. The automatic skill evolution method according to claim 5 or 6, wherein the first self-evolutionary link further comprises: Before storing the third skill into the first skill library, the third skill is automatically executed for testing. The second self-evolutionary link further includes: performing an automated execution test on the second skill before storing the second skill in the first skill library.

8. The automatic skill evolution method according to claim 4, wherein evaluating the agent's response to the target task based on the execution trajectory to obtain the first score and the second score includes: Based on the execution trajectory, evaluate at least one of the following during the execution process of the agent's response to the target task: skill hit rate, user feedback information, and execution result quality; The first score and the second score are determined based on at least one of the skill hit rate, user feedback information, and execution result quality.

9. An intelligent agent system, the system comprising: The first automatic evolutionary link module has the ability to repair itself. The second automatic evolutionary link module has the ability to create skills; An evaluation module is used to evaluate the agent's execution process in response to a target task, wherein the execution process of the target task involves the invocation of a skill; The control module is used to repair the first skill from the first skill library through the first self-evolution link module in response to the first evaluation result meeting the skill defect condition, wherein the first skill corresponds to the skill. In response to the second assessment result meeting the skill deficiency condition, a second skill is created through the second self-evolutionary link module and stored in the first skill library; The second skill corresponds to the aforementioned skill.

10. A terminal device, comprising: Storage device for storing intelligent agent systems; A processor is configured to execute instructions of the agent system to: if the agent responds to a target task, evaluate the agent's execution process in response to the target task, the execution process of the target task involving the invocation of a skill; In response to the first assessment result meeting the skill defect condition, the first skill from the first skill library is repaired through the first self-evolutionary link, and the first skill corresponds to the skill. In response to the second assessment result meeting the skill deficiency condition, a second skill is created through the second self-evolutionary link and stored in the first skill library; The second skill corresponds to the aforementioned skill.

Images