Skill self-evolution method of agent system and agent system

Abstract

This application discloses a skill self-evolution method and intelligent agent system, comprising: transmitting a first skill stored in a local skill library to a second self-evolution link based on a collaborative mechanism of a first automatic evolution link and a second self-evolution link; obtaining a second skill from the second self-evolution link, wherein the second skill is a skill processed based on the second automatic evolution link; determining whether to replace a third skill that meets the similarity condition with the second skill and is located in the local skill library based on the second skill; wherein the first automatic evolution link is used to perform personalized optimization processing on the skill based on local data, the local data including at least the user input feedback data recorded by the intelligent agent system during the execution of the skill; the second automatic evolution link is used to perform general function optimization processing on the skill based on general principles, the general principles being at least principles determined through skills uploaded by multiple different users.

Description

Technical Field

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

[0002] With the development of artificial intelligence technology, agent systems are gradually being applied to various scenarios such as task execution, content generation, information processing, and tool invocation. To handle different types of tasks, agent systems typically pre-configure or load multiple skills. These skills can exist in the form of code, parameter configurations, prompt templates, workflows, or tool encapsulations, and are stored in a local skill library.

[0003] In related technologies, if an intelligent agent system invokes a skill in response to a user's request but the output does not meet the user's needs, the user needs to continuously interact with the intelligent agent to adjust the output. Alternatively, the user can search for new skills (including but not limited to those with the same function or type) in a publicly available online skill library and download them locally. However, using this method, after downloading multiple skills from the internet, results in a large accumulation of redundant skills in the device's local skill library. Furthermore, when the intelligent agent system invokes a skill, it needs to utilize significant resources to determine which skill to use. Summary of the Invention

[0004] This application provides at least one method for skill self-evolution of an intelligent agent system and an intelligent agent system.

[0005] The technical solution of this application embodiment is implemented as follows: On one hand, embodiments of this application provide a method for the self-evolution of skills in an intelligent agent system, the method comprising: Based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link, the first skill stored in the local skill library is transmitted to the second self-evolution link, wherein the first skill is a skill processed based on the first automatic evolution link; The second skill is obtained from the second self-evolutionary link, wherein the second skill is a skill processed based on the second automatic evolutionary link; Whether to replace the third skill is determined based on the second skill, wherein the third skill is a skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link is used to perform personalized optimization processing on skills based on local data, wherein the local data includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link is used to perform general function optimization processing on skills based on general principles, wherein the general principles are at least principles determined by skills uploaded by multiple different users.

[0006] In some embodiments, the method further includes: responding to a self-evolution event involving a target skill, processing the target skill using local data based on a first automatic evolution link to generate a first skill; and updating the local skill library with the first skill; wherein the local data includes at least user input feedback data recorded by the agent system during the execution of the target skill in response to a task; determining whether to replace a third skill that satisfies a similarity condition with the second skill and is located in the local skill library based on the second skill includes: determining a skill that satisfies a similarity condition with the second skill as the first skill, and the third skill that satisfies a similarity condition with the second skill as the first skill; determining whether to replace the first skill with the second skill, wherein the second skill is a skill generated by the second automatic evolution link based on general principles to optimize the first skill's general functions; wherein the self-evolution event is when at least one of the following events is detected, initiating evolution processing of the target skill: task execution failure, user dissatisfaction with task results, detection of skill defects, or target skill performance falling below a preset threshold.

[0007] In some embodiments, the step of transmitting the first skill stored in the local skill library to the second self-evolution link based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link includes: transmitting the local data corresponding to the first skill to the second self-evolution link based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link, so that the second self-evolution link can extract the common functions of the same category of skills and determine the general principles through the skills uploaded by multiple different users and the local data corresponding to the skills.

[0008] In some embodiments, determining whether to replace a third skill that satisfies the similarity condition to the second skill and is located in the local skill library based on the second skill includes: determining that the third skill satisfies the similarity condition to the second skill based on the second skill; if the user feedback first confirmation information is obtained, replacing the third skill with the second skill; if the user feedback second confirmation information is obtained, continuing to keep the third skill in the local skill library.

[0009] In some embodiments, determining whether to replace a third skill that satisfies the similarity condition with the second skill and is located in the local skill library based on the second skill includes: determining that the third skill satisfies the similarity condition with the second skill based on the second skill; in response to a fusion instruction, adjusting the third skill based on the second skill through the first self-evolutionary link to generate a fourth skill, and replacing the third skill with the fourth skill.

[0010] In some embodiments, the second self-evolutionary link is deployed in the cloud, and the method further includes: performing de-identification processing on the first skill and the local data corresponding to the first skill; performing a quality assessment on the de-identified first skill and the de-identified local data, wherein the quality assessment is used to determine whether the functional integrity, performance indicators and user satisfaction indicators meet preset requirements; if the de-identified first skill and the de-identified local data meet the quality assessment, uploading the de-identified first skill and the de-identified local data to the second self-evolutionary link in the cloud.

[0011] In some embodiments, before outputting the second skill, the second automatic evolution chain performs at least one generalization process on the first skill and the local data of the first skill, the generalization process including removing user-personalized parameters, abstracting task processes, and retaining at least one of reusable functions; the generalization process is used to cluster the same type of skills with other users to generate general principles.

[0012] In some embodiments, before outputting the second skill, the second automatic evolution chain optimizes the first skill based on the general principles to generate the second skill, and performs cross-scenario verification, a second quality assessment, and a security assessment on the second skill; wherein, the cross-scenario verification includes at least one of functional regression testing in multiple task scenarios, compatibility testing in different user environments, and performance benchmark testing; the second quality assessment includes at least judging whether functional completeness, performance indicators, and user satisfaction indicators meet one of the preset requirements, and the security assessment includes judging whether code security, permission usage compliance, and sandbox isolation verification results meet one of the preset requirements.

[0013] Secondly, an intelligent agent system is provided, comprising: First automatic evolutionary link module; Second automatic evolutionary link module; A collaborative processing module is used to transmit a first skill stored in the local skill library to the second self-evolution link module, wherein the first skill is a skill processed by the first automatic evolution link module; and to obtain a second skill from the second self-evolution link module, wherein the second skill is a skill processed by the second automatic evolution link module. An update module is used to determine whether to replace the third skill based on the second skill, wherein the third skill is a skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link module is used to perform personalized optimization processing on skills based on local data, wherein the local data includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link module is used to perform general function optimization processing on skills based on general principles, wherein the general principles are at least those determined by skills uploaded by multiple different users.

[0014] Secondly, a terminal device is provided, comprising: Communication devices used to establish transmission channels with cloud devices; A processor is configured to transmit a first skill stored in the local skill library to the second self-evolutionary link via the transmission channel, wherein the first skill is a skill processed based on the first automatic evolutionary link; The second skill is obtained from the second self-evolutionary link through the transmission channel, wherein the second skill is a skill processed based on the second automatic evolutionary link; Based on the second skill, determine whether to replace the third skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link is used to perform personalized optimization processing on skills based on local data, wherein the local data includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link is used to perform general function optimization processing on skills based on general principles, wherein the general principles are at least principles determined by skills uploaded by multiple different users. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the implementation process of a skill self-evolution method for an intelligent agent system provided in an embodiment of this application; Figure 2 This is a schematic diagram of the composition structure of an intelligent agent system provided in an embodiment of this application; Figure 3 This is a schematic diagram of the composition structure of a terminal device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the composition structure of a dual-track self-evolution system for an intelligent agent provided in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the following description, references to "some embodiments" refer to a subset of all possible embodiments. It is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "first / second / third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for descriptive purposes only and is not intended to limit the scope of this application.

[0019] This application provides a method for the self-evolution of skills in an intelligent agent system, a system applying the method, and a user terminal applying the method, aiming to improve the user experience when the intelligent agent on the user terminal performs specific tasks.

[0020] The user terminal in this application embodiment can be a mobile phone, tablet computer, laptop computer, handheld computer, mobile internet device (MID), wearable device, virtual reality (VR) device, augmented reality (AR) device, in-vehicle device, computing device, or other processing device capable of connecting to a wireless communication network. The common feature of these various user terminals is that they all possess data processing and network communication capabilities, enabling them to deploy and run intelligent agent systems.

[0021] In related technologies, when users face new task requirements that exceed the processing capacity of their local skill library, they can typically access a cloud-based skill platform to download the required skills from a wide range of cloud-based tools to their local device. Each downloaded skill contains its code, model, configuration files, etc., consuming device storage space and memory. Downloading multiple skills from the cloud results in a large accumulation of redundant skill copies in the device's local skill library. This excessive redundancy leads to the ineffective use of the user terminal's intelligent agent's resources. Consequently, when the user terminal's intelligent agent performs a specific task, it may discover multiple similar skills locally, making it difficult to automatically select the optimal skill. This results in unstable task execution and a degraded user experience.

[0022] Furthermore, if the intelligent agent system invokes a skill in response to a user's request but the output does not meet the user's needs, the user will need to continuously interact with the intelligent agent to adjust the output. This will cause the intelligent agent system to use a lot of resources to determine a skill that meets the user's expectations when invoking the skill, resulting in a decline in user experience.

[0023] In view of the above problems, embodiments of this application provide a method for skill self-evolution of an intelligent agent system and an intelligent agent system. The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0024] Figure 1 This is a schematic diagram illustrating the implementation process of a skill self-evolution method for an intelligent agent system provided in an embodiment of this application. The following will be combined with... Figure 1 The steps shown will be explained. It should be noted that... Figure 1 The skill self-evolution method of the intelligent agent system can be implemented through intelligent agents on the user terminal, such as... Figure 1 As shown, the method includes the following steps S101 to S103: Step S101: Based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link, the first skill stored in the local skill library is transferred to the second self-evolution link.

[0025] The first skill is a skill processed based on the first automatic evolutionary link. The first automatic evolutionary link is used to perform personalized optimization processing on the skill based on local data, which includes at least the user input feedback data recorded by the intelligent agent system during the execution of the skill.

[0026] In this embodiment of the application, the intelligent agent of the user terminal can transmit the first skill stored in the local skill library to the second self-evolution link based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link.

[0027] An intelligent agent on a user terminal can refer to a software entity or system deployed on the user terminal. It can receive user commands or perceive environmental requirements, and plan solutions based on these commands or requirements to complete the tasks corresponding to those commands or requirements. When completing a task, if the intelligent agent determines that a specific function is needed, it can search for and invoke the corresponding skill from its local skill library to accomplish the task.

[0028] Skills can refer to specific capability units or tools possessed by an intelligent agent to complete a particular type of task. Each skill can encapsulate dedicated logic, algorithms, or workflows for solving a certain type of problem. In actual intelligent agent systems, skills can take many forms. For example, skills can include, but are not limited to, information processing skills such as algorithms for extracting core content from long documents, content generation skills such as generating corresponding images based on text descriptions, and code generation skills such as predicting and generating subsequent code snippets based on context and programming language.

[0029] In some embodiments, the agent may be configured with a first automatic evolutionary link. During the execution of a skill, the agent system may record feedback data from user input, and then perform personalized optimization of the skill based on local data, including at least the feedback data, through the first automatic evolutionary link.

[0030] In the specific implementation, when an agent invokes and executes a skill in response to a user task, the agent can record local data related to that execution. This local data may include, but is not limited to, user input feedback data, skill execution context data, and skill execution performance data.

[0031] After the agent obtains local data, it can perform personalized optimization of the executed skills based on the local data through the first automatic evolution link, so as to obtain the first skill after processing based on the first automatic evolution link.

[0032] In one possible implementation, if local data is detected to meet at least one of the following conditions, a first automatic evolutionary link can be triggered to perform personalized optimization processing on the executed skills based on the local data.

[0033] Condition 1: Local data indicates performance degradation. For example, when the performance data of a skill in the local data from multiple executions of a certain skill indicates that its performance is lower than a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on the local data.

[0034] Condition 2: Local data indicates negative feedback. For example, if the feedback data from multiple executions of a skill indicates that the number of times negative user feedback occurs exceeds a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on the local data. Similarly, if the feedback data from multiple executions of a skill indicates that the number of times the user manually modifies the output of that skill exceeds a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on the local data.

[0035] Condition 3: Local data indicates task failure. For example, when the number of times the agent system state in the local data of multiple executions of a certain skill indicates that the task has failed exceeds a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on the local data.

[0036] When the agent determines to trigger the first automatic evolutionary link to perform personalized optimization of the executed skill based on local data, it can analyze the local data from multiple executions of a certain skill to obtain optimization targets, such as common patterns of user behavior modification or identifying input features that lead to task failure. Then, based on the analysis results, the skill can be adjusted, and the adjustment methods can include, but are not limited to, parameter adjustments, logic adjustments, and prompt word adjustments.

[0037] For example, parameter adjustments can include adjusting configuration parameters within a skill; logic adjustments can include modifying rules or conditional decision branches within a skill, such as adding a rule to omit that content by default during generation if analysis finds that users always delete a certain item in the output; and prompt word adjustments can include optimizing the system prompt words for skills based on large language models, incorporating preferences learned from user feedback.

[0038] After the agent performs personalized optimization on the skill, a first skill processed based on the first automatic evolution link can be obtained, and then the first skill can be transferred to the second self-evolution link.

[0039] Step S102: Obtain the second skill from the second self-evolutionary link.

[0040] The second skill is a skill processed based on the second automatic evolution link. The second automatic evolution link is used to optimize the general functions of the skill based on general principles. The general principles are at least those determined by skills uploaded by multiple different users.

[0041] In this embodiment, a second self-evolutionary link can be used to optimize general functions based on local data of multiple candidate skills received, resulting in a second skill processed by the second self-evolutionary link. The multiple candidate skills can be skills of the same type as the first skill uploaded by multiple different users, or multiple skills of the same type as the first skill uploaded by the same user.

[0042] In practical applications, the second self-evolutionary link can first perform cluster analysis based on local data of multiple candidate skills to obtain general principles for configuration parameters used to characterize skills of the same type as the first skill. Then, the second skill can be generated based on these general principles and distributed to the agent.

[0043] The general principle for configuring parameters to characterize skills of the same type as the first skill can refer to a set of configuration parameters for that type of skill extracted from local data of multiple candidate skills through statistical analysis and machine learning methods. This set of configuration parameters can be defined as a parameter configuration paradigm that enables that type of skill to achieve a better balance between performance, effectiveness, and resource consumption in a general scenario.

[0044] In one possible implementation, the second self-evolutionary link can pre-configure skill templates applicable to different skill categories. These skill templates can define the standard processing flow, input / output interfaces, and framework for that skill category, but they will have parameter gaps. When generating a second skill, the configuration parameters from the general principles can be filled into the corresponding parameter gaps in the skill template. After completing the parameter configuration, the second skill processed based on the second automatic evolutionary link can be obtained.

[0045] In another possible implementation, the second self-evolutionary link can be trained to obtain a new skill by using the configuration parameters in the general principle as the constraint target. This new skill is the second skill processed by the second automatic evolutionary link.

[0046] Step S103: Based on the second skill, determine whether to replace the third skill that meets the similarity condition with the second skill and is located in the local skill library.

[0047] The third skill can refer to a skill in the local skill library that meets the similarity condition to the second skill, or it can refer to the first skill uploaded to the second self-evolutionary link.

[0048] In one possible implementation, a third skill that meets the similarity condition of the second skill and is located in the local skill library can be replaced if the user explicitly authorizes the second skill. Specifically, when the agent receives the second skill, it can initiate a confirmation to the user, and upon receiving a positive confirmation from the user, replace the third skill that meets the similarity condition of the second skill and is located in the local skill library.

[0049] Among them, a skill that meets the similarity condition with the second skill can refer to a skill located in the local skill library that is consistent with the second skill in terms of functional category; for example, a skill located in the local skill library that is consistent with the skill identifier of the second skill can be a skill that meets the similarity condition with the second skill.

[0050] In another possible implementation, if the confidence level of the second skill meets a preset threshold, a third skill that meets the similarity condition of the second skill and is located in the local skill library can be replaced.

[0051] The confidence level of the second skill can refer to the test score obtained by testing and verifying the second skill after it is obtained in the second self-evolutionary link. This verification test can be used to verify the credibility of the second skill in terms of quality, reliability, security and expected utility.

[0052] By adopting the above technical solution, a second skill that can reflect the evolutionary achievements of multiple users is obtained, and it is determined whether to replace the local redundant skills of the same kind based on the second skill. This can effectively simplify the local skill library and avoid the accumulation of a large number of redundant skill copies in the device's local skill library. This allows the intelligent agent to call the general functional skills that can reflect the evolutionary achievements of multiple users when performing specific tasks, thereby improving the stability and reliability of task execution and enhancing the user experience.

[0053] In some embodiments, the method further includes the following steps.

[0054] Step S104: In response to a self-evolution event involving a target skill, process the target skill using local data based on the first automatic evolution link to generate a first skill; and update the local skill library with the first skill.

[0055] The local data includes at least the user input feedback data recorded by the agent system during the execution of the target skill in response to the task. The self-evolutionary event is triggered when at least one of the following events is detected, initiating evolutionary processing of the target skill: task execution failure, user dissatisfaction with the task result, detection of skill defects, or the target skill performance falling below a preset threshold.

[0056] In this embodiment, the target skill can be a skill invoked by the user terminal's intelligent agent when performing a specific task. When the intelligent agent invokes the target skill to perform a specific task, the intelligent agent can record local data related to that execution.

[0057] The local data may include, but is not limited to, user-input feedback data. For example, feedback data may represent explicit task evaluations given by users to the skill execution results, such as a five-star rating, a like, or the text "the summary is not concise enough." Feedback data may also represent implicit task evaluations given by users to the skill execution results, such as when a user manually modifies the text or solution generated by the skill; when a user selects one of multiple output results and ignores the others; or when a task is canceled or re-executed.

[0058] This local data may also include skill execution context data and skill execution performance data. Specifically, skill execution context data may include the specific user instruction triggered when the skill is activated, the raw input data, the agent system state, and other data; skill execution performance data may include the execution time, CPU / memory usage, success rate, and other data.

[0059] In one possible implementation, a self-evolution event of the target skill can be triggered when the agent detects that the local data of the target skill meets at least one of the following conditions.

[0060] Condition 1: Task execution failure. For example, when the number of times the agent system state in the local data of multiple executions of a certain skill indicates that the number of task failures exceeds a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on the local data.

[0061] Condition 2: The user is dissatisfied with the task result. For example, when the feedback data from multiple executions of a certain skill indicates that the number of times negative user feedback occurs exceeds a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing of the executed skill based on the local data; or, for example, when the feedback data from multiple executions of a certain skill indicates that the number of times the user manually modifies the output result of the skill exceeds a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing of the executed skill based on the local data.

[0062] Condition 3: Skill defects are detected. For example, when local data from multiple executions of a skill, or during a special test of the skill, detects code anomalies, logical errors, security vulnerabilities, or behaviors that do not conform to expected specifications, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on relevant data.

[0063] Condition 4: The performance of the target skill is below a preset threshold. For example, when the performance data of a skill in local data from multiple executions of a certain skill indicates that its performance is below a preset threshold, the first automatic evolution link can be triggered to perform personalized optimization processing on the executed skill based on the local data.

[0064] When an agent detects a self-evolution event involving a target skill, it can process the target skill using local data based on the first automatic evolution link to generate a first skill; and update the local skill library with the first skill.

[0065] Specifically, when updating the local skill library with the first skill, it can be updated in real time. For example, the first skill can be generated in real time based on local feedback data, and the local skill library can be updated in real time when the first skill is generated.

[0066] Non-real-time updates can also be used. For example, the first automatic evolutionary link can continuously collect and analyze local data, but instead of triggering an update immediately, it processes the local data in batches according to a preset period or after the accumulated optimization samples reach a preset threshold, in order to generate a first skill that integrates all recent optimization points, and then updates the local skill library all at once. This can reduce the system overhead caused by frequent updates.

[0067] After generating the first skill, you can choose not to write it directly to the local skill library. Instead, you can perform a quick regression test, performance benchmark test, or security scan on it in a local sandbox or test environment. Only after the generated first skill passes all verifications, confirming that it functions correctly, meets performance standards, and has no security issues, should you perform an update operation to replace the old version in the local skill library. This adds a security gate to the generated first skill.

[0068] After the first skill generated passes all verifications, an interactive confirmation can be initiated with the user through the intelligent agent. For example, a prompt message can be output: "Skill A has been optimized based on your usage habits. Do you want to activate the new version immediately?" With the user's affirmative authorization, the update operation is then performed to replace the old version in the local skill library.

[0069] In this embodiment, step S103 can be implemented by steps S1031 and S1032.

[0070] Step S1031: Based on the second skill, determine the skill that satisfies the similarity condition with the second skill as the first skill.

[0071] Among them, the third skill that meets the similarity condition with the second skill is the first skill.

[0072] In this embodiment, when the agent receives the second skill, it can parse the data packet of the second skill to obtain metadata such as the unique identifier of the second skill, the skill category code, and the functional description. Then, based on this metadata, it can perform a query and match in the local skill library. After the query, it can be determined that the target skill that triggers the self-evolution event is the third skill that meets the similarity condition with the second skill, which is the first skill processed based on the first automatic evolution link.

[0073] Step S1032: Determine whether to replace the first skill with the second skill.

[0074] Among them, the second skill is a skill generated by optimizing the general functions of the first skill based on general principles in the second automatic evolution link.

[0075] In this embodiment, the second automatic evolutionary link can be deployed in the cloud. Therefore, after collecting sufficient and diverse user terminal evolutionary data in the second automatic evolutionary link, centralized analysis and global optimization can be performed to generate a second skill with more global representativeness. This allows for efficient use of cloud computing resources. In other words, the second automatic evolutionary link can adopt an update mechanism of delayed aggregation and timed batch synchronization.

[0076] This avoids making decisions based on limited and accidental evolutionary data, ensuring that the extracted general principles and generated second skills are more universal; moreover, batch processing can maximize the parallel capabilities of cloud-based distributed computing, reducing the processing cost per skill; and the timed and batch-based skill distribution provides a stable and predictable update cycle for terminal devices, which is beneficial to the overall stability management of the system.

[0077] However, considering that the second automatic evolution link set up in the cloud has a cycle for skill generation, verification and distribution, while the first skill processed by the first automatic evolution link can evolve in real time based on local feedback, the general principle on which the distributed second skill is based is derived from global data including a local uploaded version of the first skill from a certain historical version. However, the local first skill has evolved further based on new data within this cycle. Therefore, it is necessary to determine whether to replace the first skill with the second skill.

[0078] In one possible implementation, it can be determined whether to replace the first skill with the second skill by comparing the global data cutoff time on which the second skill is based with the last evolution time of the local first skill.

[0079] For example, if the last evolution time of the first skill is later than the global data deadline of the second skill, it indicates that the local skill has undergone new personalized optimizations after the cloud aggregation period. In this case, it can be determined not to replace the first skill with the second skill.

[0080] In another possible implementation, it can be determined whether to replace the first skill with the second skill by comparing the first confidence level of the second skill with the second confidence level of the local first skill. Here, the first confidence level can refer to the test score of the second skill after verification testing, and the second confidence level can refer to the test score of the first skill after verification testing.

[0081] For example, if the first confidence level of the second skill is greater than the second confidence level of the local first skill, the second skill can be used to replace the skill targeting the first skill; if the first confidence level of the second skill is equal to the second confidence level of the local first skill, the second skill can still be used to replace the skill targeting the first skill, because the fusion of the second skill has general rules and is more universal; if the first confidence level of the second skill is less than the second confidence level of the local first skill, the second skill can not be used to replace the skill targeting the first skill.

[0082] By adopting the above technical solution, and by determining that the third skill is the same as the first skill, each local evolution result (the first skill) can be used as an input and adjustment object in the global evolution process. This allows local skills to undergo global optimization through the second automatic evolution link after self-optimization through the first automatic evolution link. This avoids isolated evolution of skills, improves the collective evolution of skills, and thus enhances the reliability of task execution and user experience.

[0083] In some embodiments, step S101 described above can be implemented in the following manner.

[0084] In this embodiment, based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link, the local data corresponding to the first skill can be transmitted to the second self-evolution link, so that the second self-evolution link can extract the common functions of the same category of skills and determine the general principles through the skills uploaded by multiple different users and the local data corresponding to the skills.

[0085] The collaborative mechanism between the first automatic evolution link and the second self-evolution link can refer to a skill update logic that ensures the skill can simultaneously achieve local personalized evolution and global standardized evolution. This skill update logic can define the conditions under which local data is transferred from the first automatic evolution link to the second automatic evolution link.

[0086] For example, if an agent triggers a first automatic evolution link to perform personalized optimization processing on the executed skill based on local data, the local data of the first skill of the locally personalized evolution can be transmitted from the first automatic evolution link to the second automatic evolution link.

[0087] In the embodiments of this application, the common function of skills of the same category can refer to a standardized description of the common problems to be solved and the common goals to be achieved for a certain type of skill. For example, for the skill category of text summarization, its common function may include, but is not limited to, the intent of the skill, the extraction algorithm used, the style of the generated text, and other functions.

[0088] The general rules are a set of configuration parameters for implementing the general function, which are extracted by analyzing the skills uploaded by multiple different users and the corresponding local data.

[0089] For example, if the second automatic evolutionary link, after acquiring multiple text summary skills uploaded by different users and the corresponding local data for each text summary skill, analyzes the user input feedback data in each local data set, it can conclude that in instances with high user satisfaction, 78% of the skills used generative summarization algorithms rather than extraction-based ones; therefore, the second automatic evolutionary link can conclude that generative summarization algorithms should be prioritized. Alternatively, when the summary length parameter is set to 15% to 20% of the original text length, the skill receives the highest average score; therefore, the second automatic evolutionary link can conclude that the recommended optimal range for the summary length parameter is 15%-20% of the original text length. Or, for texts from the field of scientific papers, skills with the instruction to "prioritize retaining methodology and conclusions" added to the prompt words significantly reduce the number of times the generated results are modified by users; therefore, the second automatic evolutionary link can conclude that when the input text is identified as academic, instructions to prioritize extracting methodology and conclusions should be embedded in the processing logic.

[0090] By adopting the above technical solution, through the collaborative mechanism of the first automatic evolution link and the second self-evolution link, the general functions generated by local personalized evolution can be aggregated, analyzed and extracted to generate general rules, and the second skill can be generated and distributed to the intelligent agent based on the general rules. This can ensure that the intelligent agent's skills can simultaneously achieve local personalized evolution and global standardized evolution.

[0091] Considering that the second skill delivered from the cloud is based on general rules extracted from global data, while the third skill that meets the similarity condition and is located in the local skill library is a personalized version that has evolved over a long period of time based on personal data and is highly adapted to the user's unique habits, workflow or specific scenarios, if the principle of simple replacement is used to cover the personal best with the global best, it may in some cases disrupt the stable experience and work efficiency that the user has already formed.

[0092] Therefore, in some embodiments, the third skill can first be determined based on the second skill and the similarity condition of the second skill. Then, if the user provides first confirmation information, the third skill is replaced with the second skill. If the user provides second confirmation information, the third skill is kept in the local skill library.

[0093] In this embodiment, the third skill can first be determined based on the second skill, provided that it meets the similarity condition. The method for determining the third skill can be found in step S103 above; the specific implementation will not be repeated here.

[0094] After identifying a third skill that meets the similarity criteria to the second skill and is located in the local skill library, the agent can generate and output a user interaction notification. This notification prompts the user that there is an official skill update (i.e., the second skill) that has been globally optimized in the cloud and verified in terms of security and general performance. Its function is similar to a local skill (the third skill) currently used by the agent, and asks the user whether to replace the old skill with the new skill.

[0095] The intelligent agent can output the user interaction notification to the user through the user interface and detect the user information input by the user based on the user interaction notification. The user information may include first confirmation information and second confirmation information.

[0096] If the agent detects the first confirmation information, it can determine that the user's explicit authorization conditions are met, and then the replacement process can be triggered to replace the third skill in the local skill library with the second skill to complete the update; if the agent detects the second confirmation information, it can determine that the user chooses to maintain the status quo, and the replacement operation will not be performed. The third skill in the local skill library can remain unchanged and continue to be called as the default skill.

[0097] In one possible implementation, the agent can record the user's rejection decision and avoid popping up notifications for the same version of the same skill again during a preset cooldown period to improve the user experience.

[0098] By adopting the above technical solutions, we can ensure the user's final choice during the use of the intelligent agent, ensure the stability of the user experience, and improve the acceptance of intelligent agent skill updates.

[0099] Considering that the user-confirmed replacement strategy provided in the above embodiments is essentially a simple choice between two options, in practice, the second skill usually has better general performance, while the third skill will contain personalized logic that has been refined over a long period of time and is extremely well-suited to the user's specific scenario. If a simple replacement strategy is adopted, the advantage of one of them will be lost.

[0100] Therefore, in some embodiments, the third skill can first be determined based on the second skill and satisfy the similarity condition of the second skill. Then, in response to the fusion command, the third skill can be adjusted based on the second skill through the first self-evolution link to generate the fourth skill and replace the third skill with the fourth skill.

[0101] In this embodiment, the third skill can first be determined based on the second skill, provided that it meets the similarity condition. The method for determining the third skill can be found in step S103 above; the specific implementation will not be repeated here.

[0102] In one possible implementation, after the agent determines a third skill that meets the similarity condition with the second skill and is located in the local skill library, the agent can provide a fusion option when generating user interaction notifications, in addition to providing options to update immediately (replace) and not update temporarily (maintain). When the user selects this option, it is considered that a fusion instruction has been received.

[0103] In another possible implementation, after the agent determines a third skill that meets the similarity condition with the second skill and is located in the local skill library, the agent can automatically generate a fusion instruction.

[0104] Upon receiving the fusion command, the agent can respond to the fusion command by adjusting the third skill based on the second skill through the first self-evolution link, generating a fourth skill, and replacing the third skill with the fourth skill.

[0105] In this embodiment, after the agent receives the fusion instruction, it can invoke the first automatic evolution link and use the second skill and the third skill as inputs to the first automatic evolution link to obtain the fourth skill generated after adjusting the third skill based on the second skill.

[0106] In one possible implementation, when adjusting the third skill based on the second skill, a difference analysis can first be performed on the configuration parameter sets of the second and third skills to obtain the different difference parameters between the second and third skills. Then, based on each difference parameter, a new value can be calculated using a confidence-based weighted strategy to obtain a fourth skill that includes each new difference parameter. For example, for parameter A, its confidence level in the second skill is A1 and its parameter value is B1, and its confidence level in the third skill is A2 and its parameter value is B2. Then, in the generated fourth skill, the parameter value of parameter A is (A1×B1+A2×B2) / (A1+A2).

[0107] In another possible implementation, when adjusting the third skill based on the second skill, a difference analysis can first be performed on the configuration parameter sets of the second and third skills to obtain the different difference parameters between them. Then, for each difference parameter in the third skill, the agent can query the user feedback data recorded in the historical local execution log under the current value of that difference parameter; it can also query whether the difference parameter under the current value is strongly bound to certain specific, high-frequency local task scenarios. For example, data mining reveals that when the parameter is set to High, instances exceeding a preset threshold all occur in the scenario of processing chat log summaries from internal instant messaging tools, and in this scenario, the user feedback data is satisfactory.

[0108] If the user feedback data corresponding to the difference parameter is determined to be satisfactory, or if the difference parameter is strongly bound to certain specific and frequently occurring local task scenarios under its current value, the value of the parameter in the third skill will be maintained; if the user feedback data corresponding to the difference parameter is determined to be unsatisfactory, or if the difference parameter is not bound to certain specific and frequently occurring local task scenarios under its current value, the value of the parameter in the third skill will be updated to the value of the parameter in the second skill, and the adjusted third skill can then be used as the fourth skill.

[0109] By adopting the above technical solution, the third skill, which embodies local personalized evolution, can be integrated with the second skill, which embodies global standardized evolution. This ensures that the integrated skill achieves a dynamic balance between global and individual optimality at the parameter level, avoiding the loss of advantages due to simple replacement.

[0110] In some embodiments, the second self-evolutionary link can be deployed in the cloud. Therefore, when using the second self-evolutionary link to extract the common functions of the same category of skills and determine the general principles through the skills uploaded by multiple different users and the local data corresponding to the skills, the following steps need to be performed by the intelligent agent of the user terminal on the skills uploaded by multiple different users and the local data corresponding to the skills.

[0111] Step S1021: Desensitize the first skill and the corresponding local data.

[0112] De-identification processing may include removing or obfuscating information in local data that can directly or indirectly identify an individual.

[0113] In this embodiment, the intelligent agent can first remove personally identifiable information from the local data. Specifically, it deletes or replaces all information that can directly pinpoint a specific natural person from the local data corresponding to the first skill. For example, it deletes fields such as username, real email address, precise GPS coordinates, and device unique identifier.

[0114] Furthermore, intelligent agents can generalize numerical data. Specifically, fields containing specific values ​​can be converted into broader ranges or categories to eliminate their precise targeting.

[0115] In one possible implementation, differential privacy noise can be added to the skill parameter increments. Specifically, before uploading the locally optimized skill parameters, random noise conforming to the definition of differential privacy can be added. For example, for a set of parameter vectors, noise sampled from a Laplace or Gaussian distribution can be added.

[0116] Step S1022: A quality assessment is performed on the first de-identified skill and the de-identified local data. The quality assessment is used to determine whether the functional integrity, performance indicators and user satisfaction indicators meet the preset requirements.

[0117] In this embodiment, after the second automatic evolution link de-identifies the first skill and its corresponding local data, the de-identified first skill can be loaded and run in an isolated sandbox environment. During runtime, basic functional test cases corresponding to the skill category can be used. These test cases use general test data and do not contain sensitive user data. After runtime, it can be checked whether the first skill can start normally, receive standard input, execute core processing logic, and produce output in the expected format without any unhandled runtime errors. If all basic test cases pass, the functional integrity is deemed satisfactory.

[0118] Skill execution performance data can also be extracted from the local data corresponding to the first skill after anonymization. Then, the skill execution performance data can be compared with the preset performance threshold. If the skill execution performance data meets the preset performance threshold, the performance indicator is determined to be up to standard.

[0119] User feedback data can also be extracted from the local data corresponding to the first skill after anonymization. Then, the user feedback data can be quantified and statistically analyzed to obtain a user satisfaction index. If the user satisfaction index meets the preset satisfaction threshold, it can be determined that the user satisfaction index has met the standard.

[0120] Step S1023: If the desensitized first skill and the desensitized local data meet the quality assessment, upload the desensitized first skill and the desensitized local data to the second self-evolutionary link in the cloud.

[0121] By adopting the above technical solution, on the one hand, by de-identifying the first skill and its corresponding local data before sending the local evolution results to the cloud on the user terminal, it can be ensured that sensitive data does not leave the user device, and the cloud always receives de-identified data, thus avoiding the leakage of user privacy; on the other hand, by conducting quality assessment on the de-identified data, it can be ensured that only evolution results that have been verified locally as truly effective, functioning well, and recognized by users participate in global standardized evolution, which can guarantee that the final generated second skill has a higher quality benchmark and practicality.

[0122] Considering that the first skill and its local data uploaded by different user terminals after anonymization belong to the same category in terms of function, they will still include a large number of personalized parameters that are only meaningful to the original user device or environment in the specific implementation, such as local file path, specific network address, and personalized parameters of personal account configuration. If they are directly generalized, it will lead to the inability to accurately identify the common rules between skills. Therefore, in some embodiments, the second automatic evolution link performs generalization processing on at least one of the first skill and the local data of the first skill before outputting the second skill. The generalization process includes removing user-specific parameters, abstracting task processes, and retaining at least one of the following: generalization is used to cluster the same type of skills with other users to generate general principles.

[0123] In this embodiment, before outputting the second skill, the second automatic evolution link can scan the first skill and its local data for data such as skill configuration files, environment variables, and execution logs. Using predefined pattern recognition rules, it locates and marks configuration items that are only meaningful to the original user device or environment. Then, it replaces the specific values ​​of these personalized parameters with generic placeholders, thereby avoiding environmental dependency noise unrelated to the core function in the data and achieving uniformity in configuration across different user skills. The second automatic evolution link can also parse the execution step sequence of the first skill, determine the abstract task flow of the first skill during execution, and identify reusable functions of the first skill during execution.

[0124] The second automatic evolution link can be based on an abstract task process that removes user-personalized parameters and reusable functions, and generate general principles for the first skill by clustering the same type of skills of other users.

[0125] By adopting the above technical solution, the second automatic evolution link can accurately perform cluster analysis based on the essential similarity of skills, thereby generating a general principle for each first skill.

[0126] In some embodiments, before outputting the second skill, the second automatic evolution link optimizes the first skill based on general principles to generate the second skill, and performs cross-scenario verification, as well as a second quality assessment and a safety assessment, for the second skill. Among them, cross-scenario verification includes at least one of the following: functional regression testing under multiple task scenarios, compatibility testing under different user environments, and performance benchmark testing; the second quality assessment includes at least one of the following: judging whether the functional integrity, performance indicators, and user satisfaction indicators meet the preset requirements; and the security assessment includes judging whether the code security, permission usage compliance, and sandbox isolation verification results meet the preset requirements.

[0127] In this embodiment, after generating candidate skills, the second automatic evolution link does not directly distribute them to the agent. Instead, it performs cross-scenario verification, a second quality assessment, and a security assessment on the generated candidate skills to ensure that they have high reliability, broad adaptability, and security.

[0128] Specifically, when conducting cross-scenario verification, test scenarios corresponding to at least one related skill with contextual association with the candidate skill can be selected from a pre-set multi-scenario test case library. The candidate skill is then executed one by one in the test scenarios to complete functional regression testing. For example, if the candidate skill is the "automatic generation of meeting minutes" skill, the at least one related skill with contextual association with the candidate skill could be the "meeting audio transcription" skill, which is used before generating meeting minutes, and the "summary text abstraction" skill.

[0129] When performing cross-scenario verification, tests can be executed sequentially from at least one associated skill that has a contextual relationship with the candidate skill to complete the full test. Cross-scenario verification can include at least one of the following: functional regression testing in multiple task scenarios, compatibility testing in different user environments, and performance benchmark testing.

[0130] In this embodiment, a second quality assessment can also be performed on the candidate skill. Specifically, if the candidate skill is successfully executed in all selected test scenarios and produces results that meet the preset output specifications of each scenario, the candidate skill's functional completeness can be determined to be up to standard; if the key indicator statistics of the candidate skill in all performance benchmark test scenarios meet the preset global performance release threshold, the candidate skill's performance indicators can be determined to be up to standard; if the candidate skill obtains user satisfaction that meets the preset satisfaction threshold in all test scenarios that include simulated user interaction or result evaluation, the candidate skill's user satisfaction indicator can be determined to be up to standard.

[0131] A security assessment can also be conducted on the candidate skill. Specifically, firstly, static application security testing can be performed on the code or configuration that constitutes the candidate skill to scan for known code vulnerabilities or malicious code patterns. Then, the skill is dynamically run in a sandbox environment with strict resource limitations and network isolation, and its system calls, file accesses, network requests, and other behaviors throughout its entire lifecycle are monitored to audit the compliance of its permission usage and verify the effectiveness of sandbox isolation, preventing unauthorized or malicious operations. Only when the candidate skill successfully passes all the above cross-scenario verifications, the second quality assessment, and the security assessment, and all results meet the preset stringent requirements of their respective modules, will the second automatic evolution chain ultimately determine the skill version as a releasable second skill, generate a corresponding global confidence score for it, and then include it in the global skill library awaiting distribution.

[0132] By adopting the above technical solution, the second skill is obtained by performing multi-level verification on the generated candidate skills, which can ensure that the second skill can be securely distributed to all user terminals as a reliable global standard skill.

[0133] Based on the above embodiments, this application also provides an intelligent agent system. Figure 2 This is a schematic diagram of the composition structure of an intelligent agent system provided in an embodiment of this application, such as... Figure 2 As shown, the intelligent agent system 200 includes a first automatic evolution link module 201, a second automatic evolution link module 202, a cooperative processing module 203, and an update module 204, wherein: The collaborative processing module 203 is used to transmit a first skill stored in the local skill library to the second self-evolution link module 202, wherein the first skill is a skill processed based on the first automatic evolution link module 201; and to obtain a second skill from the second self-evolution link module 202, wherein the second skill is a skill processed based on the second automatic evolution link module 202. The update module 204 is used to determine whether to replace the third skill based on the second skill, wherein the third skill is a skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link module 201 is used to perform personalized optimization processing on skills based on local data, which includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link module 202 is used to perform general function optimization processing on skills based on general principles, which are at least principles determined by skills uploaded by multiple different users.

[0134] This application also provides a terminal device. Figure 3 This is a schematic diagram of the composition structure of a terminal device provided in an embodiment of this application, such as... Figure 3 As shown, the terminal device 300 includes a communication device 301 and a processor 302.

[0135] In this embodiment, the processor 302 can be used to transmit a first skill stored in the local skill library to a second self-evolving link via a transmission channel, wherein the first skill is a skill processed based on a first automatic evolution link; obtain a second skill from the second self-evolving link via the transmission channel, wherein the second skill is a skill processed based on a second automatic evolution link; determine whether to replace a third skill that meets the similarity condition with the second skill and is located in the local skill library based on the second skill; wherein the first automatic evolution link is used to perform personalized optimization processing on the skill based on local data, the local data including at least the user input feedback data recorded by the intelligent agent system during the skill execution process; the second automatic evolution link is used to perform general function optimization processing on the skill based on general principles, the general principles being at least principles determined by skills uploaded by multiple different users.

[0136] The descriptions of the system-side and device-side embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the system-side and device-side embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0137] The following describes the application of the skill self-evolution method of the intelligent agent system provided in the embodiments of this application in a real-world scenario, mainly involving a dual-track self-evolution method based on the intelligent agent system and user interaction.

[0138] In current agent system technology practices, skills are usually pre-installed or pre-loaded in terminal devices, and the logic and function of the skills are basically fixed after deployment.

[0139] When users face new task requirements that exceed the scope of their existing local skills, the usual solution is to access a cloud-based skills platform, manually compare and filter the numerous skills offered by the platform based on their function descriptions, and then download the selected skills package to their local device.

[0140] However, this combination of a static local skill library and manual cloud downloads has the following limitations in practical applications: First, to find the most suitable skill, users often need to download multiple similar skills from the cloud, resulting in a large accumulation of redundant skill copies in the device's local skill library. When performing specific tasks, faced with multiple similar skills, there is a lack of an effective evaluation and decision-making mechanism to automatically call the optimal skill, causing the effectiveness, efficiency, and stability of task execution to be compromised. Second, standardized skills downloaded from the cloud are designed for general scenarios and cannot be self-optimized and dynamically adjusted based on continuous user feedback. This makes it difficult to adapt to users' unique operating habits and business scenarios, leading to a rigid user experience in the long run.

[0141] In some embodiments, within the relevant art, the skill evolution of an agent system primarily follows two implementation paths.

[0142] One approach is an agent skill evolution scheme based on open-source frameworks. Such schemes are typically implemented by mainstream open-source agent frameworks such as AutoGPT (Autonomous Generative Pre-trained Transformer), LangChainAgents (a popular open-source framework for developing language model-driven applications), and MetaGPT (Multi-Agent Generative Pre-trained Transformer). Their core feature is providing a single-track, local evolutionary path.

[0143] Specifically, under this scheme, the agent's skills are stored and managed in the user's local device environment in the form of code files, configuration files, or database records, forming a local skill library. The skill evolution process is triggered and executed entirely within a local closed loop: after the agent performs a task, it automatically analyzes the reasons and generates modification plans for the skill code or parameters based on the task execution result (such as success / failure) or direct feedback from the user. This scheme adopts an instant-effect mechanism, meaning that once a skill is modified, it immediately takes effect on the current device for subsequent tasks to invoke, without requiring additional review, verification, or synchronization processes. This single-track evolution scheme can quickly respond to users' personalized needs, but strictly limits the evolution scope to a single device.

[0144] Therefore, this approach has the following limitations: First, the evolution process lacks a systematic verification mechanism, and the modified skills may introduce errors, security vulnerabilities, or performance issues, i.e., there is a lack of verification loop; second, the evolution results cannot be shared between devices, causing the same optimization problems to be solved repeatedly by different users, forming information silos and reducing the overall evolution efficiency; finally, local evolution may produce low-quality or even harmful skill changes due to data quality or malicious input, posing a risk of uncontrollable quality.

[0145] Another approach is the cloud-based skills marketplace solution offered by commercial platforms. This provides a centralized cloud-based skills repository, with all skills managed uniformly by the platform. Skills typically undergo manual review before being listed and published, and version control and rollback are supported. This centralized management model is beneficial for ensuring the overall quality and security of skills, but it also brings new problems: First, reliance on manual review leads to a long skill release or update cycle, making it difficult to meet users' needs for timely and rapid skill evolution; second, a unified cloud-based skills repository is difficult to adapt to the specific usage scenarios and personalized preferences of different users; third, users usually need to upload complete information containing their task data to the cloud for skill training or optimization, which brings potential privacy risks.

[0146] In summary, existing skill evolution schemes, whether fully localized single-track evolution or fully centralized cloud marketplaces, struggle to achieve a good balance across multiple dimensions, including evolution efficiency and quality security, personalized response and standardized management, and rapid iteration and privacy protection. Therefore, a new skill evolution architecture is urgently needed to systematically address these issues.

[0147] This application proposes a dual-track self-evolution method for intelligent agents. This method is based on a dual-track self-evolution architecture, which can systematically divide the complete skill evolution process into a local personalized evolution track (i.e., the first automatic evolution link described in the above embodiments) and a global standardized evolution track (i.e., the second automatic evolution link described in the above embodiments). These two evolution tracks are logically and physically separated, but coupled through a cooperative mechanism.

[0148] In this embodiment, the local personalized evolutionary trajectory can be deployed and run locally on the user terminal device. The executing entity is the intelligent agent system itself on the user terminal, and the scope of the evolutionary operation is the user's exclusive local skill library.

[0149] The evolution of this track can be triggered by specific events, such as detecting task execution failure, receiving explicit user dissatisfaction feedback, or the system itself detecting logical flaws or performance bottlenecks in the skill. Once triggered, the localized evolution track can first automatically analyze the root cause of the problem, then generate a repair or optimization plan based on local interaction data, and then perform rapid regression testing and security scanning on the modified skill in a locally built sandbox environment. After successful verification, the new skill is immediately deployed to the local skill library and takes effect.

[0150] The verification mechanism employed in this track emphasizes immediacy and specificity, including verifying functional correctness in a resource-constrained sandbox, performing regression tests on historical use cases, and conducting basic security code scans. The entire evolution process and its scope are entirely confined to the single user device that initiated the evolution, thereby achieving immediate response to users' personalized needs.

[0151] The global standardized evolution track can be deployed on a cloud server. Its execution entity is the cloud-based skill evolution platform, and the goal of evolution is to maintain and optimize a global standard skill library for all users.

[0152] The evolution of this trajectory may not be directly triggered by local events. Its data comes from high-quality evolutionary results aggregated from local personalized evolutionary trajectories of a massive number of user terminals after explicit user authorization. First, the aggregated data can be de-identified and anonymized through differential privacy processing. Then, algorithms such as cluster analysis and pattern mining can be used to extract common optimization patterns and abstract principles of a certain type of skill from the anonymized data. Based on the extracted patterns, candidate skills are generated or optimized, and then subjected to extremely rigorous cross-scenario regression testing, performance benchmark testing, security auditing, and A / B testing, among other multi-dimensional verification and quality assessments. Only when a candidate skill passes all verifications is it released as a new version of the global skill library.

[0153] The verification mechanism of this track is characterized by its comprehensiveness and rigor. The scope of its evolution results is global, and it can be securely distributed and fed back to all user terminals connected to the platform.

[0154] Through the aforementioned dual-track architecture, the local personalized evolution track and the global standardized evolution track can work together. The local personalized evolution track can quickly perceive and adapt to individual differences, while the global standardized evolution track can form a better general solution based on local evolution, and then feed back to enhance all user terminals.

[0155] The scheme described in the above embodiments achieves the following: First, through a dual-track collaborative mechanism, the local personalized evolution track can quickly respond to users' personalized needs, while the global standardized evolution track can accumulate collective wisdom. The two achieve deep collaboration through a rigorous authorization, desensitization, verification, and feedback process. Second, through a layered verification system, the local personalized evolution track focuses on performing rapid verification, basic security scanning, and functional regression testing in a sandbox to ensure the immediate security and usability of personalized modifications; the global standardized evolution track performs stringent verification, including multi-scenario testing, deep security scanning, performance benchmarking, and compatibility testing, to ensure the security of the overall skills. Third, through privacy-preserving knowledge extraction methods, user-authorized data is desensitized, focusing on extracting abstract patterns of skill changes. This can be combined with federated learning or differential privacy technology to mathematically prevent the inference of users' original information. Finally, a quality gate is used to evaluate the effectiveness of skills from three dimensions: functionality, performance, and user satisfaction. A security gate is used to ensure that skills are not malicious and their behavior is controllable through code scanning, permission auditing, and sandbox verification.

[0156] In some embodiments, such as Figure 4 As shown, this application provides a dual-track self-evolution system for an intelligent agent. The dual-track self-evolution system 400 for the intelligent agent includes a local personalized evolution unit 401 and a global standardized evolution unit 402.

[0157] The Local Personalization Evolution Unit 401, deployed on the user terminal side, is primarily used to respond to the user's immediate task requirements and rapidly iterate the agent's skills locally. Specifically, the Local Personalization Evolution Unit 401 includes: a user device module for capturing user commands and system feedback, and a task execution module for executing specific tasks; when task execution deviates from expectations or the user issues a correction command, the problem detection module extracts the corresponding abnormal signal; the Local Evolution Engine receives this abnormal signal and drives the problem diagnosis and skill generation / repair module within the Local Evolution Engine to generate targeted skill patches or logic rewriting schemes; the generated schemes then enter a local verification sandbox for functional verification and security scanning to ensure basic security. Simultaneously, the generated schemes can be written to the local skill library for immediate effect. The verified schemes can then be transmitted to the authorization and desensitization module for desensitization, and with user authorization, the local operation logs are uploaded to the cloud via quality gate 1.

[0158] The global standardization evolution unit 402, deployed on a cloud server, is primarily used to aggregate local evolution data from various terminals, extract common optimization patterns, and form standardized global skills. Specifically, the global standardization evolution unit 402 includes: a cloud evolution platform that receives and aggregates data uploaded from various user terminals; a data aggregation and analysis module for in-depth mining of the aggregated data, a multi-user pattern extraction module, and a scheme optimization and generalization modification module for extracting universally applicable skill evolution patterns; the extracted and modified new skill schemes enter the global skill cluster for multi-stage functional testing, in-depth security auditing, performance benchmark testing, compatibility testing, A / B testing, and canary release verification tests; during this process, the quality gate control module is responsible for intercepting and allowing access based on functionality, performance, and user satisfaction dimensions, while the security gate control performs in-depth code auditing and malicious behavior detection; finally, standardized skills that pass dual gate control verification will be solidified into the global standard skill library and further distributed to various user terminals through a skill distribution network, thus completing a complete self-evolutionary closed loop from local innovation to global accumulation and then to capability feedback.

[0159] In this embodiment, the local personalized evolution unit 401 in the dual-track self-evolution system 400 of the intelligent agent can be a physical carrier or a set of software modules that implements the local personalized optimization process described in the first automatic evolution link in the above embodiment, and the global standardized evolution unit 402 can be a physical carrier or a set of software modules that implements the cloud-based global refinement and verification process described in the second automatic evolution link in the above embodiment.

[0160] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

[0161] It should be understood that the phrase "one embodiment" or "an embodiment" 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 an 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 numbers of the above steps / processes do not imply a sequential order of execution; the execution order of each step / 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 embodiments of this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0162] 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, article, 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, article, or apparatus. Unless otherwise specified, 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.

[0163] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0164] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0165] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0166] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0167] The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for the self-evolution of skills in an intelligent agent system, the method comprising: Based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link, the first skill stored in the local skill library is transmitted to the second self-evolution link, wherein the first skill is a skill processed based on the first automatic evolution link; The second skill is obtained from the second self-evolutionary link, wherein the second skill is a skill processed based on the second automatic evolutionary link; Based on the second skill, determine whether to replace the third skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link is used to perform personalized optimization processing on skills based on local data, wherein the local data includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link is used to perform general function optimization processing on skills based on general principles, wherein the general principles are at least principles determined by skills uploaded by multiple different users.

2. The method according to claim 1, further comprising: In response to a self-evolution event involving a target skill, the target skill is processed using local data based on a first automatic evolution link to generate a first skill; The local skill library is updated with the first skill; wherein the local data includes at least the user input feedback data recorded by the agent system during the execution of the target skill in response to the task; The step of determining whether to replace a third skill that meets the similarity condition to the second skill and is located in the local skill library based on the second skill includes: Based on the second skill, the skill that meets the similarity condition with the second skill is identified as the first skill, and the third skill that meets the similarity condition with the second skill is identified as the first skill; Determine whether to replace the first skill with the second skill, wherein the second skill is a skill generated by the second automatic evolution link based on general principles and by optimizing the general functions of the first skill; The self-evolution event is the initiation of evolutionary processing of the target skill when at least one of the following events is detected: task execution failure, user dissatisfaction with task results, detection of skill defects, or target skill performance falling below a preset threshold.

3. The method according to claim 1, wherein the step of transmitting the first skill stored in the local skill library to the second self-evolutionary link based on the collaborative mechanism of the first automatic evolutionary link and the second self-evolutionary link comprises: Based on the collaborative mechanism of the first automatic evolution link and the second self-evolution link, the local data corresponding to the first skill is transmitted to the second self-evolution link, so that the second self-evolution link can extract the common functions of the same category of skills and determine the general principles through the skills uploaded by multiple different users and the local data corresponding to the skills.

4. The method according to claim 1 or 2, wherein determining whether to replace a third skill that satisfies a similarity condition to the second skill and is located in the local skill library based on the second skill comprises: Based on the second skill, the skill that meets the similarity condition with the second skill is identified as the third skill; If the user feedback receives first confirmation, the third skill is replaced with the second skill. If a second confirmation is received from the user, the third skill will remain in the local skill library.

5. The method according to claim 1 or 2, wherein determining whether to replace a third skill that satisfies a similarity condition to the second skill and is located in the local skill library based on the second skill comprises: Based on the second skill, the skill that meets the similarity condition with the second skill is identified as the third skill; In response to the fusion command, the third skill is adjusted based on the second skill through the first self-evolutionary link to generate a fourth skill, and the fourth skill replaces the third skill.

6. The method according to any one of claims 1-3, wherein the second self-evolving link is deployed in the cloud. The method further includes: The first skill and its corresponding local data are anonymized. A quality assessment is conducted on the first de-identified skill and the de-identified local data. The quality assessment is used to determine whether the functional integrity, performance indicators and user satisfaction indicators meet the preset requirements. If the de-identified first skill and the de-identified local data meet the quality assessment, the de-identified first skill and the de-identified local data are uploaded to the second self-evolutionary link in the cloud.

7. The method according to claim 6, wherein before outputting the second skill, the second automatic evolution link performs a generalization process on at least one of the first skill and local data of the first skill, the generalization process including removing user-personalized parameters, abstracting task processes, and retaining reusable functions; the generalization process is used to cluster the same type of skills with other users to generate general principles.

8. The method according to claim 7, wherein before outputting the second skill, the second automatic evolutionary link optimizes the first skill based on the general principle to generate the second skill, and performs cross-scenario verification, a second quality assessment, and a security assessment on the second skill; in, The cross-scenario verification includes at least one of the following: functional regression testing under multiple task scenarios, compatibility testing under different user environments, and performance benchmark testing; the second quality assessment includes at least judging whether functional integrity, performance indicators, and user satisfaction indicators meet one of the preset requirements; the security assessment includes judging whether code security, permission usage compliance, and sandbox isolation verification results meet one of the preset requirements.

9. An intelligent agent system, comprising: First automatic evolutionary link module; Second automatic evolutionary link module; A collaborative processing module is used to transmit a first skill stored in the local skill library to the second self-evolution link module, wherein the first skill is a skill processed by the first automatic evolution link module; and to obtain a second skill from the second self-evolution link module, wherein the second skill is a skill processed by the second automatic evolution link module. An update module is used to determine whether to replace the third skill based on the second skill, wherein the third skill is a skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link module is used to perform personalized optimization processing on skills based on local data, wherein the local data includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link module is used to perform general function optimization processing on skills based on general principles, wherein the general principles are at least those determined by skills uploaded by multiple different users.

10. A terminal device, comprising: Communication devices used to establish transmission channels with cloud devices; A processor is configured to transmit a first skill stored in the local skill library to the second self-evolutionary link via the transmission channel, wherein the first skill is a skill processed based on the first automatic evolutionary link; The second skill is obtained from the second self-evolutionary link through the transmission channel, wherein the second skill is a skill processed based on the second automatic evolutionary link; Based on the second skill, determine whether to replace the third skill that meets the similarity condition with the second skill and is located in the local skill library; The first automatic evolution link is used to perform personalized optimization processing on skills based on local data, wherein the local data includes at least the user input feedback data recorded by the intelligent agent system during the execution of skills; the second automatic evolution link is used to perform general function optimization processing on skills based on general principles, wherein the general principles are at least principles determined by skills uploaded by multiple different users.

Images