Multi-agent skill automatic evolution method and system

Abstract

The application belongs to the technical field of artificial intelligence, and particularly relates to a multi-agent skill automatic evolution method and system. The method comprises the following steps: constructing multiple special agents, wherein the multiple special agents each have different professional functions and share the same skill library; recording interactive feedback data of a user on an output result of an agent in a conversation process of any agent in the multiple special agents in real time; classifying the interactive feedback data recorded in a preset time period according to a preset business scenario, and identifying a feedback mode; automatically generating a candidate skill according to the identified feedback mode; pushing a confirmation prompt of the candidate skill to the user, and incorporating the candidate skill into the skill library in response to a confirmation operation of the user, so that the candidate skill can be called by any agent in the multiple special agents. The application can solve the problem that there is a lack of AI auxiliary tools in the market which are specially oriented to system integrators, can automatically learn and evolve from user use, and can be used in a full business chain.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to a method and system for automatic evolution of skills of multiple agents. Background Technology

[0002] A system integrator is a technology service provider that offers turnkey automation system projects, covering the entire chain from pre-sales planning and solution design to software development, electrical control, project implementation, and project auditing. Their technical services have the following key characteristics:

[0003] First, multiple projects are executed in parallel. System integrators typically undertake multiple client projects simultaneously, each involving different industry sectors (logistics, manufacturing, pharmaceuticals, etc.) and technical requirements. Engineers need to switch between multiple projects, and a large amount of repetitive work (such as solution creation, drawing, and equipment selection) consumes a lot of manpower.

[0004] Second, there is a high demand for knowledge reuse. The planning schemes, equipment lists, and design drawings of similar projects are highly similar, but there is a lack of effective knowledge accumulation and reuse mechanisms, resulting in "starting from scratch for each project".

[0005] Third, multi-disciplinary collaboration. A complete system integration project requires the collaboration of multiple roles, including pre-sales planning engineers, solution design engineers, software development engineers, electrical engineers, and implementation engineers. The efficiency of knowledge transfer and collaboration among these roles directly affects the project delivery cycle.

[0006] Fourth, there is a severe lack of intelligent tools. Currently, the system integration industry still relies heavily on traditional manual work methods. The penetration rate of AI tools in this industry is even lower; many engineers are still manually analyzing data using Excel, meticulously drawing CAD blueprints, and writing PLC programs from scratch. Overall, AI penetration at the enterprise level is low.

[0007] Currently, the main types of technical tools available on the market to assist system integrators include the following:

[0008] (a) General AI Programming Assistant

[0009] AI-powered programming assistants, such as GitHub Copilot and Cursor, can help engineers write code. Their technical solution involves automatically generating code snippets based on large language models and user-input comments or context.

[0010] Its shortcomings are:

[0011] (1) It can only handle code generation tasks and cannot cover non-code tasks such as solution design, drawing, and equipment selection for system integrators;

[0012] (2) Lack of knowledge of the system integration industry and lack of understanding of the business specifications and design standards of system integrators (such as the aisle width standard of logistics warehouses, the selection rules of automated equipment, etc.).

[0013] (3) The user's correction behavior to the output results cannot be recorded and learned by the system, and guidance is required every time it is used;

[0014] (ii) General conversational AI assistant

[0015] General conversational AI assistants, such as ChatGPT and Wenxin Yiyan, can answer various questions and generate text content.

[0016] Its shortcomings are:

[0017] (1) Lack of understanding of the specific business scenarios of system integrators, resulting in solutions that are often too general and fail to meet the requirements of engineering practice;

[0018] (2) It does not include professional knowledge such as design specifications, equipment parameters, and calculation standards in the system integration industry;

[0019] (3) It lacks the ability to collaborate with multiple agents and cannot simultaneously support the needs of multiple professional roles such as pre-sales planning, solution design, and software development.

[0020] (III) Fine-tuning techniques for general large models

[0021] AI model continuous optimization technologies, such as Baidu's "Large Model Data Flywheel Technology" and Huawei's "Data Flywheel Fine-tuning Method and Device" (publication number CN120234408A), use user feedback data to fine-tune or reinforce learning training large models.

[0022] The shortcomings of this type of technology are:

[0023] (1) Model fine-tuning requires a large amount of labeled data and expensive GPU computing resources, which is not suitable for the actual conditions of small and medium-sized system integrators;

[0024] (2) The fine-tuning cycle is long (usually several days to several weeks), and the user's corrective behavior cannot be immediately translated into an improvement in the agent's capabilities;

[0025] (3) Using raw user feedback data directly for model training poses a risk of leaking customer project data;

[0026] (4) The learning outcomes are implicit (model parameter changes), and users cannot perceive what the agent has "learned", nor can they review and control the learning outcomes.

[0027] (iv) Specialized engineering intelligent agents in the field of industrial automation

[0028] In recent years, specialized intelligent agents for engineering scenarios have emerged in the field of industrial automation. For example, Siemens' Eigen Engineering Agent, released in 2025, is an engineering intelligent agent deeply integrated with TIA Portal. It can assist engineers in completing PLC code writing and HMI visualization generation through natural language interaction. This intelligent agent can plan, execute, and verify industrial automation engineering tasks, achieving an efficiency improvement of 2 to 5 times compared to traditional manual methods.

[0029] However, this type of technology still has the following shortcomings:

[0030] (1) It is deeply bound to the industrial software ecosystem of a specific vendor (such as Siemens TIA Portal) and does not have cross-platform and cross-industry universality;

[0031] (2) It is mainly geared towards PLC programming and HMI design in the field of industrial automation, and cannot cover the full business chain needs of system integrators from pre-sales planning to project auditing;

[0032] (3) It also does not have the ability to “automatically learn new skills from user usage”, and the expansion of its knowledge base still depends on manual updates by the manufacturer.

[0033] In summary, the core contradiction currently facing the system integrator industry is that its business is highly dependent on the reuse of human experience and knowledge, but the market lacks AI-assisted tools that are specifically designed for the entire business chain of system integrators and can automatically learn and evolve from user usage. Summary of the Invention

[0034] To address the aforementioned issues, this application aims to provide a method for the automatic evolution of multi-agent skills. This method can solve the problem that businesses heavily rely on the reuse of human experience and knowledge, but there is a lack of AI-assisted tools on the market that are specifically designed for the entire business chain of system integrators and can automatically learn and evolve from user usage.

[0035] The purpose of this application is also to provide a multi-agent skill automatic evolution system to solve the problem that business relies heavily on human experience and knowledge reuse, but there is a lack of AI-assisted tools on the market that are specifically designed for the entire business chain of system integrators and can automatically learn and evolve from user usage.

[0036] To achieve the above objectives, the technical solution of the present invention is as follows.

[0037] A method for automatic skill evolution of multi-agent systems, characterized by comprising:

[0038] Construct multiple specialized intelligent agents, each with different professional functions, and share the same skill set; pre-build a collaborative system comprising multiple specialized intelligent agents, each with different professional functions, including but not limited to:

[0039] Pre-sales planning intelligent agent: used to assist engineers in data analysis, conceptual solution design and customer communication during the pre-sales stage;

[0040] Solution Integration Intelligent Agent: Used to assist engineers in drawing construction drawings and compiling bill of quantities after winning the bid;

[0041] Project management intelligent agent: used to assist project managers in WBS decomposition, Gantt chart generation, and resource scheduling;

[0042] Software integrated intelligent agents: used to assist software engineers in generating system interface code and data migration scripts;

[0043] Electrical control intelligent agent: used to assist electrical engineers in generating electrical drawings and PLC programs;

[0044] Implementation delivery agent: used to assist implementation engineers in compiling deliverables and archiving projects;

[0045] Project Audit Agent: Used to assist audit specialists in conducting project cost audits and compliance checks.

[0046] The multiple dedicated intelligent agents share the same skill library, and the candidate skills generated by each intelligent agent during operation can be included in the skill library for all intelligent agents to call.

[0047] The system records interactive feedback data of the user on the output results of any of the multiple dedicated intelligent agents during the conversation process; during the collaborative work between the system integrator's engineer and any dedicated intelligent agent, when the engineer is dissatisfied with the output result of the intelligent agent and issues a correction command, the system captures and stores the following correction behavior data in real time:

[0048] The engineer's correction instructions include text such as "Change the channel width from 3.5 meters to 4 meters", "Use the median for calculation", and "Add a growth rate field".

[0049] The output before revision includes scheme description, code snippets, drawing parameters, equipment list, calculation reports, etc.

[0050] The corrected output content;

[0051] Correct the context information of the occurrence, including the current project identifier, the current agent identifier, the current task type, and the industry sector;

[0052] Correct the execution result of the instruction, such as success or failure;

[0053] The types of projects affected by this correction include e-commerce logistics centers, pharmaceutical warehouses, and manufacturing production lines.

[0054] The aforementioned correction behavior data is stored in a structured form in the session correction record table. Each correction record includes a project identifier, user identifier, agent identifier, session identifier, timestamp, correction instruction, comparison of content before and after correction, and execution status, and is associated with the corresponding industry domain tag.

[0055] Specifically, the correction behavior data does not include specific business data of the customer project, such as customer name, specific inventory data, specific purchase price, etc. It only records the pattern characteristics and parameter changes of the correction behavior to ensure the privacy and security of customer project data.

[0056] The interactive feedback data recorded within a preset time period is categorized according to preset business scenarios, and feedback patterns whose frequency and success rate both reach preset thresholds are identified within each business scenario. At a preset time, preferably during the early morning off-peak business period, the system performs an offline analysis task and performs the following operations:

[0057] Read unprocessed correction records within a preset time period, preferably the last 7 days, from the correction record table;

[0058] Correction records are categorized according to the system integrator's business scenarios, which include, but are not limited to:

[0059] Pre-sales planning scenarios include adjustments to parameters such as area and throughput, adjustments to the layout of the solution, and changes in equipment selection.

[0060] Scheme design scenarios: drawing dimension correction, equipment positioning adjustment, and increase / decrease in project scope;

[0061] Software development scenarios: changes in calculation methods, adjustments to data structures, and modifications to interface parameters;

[0062] Project implementation scenario: master data field mapping adjustment, import format change;

[0063] Project management scenarios: schedule adjustments, resource allocation changes, milestone modifications;

[0064] Within each business scenario, further grouping and clustering are performed according to a modified pattern, which includes, but is not limited to:

[0065] Parameter threshold adjustment, such as adjusting the anomaly detection threshold from 3 times the standard deviation to 2.5 times the standard deviation;

[0066] Changes in calculation methods, such as changing the calculation of the average to the calculation of the median;

[0067] Changes to the output format, such as adding specific fields, adjusting the output order, or adding language versions;

[0068] Process changes, such as adding data preprocessing steps or adjusting approval nodes;

[0069] Design specifications have been changed, such as adjusting the passageway width from 3.5 meters to 4 meters;

[0070] Code structure changes, such as changing synchronous calls to asynchronous calls or adding exception handling logic;

[0071] Statistics were compiled on the frequency of occurrence and success rate of each correction mode in the same business scenario.

[0072] When the number of occurrences of a certain correction pattern reaches the first preset threshold, preferably 3 times, and the execution success rate exceeds the second preset threshold, preferably 80%, the correction pattern is marked as a high-frequency effective correction pattern, and its business scenario and applicable project type are recorded.

[0073] Based on the identified feedback patterns, candidate skills are automatically generated; for correction behaviors marked as high-frequency effective correction patterns, the system automatically performs candidate skill generation operations.

[0074] Extract common correction logic from multiple correction records of this correction pattern;

[0075] The specific correction operation is abstracted into parameterizable general logic, and the parameters include, but are not limited to: input data format, threshold parameter, and output format option;

[0076] Identify the business scenarios and agent types to which this correction mode applies;

[0077] Based on the abstracted correction logic, a code template for candidate skills is automatically generated. The code template includes:

[0078] Skill names and descriptions are automatically generated, including explanations of applicable scenarios;

[0079] Input parameter definition;

[0080] The core processing logic is extracted from the correction mode;

[0081] Output format definition;

[0082] Applicable business scenario tags and intelligent agent type tags;

[0083] Calculate a comprehensive score for the candidate skills, the comprehensive score being weighted based on at least one of the following factors:

[0084] The frequency of occurrence of this correction pattern, with a weight of 40%;

[0085] Execution success rate, weight: 30%;

[0086] Engineers' proactive positive feedback, such as liking, saving, or explicitly stating "this is useful," carries a weight of 20%.

[0087] The degree of quality improvement of the results before and after correction, weighted at 10%;

[0088] Candidate skills whose comprehensive scores reach the third preset threshold are stored in the candidate skill table, marked as "pending confirmation", and associated with the corresponding business scenario tags and the list of applicable intelligent agents.

[0089] The system pushes a confirmation prompt for the candidate skill to the user and, in response to the user's confirmation, adds the candidate skill to the skill library, making the candidate skill available for use by any of the multiple dedicated agents. The system checks whether the engineer has any candidate skills to be confirmed the next time they use any agent.

[0090] If there are candidate skills that need to be confirmed, the system will send a confirmation prompt to the engineer. The confirmation prompt includes:

[0091] The functional descriptions of candidate skills should be presented in a way that the engineer can understand.

[0092] Explain the source of this skill, including which corrective actions by the engineer it was generated from, and which business scenario it belongs to;

[0093] Expected effects and applicable scenarios for using this skill;

[0094] The scope of this skill's benefits is limited to individual use / team sharing / enterprise-wide sharing only;

[0095] In response to the engineer's confirmation, the candidate skill is officially added to the engineer's personal skill library, and can be added to the enterprise skill library after review by the enterprise administrator;

[0096] In response to the engineer's rejection action, the candidate skill is marked as rejected and will no longer be pushed to that engineer.

[0097] If the engineer does not confirm or refuse the operation within the preset time, preferably 7 days, the candidate skill will automatically expire and the status will be marked as expired.

[0098] System integrator skills: scenario adaptation and continuous iteration

[0099] Record the number of times each skill is invoked, the success rate of execution, the average execution time, and the engineer's rating;

[0100] The effectiveness of skill usage is categorized and statistically analyzed based on business scenarios, such as pre-sales planning, solution design, and software development, and project types, such as e-commerce logistics, pharmaceutical warehousing, and manufacturing.

[0101] When the effectiveness of a skill in a specific business scenario or project type is lower than the fourth preset threshold, the skill is marked as needing optimization in that scenario.

[0102] Based on skill effect data, automatically generate skill optimization suggestions or trigger automatic skill version updates;

[0103] Automatically recommend high-value skills to other engineers within the same company or to other projects within the same business scenario.

[0104] Optionally, the interactive feedback data includes: the user's correction instruction, the output content before correction, the output content after correction, the business scenario label when the correction occurred, and the agent identifier.

[0105] Optionally, the business scenarios include at least one of the following: pre-sales planning scenario, solution integration scenario, project management scenario, software integration scenario, electrical control scenario, implementation and delivery scenario, and project audit scenario.

[0106] Optionally, the feedback mode includes at least one of the following: parameter threshold adjustment, calculation method change, output format change, processing flow change, design specification change, and code structure change.

[0107] Optionally, the step of automatically generating candidate skills includes:

[0108] Extract common feedback logic from multiple interactive feedback records;

[0109] Abstract specific feedback operations into parameterizable general logic;

[0110] Identify the business scenarios and agent types to which the feedback pattern applies;

[0111] Generate a code template for candidate skills. The code template includes a skill name, skill description, input parameter definition, core processing logic, output format definition, applicable business scenario tags, and applicable intelligent agent type tags.

[0112] Optionally, it also includes: calculating a comprehensive score for candidate skills, the comprehensive score being calculated based on a weighted average of at least one of the following factors: frequency of occurrence of feedback patterns, execution success rate, positive user feedback, and the degree of quality improvement of results before and after feedback.

[0113] Optionally, the step of pushing a candidate skill confirmation prompt to the user includes:

[0114] When the user uses any intelligent agent again, detect whether the user has any candidate skills to be confirmed;

[0115] If such a candidate skill exists, a confirmation prompt is pushed to the user. The confirmation prompt includes a functional description of the candidate skill, an explanation of the source of the candidate skill, the expected effect of using the candidate skill, and applicable scenarios.

[0116] Optionally, it also includes: recording the usage effect data of each skill in the skill library after it is invoked, classifying and statistically analyzing the usage effect data according to business scenarios; marking the skill as needing optimization in a specific business scenario when the usage effect of a skill is lower than a preset threshold; and generating skill optimization suggestions or triggering automatic updates of skill versions based on the usage effect data.

[0117] Optionally, the method is applied to a multi-agent collaborative platform of a system integrator, wherein the multiple dedicated agents include at least one of the following: pre-sales planning agent, solution integration agent, project management agent, software integration agent, electrical control agent, implementation and delivery agent, and project audit agent.

[0118] This invention also provides a multi-agent skill automatic evolution system, characterized in that it includes:

[0119] A multi-agent collaboration module is used to provide multiple specialized agents with different professional functions, and the multiple specialized agents share the same skill library;

[0120] The feedback recording module is used to capture and store the user's interactive feedback data in real time during the user's conversation with any dedicated intelligent agent;

[0121] The pattern recognition module is used to classify the stored feedback records according to preset business scenarios and identify high-frequency and effective feedback patterns in each business scenario.

[0122] The candidate skill generation module is used to automatically generate candidate skills based on the identified feedback patterns;

[0123] The user confirmation module is used to push candidate skill confirmation prompts to users and receive confirmation or rejection from users.

[0124] The skill management module is used to include candidate skills confirmed by the user into the skill library, so that the candidate skills can be invoked by any of the multiple dedicated intelligent agents.

[0125] Compared with the prior art, the beneficial effects of this application are as follows:

[0126] 1. This invention enables specialized intelligent agents to automatically learn new capabilities from users' daily use. Compared to existing technologies that require developers to manually code and define capabilities, this invention significantly reduces the maintenance cost of the capability library, and the learned capabilities are naturally adapted to users' actual work scenarios and business specifications.

[0127] 2. This invention constructs a collaborative system comprising multiple specialized intelligent agents, each sharing the same capability library. A capability taught to an agent by a user in one business scenario can be reused by other specialized intelligent agents in different scenarios, achieving cross-agent knowledge accumulation and collaboration.

[0128] 3. In this invention, feedback behavior is captured in real time, offline analysis tasks can be configured with execution cycles as needed, and candidate capability generation and user confirmation can be completed the following day. Compared to the learning cycle of several days to several weeks based on model fine-tuning in existing technologies, this invention achieves a rapid closed loop from user correction to capability availability.

[0129] 4. This invention only records the pattern characteristics and parameter changes of feedback behavior, and does not store the user's specific business data. Compared with existing technologies that directly use raw user feedback data for model training, this invention effectively protects the user's business secrets and data privacy.

[0130] 5. This invention enables a valid feedback pattern in a scenario to be identified and transformed into a capability, which can be directly reused in subsequent similar scenarios, effectively solving the pain points of users having a lot of repetitive work and difficulty in knowledge accumulation.

[0131] 6. In this invention, user confirmation is required before candidate capabilities are included in the database. Compared to the existing technology where user feedback data is directly used for model training, this invention allows users to clearly understand what the agent has learned and to review and control the learning outcomes.

[0132] 7. This invention uses a capability effectiveness tracking mechanism to classify and statistically analyze the usage effects of capabilities according to business scenarios, thereby enabling continuous monitoring and iterative optimization of capabilities. Attached Figure Description

[0133] Figure 1 This is a flowchart illustrating a multi-agent skill automatic evolution method according to the present invention. Detailed Implementation

[0134] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0135] See Figure 1This embodiment provides a method for automatic evolution of multi-agent skills, including:

[0136] S1: Construct multiple specialized intelligent agents, each with different professional functions, but sharing the same skill set; pre-build a collaborative system containing multiple specialized intelligent agents, each with different professional functions, including but not limited to:

[0137] Pre-sales planning intelligent agent: used to assist engineers in data analysis, conceptual solution design and customer communication during the pre-sales stage, mainly serving pre-sales engineers;

[0138] Solution Integration Intelligent Agent: Used to assist engineers in drawing construction drawings and compiling bills of quantities after winning the bid; primarily serving solution design engineers.

[0139] Project management intelligent agent: used to assist project managers in WBS decomposition, Gantt chart generation and resource scheduling, mainly serving project managers;

[0140] Software integrated intelligent agents: used to assist software engineers in generating system interface code and data migration scripts, primarily serving software engineers;

[0141] Electrical control intelligent agent: used to assist electrical engineers in generating electrical drawings and PLC programs, primarily serving electrical engineers;

[0142] Implementation Delivery Agent: Used to assist implementation engineers in compiling deliverables and archiving projects, primarily serving implementation engineers;

[0143] Project Audit Agent: Used to assist audit specialists in conducting project cost audits and compliance checks, primarily serving audit specialists.

[0144] Multiple specialized agents share the same skill library. Candidate skills generated by each agent during operation can be included in the skill library and used by all agents. For example, the "ABC classification-based shelf selection method" learned by the pre-sales planning agent in an e-commerce logistics center project can be directly used by the solution integration agent in subsequent projects.

[0145] S2: Real-time recording of interactive feedback data from a user during a conversation with any of the multiple dedicated agents, showing the user's responses to the agent's output. For example, a pre-sales engineer from a system integrator is using a pre-sales planning agent to generate a conceptual plan for an e-commerce logistics center project. The agent initially generates a layout plan with aisle widths of 3.5 meters. Based on the project's actual conditions, such as the warehouse primarily handling small items and requiring more shelving space, the engineer issues a correction instruction: "Change the aisle width to 4 meters."

[0146] The system captures this corrective action in real time and records the following data:

[0147] Fields Record content Project Identifier PRJ-2026-007 User ID U_ZHANG Intelligent agent identification AGENT_PRESALES Business Scenarios Pre-sales planning Project Type E-commerce logistics center Timestamp 2026-07-09 14:23:45 Correction instructions "The passageway width has been changed to 4 meters." Parameters before correction {"channel_width": 3.5} Corrected parameters {"channel_width": 4.0} Execution result success

[0148] This correction log does not contain sensitive business information such as customer names or specific SKU data; it only records the pattern characteristics and parameter changes of the correction behavior.

[0149] S3: Classify interactive feedback data recorded within a preset time period according to preset business scenarios, and identify feedback patterns whose frequency and success rate both reach preset thresholds within each business scenario; for example, the system executes an offline analysis task at 2:00 AM daily. Within the analysis window of the most recent 7 days, the system classifies and statistically analyzes correction records according to business scenarios:

[0150] Recognition results in the pre-sales planning scenario:

[0151] In pre-sales planning scenarios, the system identified the following correction patterns:

[0152] "Channel Width Adjustment" mode: The operation of adjusting the channel width from the default value of 3.5 meters to 4.0 meters occurred in 3 different projects, with a frequency of 3 times and a success rate of 100%;

[0153] "ABC category threshold adjustment" mode: In both e-commerce logistics center projects, the operation of adjusting the percentage threshold of category A items from 30% to 40% occurred twice, with a success rate of 100%.

[0154] "Gravity racking selection" mode: In both projects, the operation of changing standard racks to gravity racks occurred twice, with a success rate of 100%.

[0155] Among them, the "channel width adjustment" mode is marked as a high-frequency effective correction mode when it appears 3 times, reaching the preset threshold.

[0156] For the "channel width adjustment" high-frequency and effective correction mode, the system automatically performs candidate skill generation:

[0157] The common correction logic was extracted from the three correction records of this correction mode: In the e-commerce logistics center project, the preferred channel width is 4.0 meters, instead of the default 3.5 meters;

[0158] The specific modifications are abstracted into parameterizable general logic:

[0159] Input parameters: Project type, e-commerce logistics center;

[0160] Output parameters: Recommended channel width;

[0161] Applicable scenarios: Pre-sales planning;

[0162] The code template for generating candidate skills includes:

[0163] Skill Name: "Recommendation of Aisle Width for E-commerce Logistics Centers"

[0164] Skill Description: "Automatically recommends aisle width based on project type; 4.0 meters is recommended for e-commerce logistics centers."

[0165] Input parameter definition: [{"name": "project_type", "type": "string"}]

[0166] Core processing logic: if project_type == "e-commerce logistics center": return 4.0 else: return Applicable business scenario: pre-sales planning;

[0167] Applicable agents: pre-sales planning agents, solution integration agents;

[0168] Calculate the overall score for the candidate skill:

[0169] Frequency of occurrence: 3 times (Score: 90 points)

[0170] Success rate: 100% (Score: 100 points)

[0171] Overall score: 90 × 40% + 100 × 30% = 66 points

[0172] Once the overall score reaches the preset threshold, the candidate skill is stored in the candidate skill table, and its status is "pending confirmation".

[0173] S4: Based on the identified feedback patterns, candidate skills are automatically generated; for correction behaviors marked as high-frequency effective correction patterns, the system automatically performs candidate skill generation operations:

[0174] Extract common correction logic from multiple correction records of this correction pattern;

[0175] The specific correction operation is abstracted into parameterizable general logic, and the parameters include, but are not limited to: input data format, threshold parameter, and output format option;

[0176] Identify the business scenarios and agent types to which this correction mode applies;

[0177] Based on the abstracted correction logic, code templates for candidate skills are automatically generated. The code templates include:

[0178] Skill names and descriptions are automatically generated, including explanations of applicable scenarios;

[0179] Input parameter definition;

[0180] The core processing logic is extracted from the correction mode;

[0181] Output format definition;

[0182] Applicable business scenario tags and intelligent agent type tags;

[0183] Calculate a comprehensive score for the candidate skills, which is weighted based on at least one of the following factors:

[0184] The frequency of occurrence of this correction pattern, with a weight of 40%;

[0185] Execution success rate, weight: 30%;

[0186] Engineers' proactive positive feedback, such as liking, saving, or explicitly stating "this is useful," carries a weight of 20%.

[0187] The degree of quality improvement of the results before and after correction, weighted at 10%;

[0188] Candidate skills whose comprehensive scores reach the third preset threshold are stored in the candidate skill table, marked as "pending confirmation", and associated with the corresponding business scenario tags and the list of applicable intelligent agents.

[0189] S5: Push confirmation prompts for candidate skills to the user, and in response to the user's confirmation, add the candidate skills to the skill library, making them available for use by any of the multiple dedicated agents. The system checks whether the engineer has any candidate skills to be confirmed the next time they use any agent.

[0190] If there are candidate skills that need to be confirmed, the system will send a confirmation prompt to the engineer. The confirmation prompt includes:

[0191] The functional descriptions of candidate skills should be presented in a way that the engineer can understand.

[0192] Explain the source of this skill, including which corrective actions by the engineer it was generated from, and which business scenario it belongs to;

[0193] Expected effects and applicable scenarios for using this skill;

[0194] The scope of this skill's benefits is limited to individual use / team sharing / enterprise-wide sharing only;

[0195] In response to the engineer's confirmation, the candidate skill is officially added to the engineer's personal skill library, and can be added to the enterprise skill library after review by the enterprise administrator;

[0196] In response to the engineer's rejection action, the candidate skill is marked as rejected and will no longer be pushed to that engineer.

[0197] If the engineer does not confirm or refuse the operation within the preset time, preferably 7 days, the candidate skill will automatically expire and the status will be marked as expired.

[0198] Step Six: Skills system integrator scenario adaptation and continuous iteration;

[0199] Record the number of times each skill is invoked, the success rate of execution, the average execution time, and the engineer's rating;

[0200] The effectiveness of skill usage is categorized and statistically analyzed based on business scenarios, such as pre-sales planning, solution design, and software development, and project types, such as e-commerce logistics, pharmaceutical warehousing, and manufacturing.

[0201] When the effectiveness of a skill in a specific business scenario or project type is lower than the fourth preset threshold, the skill is marked as needing optimization in that scenario.

[0202] Based on skill effect data, automatically generate skill optimization suggestions or trigger automatic skill version updates;

[0203] Automatically recommend high-value skills to other engineers within the same company or to other projects within the same business scenario.

[0204] In this embodiment, the interactive feedback data includes: the user's correction instruction, the output content before correction, the output content after correction, the business scenario label when the correction occurred, and the agent identifier.

[0205] In this embodiment, the business scenarios include at least one of the following: pre-sales planning scenario, solution integration scenario, project management scenario, software integration scenario, electrical control scenario, implementation and delivery scenario, and project audit scenario.

[0206] In this embodiment, the feedback mode includes at least one of the following: parameter threshold adjustment, calculation method change, output format change, processing flow change, design specification change, and code structure change.

[0207] In this embodiment, the step of automatically generating candidate skills includes:

[0208] Extract common feedback logic from multiple interactive feedback records;

[0209] Abstract specific feedback operations into parameterizable general logic;

[0210] Identify the business scenarios and agent types to which the feedback pattern applies;

[0211] Generate code templates for candidate skills. The code templates include skill name, skill description, input parameter definition, core processing logic, output format definition, applicable business scenario tags, and applicable intelligent agent type tags.

[0212] In this embodiment, the method further includes: calculating a comprehensive score for candidate skills, the comprehensive score being calculated based on at least one of the following factors: frequency of occurrence of feedback patterns, execution success rate, positive user feedback, and the degree of quality improvement of results before and after feedback.

[0213] In this embodiment, the step of pushing a candidate skill confirmation prompt to the user includes:

[0214] The next time a user uses any agent, detect whether the user has any candidate skills that need to be confirmed.

[0215] If such a prompt exists, a confirmation prompt will be sent to the user. The confirmation prompt will include a description of the candidate skill's function, an explanation of the candidate skill's source, the expected effect of using the candidate skill, and the applicable scenarios.

[0216] In this embodiment, it also includes: recording the usage effect data of each skill in the skill library after it is invoked, classifying and statistically analyzing the usage effect data according to business scenarios; marking the skill as needing optimization in a specific business scenario when the usage effect of a skill is lower than a preset threshold; and generating skill optimization suggestions or triggering automatic updates of skill versions based on the usage effect data.

[0217] In this embodiment, the method is applied to a multi-agent collaborative platform of a system integrator, and the multiple dedicated agents include at least one of the following: pre-sales planning agent, solution integration agent, project management agent, software integration agent, electrical control agent, implementation and delivery agent, and project audit agent.

[0218] This invention also provides a multi-agent skill automatic evolution system, characterized in that it includes:

[0219] The multi-agent collaboration module is used to provide multiple specialized agents with different professional functions, and these agents share the same skill set.

[0220] The feedback recording module is used to capture and store the user's interactive feedback data in real time during the user's conversation with any dedicated intelligent agent;

[0221] The pattern recognition module is used to classify the stored feedback records according to preset business scenarios and identify high-frequency and effective feedback patterns in each business scenario.

[0222] The candidate skill generation module is used to automatically generate candidate skills based on the identified feedback patterns;

[0223] The user confirmation module is used to push candidate skill confirmation prompts to users and receive confirmation or rejection from users.

[0224] The skill management module is used to include user-confirmed candidate skills into the skill library, so that the candidate skills can be invoked by any of the multiple dedicated intelligent agents.

[0225] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatic skill evolution of multi-agent systems, characterized in that, include: Construct multiple specialized intelligent agents, each with different professional functions, and share the same skill set; Real-time recording of interactive feedback data from the user to the output results of any of the multiple dedicated intelligent agents during a session; The interactive feedback data recorded within a preset time period is classified according to preset business scenarios, and feedback patterns whose frequency and success rate both reach preset thresholds are identified in each business scenario. Candidate skills are automatically generated based on the identified feedback patterns; A confirmation prompt for the candidate skill is pushed to the user, and in response to the user's confirmation operation, the candidate skill is added to the skill library, so that the candidate skill can be invoked by any of the multiple dedicated intelligent agents.

2. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, The interactive feedback data includes: the user's correction instructions, the output content before correction, the output content after correction, the business scenario label when the correction occurred, and the agent identifier.

3. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, The business scenarios include at least one of the following: pre-sales planning scenario, solution integration scenario, project management scenario, software integration scenario, electrical control scenario, implementation and delivery scenario, and project audit scenario.

4. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, The feedback mode includes at least one of the following: parameter threshold adjustment, calculation method change, output format change, processing flow change, design specification change, and code structure change.

5. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, The steps for automatically generating candidate skills include: Extract common feedback logic from multiple interactive feedback records; Abstract specific feedback operations into parameterizable general logic; Identify the business scenarios and agent types to which the feedback pattern applies; Generate a code template for candidate skills. The code template includes a skill name, skill description, input parameter definition, core processing logic, output format definition, applicable business scenario tags, and applicable intelligent agent type tags.

6. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, It also includes: calculating a comprehensive score for candidate skills, the comprehensive score being calculated based on a weighted average of at least one of the following factors: frequency of occurrence of feedback patterns, execution success rate, positive user feedback, and the degree of quality improvement of results before and after feedback.

7. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, The step of pushing a candidate skill confirmation prompt to the user includes: When the user uses any intelligent agent again, detect whether the user has any candidate skills to be confirmed; If such a candidate skill exists, a confirmation prompt is pushed to the user. The confirmation prompt includes a functional description of the candidate skill, an explanation of the source of the candidate skill, the expected effect of using the candidate skill, and applicable scenarios.

8. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, Also includes: Record the usage effect data of each skill in the skill library after it is invoked, and classify and statistically analyze the usage effect data according to business scenarios; When the effectiveness of a skill in a specific business scenario is lower than a preset threshold, the skill is marked as needing optimization in that scenario. Based on the usage effect data, skill optimization suggestions are generated or automatic updates of skill versions are triggered.

9. The method for automatic evolution of multi-agent skills according to claim 1, characterized in that, The method is applied to a multi-agent collaborative platform of a system integrator, wherein the multiple dedicated agents include at least one of the following: pre-sales planning agent, solution integration agent, project management agent, software integration agent, electrical control agent, implementation and delivery agent, and project audit agent.

10. A multi-agent skill automatic evolution system for running the multi-agent skill automatic evolution method according to any one of claims 1-9, characterized in that, include: A multi-agent collaboration module is used to provide multiple specialized agents with different professional functions, and the multiple specialized agents share the same skill library; The feedback recording module is used to capture and store the user's interactive feedback data in real time during the user's conversation with any dedicated intelligent agent; The pattern recognition module is used to classify the stored feedback records according to preset business scenarios and identify high-frequency and effective feedback patterns in each business scenario. The candidate skill generation module is used to automatically generate candidate skills based on the identified feedback patterns; The user confirmation module is used to push candidate skill confirmation prompts to users and receive confirmation or rejection from users. The skill management module is used to include candidate skills confirmed by the user into the skill library, so that the candidate skills can be invoked by any of the multiple dedicated intelligent agents.

Images