An agent continuous evolution method, device, equipment and storage medium

Abstract

The application discloses an agent continuous evolution method and device, equipment and storage medium. The method comprises the following steps: constructing an empty skill library to store skill assets, generating a structured conversation record based on a structured record of a target task whole process, performing multi-time scale progressive growth processing based on the conversation record to obtain an intra-conversation temporary strategy and an optimized execution result, performing attribution analysis to generate an attribution result after the conversation ends, extracting a candidate skill definition based on multiple rounds of records and the attribution result, performing upgrade gate verification and verification evaluation on the candidate skill, writing the candidate skill passing the evaluation into the skill library, continuously monitoring the skill library and downgrading or archiving invalid skills, realizing automatic precipitation of reusable skills from conversation experience of the agent, controllable and verifiable autonomous continuous evolution, improving task execution efficiency and experience reuse rate in long-term operation of the agent, reducing memory pollution and system error rate, reducing maintenance cost, and improving behavior stability and long-term performance.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and storage medium for the continuous evolution of an intelligent agent. Background Technology

[0002] The need for intelligent agents to continuously learn and optimize from interactions during long-term operation, and to achieve controllable and verifiable continuous evolution, has become a core requirement for complex business scenarios.

[0003] Existing intelligent agent evolution schemes mainly use static prompts, simple memory storage, model fine-tuning, or manual rule bases to achieve capability iteration.

[0004] However, the above methods generally suffer from problems such as the inability to effectively reuse experience, severe memory noise pollution, uncontrollable system behavior, unclear optimization direction, lack of verification for updates, and high maintenance costs. They cannot form a closed-loop continuous learning mechanism and it is difficult to achieve long-term performance improvement of the agent while ensuring stability and reliability. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method, apparatus, device, and storage medium for the continuous evolution of intelligent agents. The aim is to improve the task execution efficiency and experience reuse capability of intelligent agents during long-term operation, reduce memory pollution and error rates, achieve controllable, verifiable, and manageable continuous evolution, while reducing system maintenance costs and enhancing the stability and long-term performance of intelligent agent behavior.

[0006] The embodiments of this application disclose the following technical solutions: In a first aspect, this application provides a method for the continuous evolution of an intelligent agent, comprising: Construct an empty skill library; the skill library is used to store versioned, traceable, and reusable agent skill assets; In response to the target task input by the user, the entire execution process of the target task is recorded in a structured manner to generate a structured session record; the structured session record includes task target information, decision trajectory information, human-computer interaction correction information, and performance indicator information. Based on the structured session record, a progressive growth process with multiple time scales is performed to obtain the session execution results after temporary policy caching and iterative optimization, and the structured session record is updated synchronously. After the session ends, attribution analysis is performed on the structured session record, and the performance changes of the target task are attributed to preset attribution categories to generate attribution analysis results. Based on the attribution analysis results and multi-round structured conversation records, skill extraction processing is performed to generate candidate skill definitions; An upgrade gating verification and validation evaluation process is performed on the candidate skill definition to obtain candidate skills that pass the validation evaluation; After assigning a unique version number to the candidate skills that have passed the verification and evaluation and binding them with the corresponding regression sample set and performance baseline, they are written into the skill library to form reusable long-term skill assets. The skills in the skill library are continuously monitored for operational status, and skills that meet the preset skill obsolescence requirements are downgraded or obsolescence archived to achieve closed-loop continuous evolution.

[0007] Optionally, as described above, the step of performing progressive growth processing across multiple time scales based on structured session records to obtain the session execution results after in-session temporary policy caching and iterative optimization, and synchronously updating the structured session records, includes: The execution process of a single decision reasoning in the structured session record is subjected to real-time self-verification and repetition error correction pattern recognition to obtain a set of repetition error correction patterns within a single reasoning; The set of repeated error correction patterns is subjected to a threshold determination of the number of repetitions and pre-fixing processing. When the number of repetitions of the same error correction pattern reaches a preset first threshold, the in-session temporary policy cache available for the current session is obtained. Based on the in-session temporary policy cache, the subsequent decision-making of the current session is pre-verified and the process is accelerated to obtain the iteratively optimized session execution result, and the session execution result is synchronously updated to the structured session record.

[0008] Optionally, in the method described above, after the session ends, attribution analysis is performed on the structured session record, and the performance changes in task execution are attributed to preset attribution categories, generating attribution analysis results, including: Based on a predefined checklist, the structured session records are automatically scanned and feature pattern matched to obtain a set of candidate attribution tags for the corresponding preset attribution categories; the preset attribution categories include knowledge gap type, tool strategy type, constraint governance type, interaction model building type, and orchestration collaboration type. For each attribution factor corresponding to a label in the candidate attribution label set, a quantitative calculation of the contribution of performance changes is performed to obtain an attribution factor sequence; the attribution factor sequence is initially sorted from high to low contribution. The initial sorted attribution factor sequence is subjected to priority calibration processing, and the safety-related attribution factors of constraint governance type are adjusted to the front of the sequence according to risk level, so as to obtain the priority-calibrated attribution factor sequence. For each attribution factor in the priority-calibrated attribution factor sequence, regression testing and ablation testing are performed sequentially to verify the causal relationship between each attribution factor and performance changes, eliminate invalid attribution factors, and obtain a set of valid attribution factors. Based on the verified set of attribution factors, the corresponding attribution categories, contribution rankings, verification results, and skill update guidance are integrated to generate attribution analysis results.

[0009] Optionally, in the method described above, the step of performing skill extraction processing and generating candidate skill definitions based on the attribution analysis results and multi-turn structured conversation records includes: For the accumulated multi-round structured session records, combined with the attribution analysis results of the corresponding sessions, session classification and effective fragment filtering are performed to obtain a set of positive example sessions where the task was successfully executed, a set of negative example sessions where the task failed, and effective execution trajectory fragments in the corresponding sessions. Perform multi-trajector alignment and common step sequence extraction processing on multiple valid execution trajectory segments in the positive example session set to obtain a standardized positive example execution flow; For the set of counterexample sessions and the attribution analysis results of the corresponding sessions, perform high-frequency error links and constraint violation scenarios to locate and extract them, and obtain standardized guardrail rules. The standardized positive example execution process and standardized guardrail rules are subjected to minimum execution unit decomposition, dependency mapping and hierarchical classification to obtain atomic skills, composite skills and guardrail skills; the composite skills are composed of multiple sub-skills; The atomic skills, the combined skills, and the guardrail skills are standardized and supplemented with additional definitions to generate candidate skill definitions.

[0010] Optionally, as described above, the step of assigning a unique version number to the candidate skills that have passed the verification and evaluation, binding them to the corresponding regression sample set and performance baseline, and then writing them into the skill library to form reusable long-term skill assets includes: The candidate skills that pass the verification and evaluation are subjected to baseline calibration of performance and standardization and completion of skill meta-information to obtain skills to be added to the database; the skills to be added to the database carry fixed performance benchmark indicators and standardized execution rules. Perform effective regression sample screening and association binding processing on the multi-round verification session records and historical valid execution trajectories corresponding to the skills to be added to the database, and obtain the versioned skills to be bound to the corresponding regression sample set; The versioned skills are obtained by performing unique version number assignment and compliance verification of dependency relationships between skills on the skills to be versioned corresponding to the bound regression sample set; each versioned skill carries a unique version number and has no dependency conflicts. The versioned skills undergo a secondary verification of existing asset conflicts in the skill library and a final review of security compliance to obtain compliant versioned skills; the compliant versioned skills are those that have passed the full database verification. The compliant versioned skills are written into the skill library and the skill index directory and version history ledger of the skill library are updated synchronously to obtain long-term skill assets.

[0011] Optionally, as described above, the step of continuously monitoring the operational status of skills in the skill library and performing downgrading or obsolescence archiving on skills that meet preset obsolescence requirements to achieve closed-loop continuous evolution includes: The full-cycle operation data of each skill in the skill library is continuously collected and processed with multi-dimensional indicators to obtain a real-time operation status sequence of the skills; the real-time operation status sequence of the skills includes the skill hit rate, performance gain after hit, abnormal fluctuations, and compliance execution records; The real-time running status sequence of the skills is processed to determine the downgrade trigger conditions, resulting in a set of skills to be reviewed and a set of abnormal skills. The set of skills to be reviewed is the set of skills that will trigger downgrade review. The set of abnormal skills is the set of skills that need to be frozen immediately. The set of skills to be reviewed and the set of abnormal skills are subjected to validity verification, functional coverage verification and compliance retrospective review to obtain the set of skills to be discarded and the set of skills to be retained and optimized. The set of skills to be discarded is marked as discarded and the full-link audit log is retained to obtain archived skill files marked as discarded; The archived skill files and the set of skills that can be retained and optimized are synchronized and updated to obtain an updated and optimized skill library, so as to complete the iterative update of closed-loop continuous evolution.

[0012] Optionally, the method described above further includes: Upon receiving a new user task, the matching degree and confidence level are calculated and evaluated based on the task characteristics of the user task and the skills in the skill library. When a suitable skill is matched from the skill library for the user task, the suitable skill is directly invoked to execute the task.

[0013] Secondly, this application provides a device for the continuous evolution of an intelligent agent, comprising: An initialization module is used to build an empty skill library; the skill library is used to store versioned, traceable, and reusable agent skill assets. The structured recording module is used to respond to the target task input by the user, and to record the entire execution process of the target task in a structured manner to generate a structured session record; the structured session record includes task target information, decision trajectory information, human-computer interaction correction information, and performance index information. The multi-timescale growth module is used to perform progressive growth processing on multiple timescales based on structured session records, obtain the session execution results after temporary policy caching and iterative optimization within the session, and synchronously update the structured session records. The attribution analysis module is used to perform attribution analysis on the structured session record after the session ends, and attribute the performance changes of the target task to a preset attribution category, generating attribution analysis results; The skill extraction module is used to perform skill extraction processing based on the attribution analysis results and multi-round structured conversation records, and generate candidate skill definitions. The upgrade gating and verification evaluation module is used to perform upgrade gating verification and verification evaluation processes on the candidate skill definitions to obtain candidate skills that pass the verification evaluation; The skill asset management module is used to write the candidate skills that have passed the verification and evaluation into the skill library after assigning a unique version number and binding the corresponding regression sample set and performance baseline, so as to form reusable long-term skill assets. The degradation control module is used to continuously monitor the operational status of skills in the skill library, and to perform degradation or obsolescence archiving on skills that have degraded performance, have redundant functions, or violate security and compliance requirements, so as to achieve closed-loop continuous evolution.

[0014] Thirdly, this application provides an electronic device, the device including: a processor, and a memory communicatively connected to the processor; The memory stores the instructions that the computer executes; The processor executes computer execution instructions stored in memory to implement the agent continuous evolution method described in any of the above embodiments.

[0015] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the agent continuous evolution method described in any of the above embodiments.

[0016] Compared with the prior art, this application has the following beneficial effects: This solution constructs a skill library for storing versioned, traceable, and reusable agent skill assets, providing a standardized asset carrier for the accumulation of agent experience. This overcomes the core pain points of traditional solutions, such as implicit experience storage and the inability to systematically manage it. Combined with structured recording of the entire execution process of user-defined tasks, it can completely retain information across all dimensions, including task objectives, decision trajectories, human-computer interaction corrections, and performance metrics. This provides complete and traceable data support for subsequent attribution analysis and skill extraction. Furthermore, through progressive growth processing across multiple time scales, it achieves real-time error correction within the inference loop and in-session error correction. The time-based strategy is solidified into a layered growth model based on long-term skill accumulation within sessions. Decision-making processes can be optimized and accelerated within a single session, effectively reducing task iteration rounds and reasoning time. Precise attribution analysis after the session ends accurately maps task performance changes to preset attribution categories, achieving targeted skill updates. This avoids ineffective updates caused by indiscriminate learning, significantly improving the efficiency of skill updates. Simultaneously, skill extraction is performed based on attribution results and multi-round structured session records. Standardized execution processes and pre-emptive guardrail rules are extracted from success and failure trajectories, constructing a hierarchical structure combining atomic and combined skills. The skills system enables fine-grained and highly flexible reuse of execution experience, significantly shortening the adaptation cycle for new scenarios. Through rigorous upgrade gating verification and full-process validation and evaluation, it employs a triple gating system—cross-session consistency, verifiable performance contribution, and security compliance—along with a multi-dimensional verification mechanism. This prevents unverified erroneous information from being written into the skills library from the outset, ensuring that every skill update provides quantifiable and reproducible performance improvements. Combined with the unique version number assignment before skills are added to the library and the binding of regression sample sets with performance baselines, it achieves traceability, rollback capability, and auditability throughout the entire lifecycle of skills assets, solving the problem of error susceptibility inherent in traditional solutions. To address the issues of location and inaccurate rollback, the system continuously monitors the operational status of skills in the skill library and performs downgrading or obsolescence archiving on failed skills. This ensures the long-term streamlined and efficient operation of the skill library, avoids system performance degradation caused by skill library expansion, and forms a complete closed-loop continuous evolution mechanism of "record-attribution-update-verify-reuse-elimination". Ultimately, this effectively improves the task execution efficiency and experience reuse capability of the agent in long-term operation, reduces memory pollution and error rate, and achieves controllable, verifiable, and manageable continuous evolution. At the same time, it reduces system maintenance costs and improves the stability and long-term performance of the agent's behavior. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for continuous evolution of an intelligent agent provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an intelligent agent continuous evolution device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and accompanying drawings. It should be particularly noted that the embodiments described in this application are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] As described earlier, with the rapid iteration of artificial intelligence technology, intelligent agent systems based on large language models have been widely applied in scenarios requiring long-term stable operation, such as enterprise-level intelligent services, professional domain decision support, and industrial intelligent control. Throughout the entire lifecycle of these scenarios, intelligent agent systems need to continuously adapt to dynamically changing business requirements and complex and diverse task scenarios, while accumulating successful experiences from historical interactions and mitigating execution risks to achieve continuous optimization of task execution efficiency and accuracy. Therefore, the continuous evolution capability of intelligent agent systems has become a key indicator determining their long-term availability, business adaptability, and core value.

[0022] Currently, the mainstream technical solutions adopted in the industry for the iterative and continuous evolution of intelligent agent systems mainly fall into four categories: First, static prompt word systems, which complete task execution through fixed system prompt words and tool call configurations, requiring the construction of a complete reasoning chain from scratch for each session; second, simple memory systems, which store all interaction records through long-term memory modules, retrieving historical content as a reference in subsequent task reasoning to achieve simple reuse of experience; third, fine-tuning-based learning systems, which periodically collect historical interaction data of intelligent agents to fine-tune the parameters of a large language model base, thereby improving the overall capabilities of the intelligent agent through model iteration; and fourth, artificial rule base systems, where successful execution experiences of intelligent agents are solidified into fixed rules by humans, and the iteration of intelligent agent execution strategies is achieved through the updating and maintenance of the rule base.

[0023] However, the aforementioned existing technical solutions all suffer from intractable technical defects in practical engineering applications, failing to achieve secure, controllable, and verifiable long-term continuous evolution of intelligent agent systems. First, static prompt word systems cannot accumulate and reuse historical successful experiences. High-frequency repetitive tasks still require executing the complete inference chain. Experimental data shows that the average inference time for repetitive tasks is 8-12 seconds, with a single round consuming 2000-3000 tokens, and the execution results are inconsistent. In enterprise-level application scenarios with 100,000 repetitive tasks, the computational cost is 5-8 times higher than the fixed solution, resulting in low efficiency in repetitive task execution. Second, simple memory systems lack strict memory writing gating and conflict resolution mechanisms, easily writing user slips of the tongue, temporary preferences, and unverified erroneous information into long-term memory, causing serious memory pollution. Long-term operational test data shows that after 100 sessions, the memory conflict rate of such systems reaches 18%, and the task execution error rate increases from the initial 12% to 27%. They are also prone to creating a hallucination amplification effect and lack a source tracing mechanism, making it difficult to locate and repair errors.

[0024] Furthermore, existing solutions lack a clear learning control mechanism, making it impossible to achieve controllable skill development. Ungated continuous learning systems are prone to unpredictable behavioral drift. User survey data shows that 65% of users reported that the behavior of such systems is becoming increasingly unpredictable, and 42% of users abandoned using them due to uncontrollable system behavior. The system is prone to solidifying accidental success paths into standard processes, leading to systemic errors across different scenarios.

[0025] Meanwhile, the existing system lacks a standardized performance verification mechanism, making it impossible to prove the authenticity and reproducibility of the capability improvement. System updates are prone to introducing unknown errors, resulting in high update risks in the production environment. Operations teams find it difficult to achieve high-frequency iterations, and the system cannot be continuously optimized.

[0026] Furthermore, in existing technologies, the execution experience of intelligent agents is mostly stored implicitly in memory content or model parameters, which cannot achieve version management, rollback, and full-link auditing. When problems occur, it is difficult to locate the source of the error, and it can only be rolled back to the initial state, losing all accumulated experience. While manual rule base systems can achieve rule governance, the number of rules is huge in complex scenarios, requiring a dedicated team for maintenance, with annual maintenance costs reaching hundreds of thousands of yuan. This results in high maintenance costs and weak scenario adaptability.

[0027] In summary, existing technologies cannot construct a closed-loop continuous evolution mechanism, making it difficult to achieve long-term and continuous improvement of intelligent agent capabilities while ensuring controllable and stable system behavior.

[0028] The inventors, through research, proposed a method, device, equipment, and storage medium for the continuous evolution of intelligent agents. This is achieved by constructing a fully closed-loop skill library bootstrapping evolution system encompassing "recording-attribution-updating-verification-reuse-elimination," with a multi-timescale, progressive three-layer upgrade mechanism at its core. Combined with a strict upgrade gating verification strategy, the system first upgrades the error-correction mode within the inference loop to a temporary policy cache with a scope limited to the current session. Only after the mode meets three core conditions—cross-session consistency, verifiable performance contribution, and security compliance—is it upgraded to a long-term skill asset applicable across the entire system. This fundamentally addresses the core pain points of existing technologies, such as severe memory pollution and uncontrollable growth processes. Through an attribution-driven precise update mechanism, changes in task performance are accurately attributed to five categories: knowledge gap type, tool strategy type, constraint governance type, interaction model building type, and orchestration collaboration type. Combined with regression testing and ablation testing attribution verification steps, attribution errors are avoided. Error cascading enables targeted and precise optimization of skill updates, significantly improving system update efficiency. A reproducibility verification framework comprised of regression testing, control experiments, and time window monitoring, coupled with a lightweight / standard / deep tiered evaluation strategy based on skill impact, ensures that each skill update has quantifiable and reproducible performance improvement evidence, effectively reducing the risk of introducing new errors and guaranteeing high availability for long-term system operation. A hierarchical skill system consisting of atomic skills, combined skills, and guardrail skills enables fine-grained decomposition and flexible reuse of execution experience, significantly reducing skill maintenance costs and adaptation cycles to new scenarios. A dual-path degradation control mechanism combining passive performance monitoring and proactive event-driven mechanisms promptly freezes, downgrades, and archives skills with performance degradation, functional redundancy, or compliance violations, ensuring a long-term streamlined and efficient skill library and preventing performance degradation after long-term system operation.

[0029] This solution enables intelligent agents to automatically accumulate conversational experience into manageable, traceable, and reusable skill assets. It effectively solves five core technical problems in existing technologies: inability to accumulate experience, high risk of memory contamination, uncontrollable growth, unverifiable improvement, and inability to manage assets. Under the premise of ensuring controllable and predictable system behavior, it can achieve true continuous evolution through learning in application. In long-term operation scenarios, it can significantly improve task execution efficiency and reduce computing and maintenance costs, and has outstanding technical advantages and production-grade application value.

[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0031] See Figure 1 This figure is a flowchart illustrating a method for the continuous evolution of an intelligent agent according to an embodiment of this application. Figure 1 As shown, the method includes: S101: Build an empty skill library.

[0032] The skill library is used to store versioned, traceable, and reusable agent skill assets.

[0033] In this embodiment, for example, in a campaign-level air combat decision-making scenario, the intelligent agent system needs to provide commanders with suggestions on troop deployment and firepower allocation. Initially, the system has no skill library and needs to build an empty skill library; the skill library is used to store versioned, traceable, and reusable intelligent agent air combat decision-making skill assets, and initially there is no skill data, no temporary policy cache, and no historical session records.

[0034] S102: In response to the target task input by the user, the entire execution process of the target task is recorded in a structured manner to generate a structured session record.

[0035] Structured session records contain task objective information, decision trajectory information, human-computer interaction correction information, and performance index information.

[0036] In this embodiment, in response to the target mission input by the commander, "enemy two-aircraft formation intrusion, generate interception deployment plan", the mission input includes enemy formation parameters, three available friendly fighter jets, engagement rules constraints, and ammunition resource constraints. The entire execution process of the target mission is recorded in a structured manner, generating a structured session record.

[0037] S103: Based on the structured session record, perform progressive growth processing at multiple time scales to obtain the session execution results after temporary policy caching and iterative optimization, and update the structured session record synchronously.

[0038] In this embodiment, a progressive growth process with multiple time scales is performed based on the structured session record. It is found that the "ammunition constraint check" is repeatedly triggered 3 times in a single inference, and it is fixed as a temporary policy cache within the session. Subsequent decisions within the session are automatically performed with pre-verification to obtain the iteratively optimized session execution result, and the structured session record is updated synchronously.

[0039] Furthermore, as an achievable approach, a progressive growth process based on multiple time scales is performed on structured session records to obtain the session execution results after temporary policy caching and iterative optimization. The specific implementation method for synchronously updating the structured session records includes the following steps: S1031: Perform real-time self-verification and repetitive error correction pattern recognition processing on the execution process of single decision reasoning in the structured session record to obtain a set of repetitive error correction patterns within a single reasoning.

[0040] In this embodiment, the execution process of a single decision reasoning in the structured session record is subjected to real-time self-verification and repeated error correction pattern recognition. It is detected that the "ammunition constraint check" is triggered three times in a row during the generation of a single decision, and the set of repeated error correction patterns within the single reasoning is obtained.

[0041] S1032: Perform repetition threshold determination and pre-fixing processing on the set of repeated error correction modes. When the repetition count of the same error correction mode reaches a preset first threshold, obtain the session-in-session temporary policy cache available for the current session.

[0042] In this embodiment, the set of repeated error correction patterns is subjected to a threshold determination of the number of repetitions and pre-fixing processing. The first threshold is preset to 2 times. When the number of repetitions of the same error correction pattern reaches the preset threshold by 3 times, the "ammunition constraint check" step is moved to the in-session default verification step to obtain the in-session temporary policy cache available in the current session.

[0043] S1033: Based on the in-session temporary policy cache, perform pre-verification and process acceleration processing on subsequent decisions of the current session to obtain the iteratively optimized session execution result, and synchronously update the session execution result to the structured session record.

[0044] In this embodiment, based on the in-session temporary policy cache, the subsequent decision-making of the current session is pre-verified and the process is accelerated. The ammunition constraint verification is automatically completed before the subsequent scheme is generated, without repeated error correction iterations, and the iteratively optimized session execution result is obtained. The number of scheme iteration rounds is reduced from 4 rounds to 2 rounds, the time consumption of a single decision is reduced by 20%, and the session execution result is synchronously updated to the structured session record.

[0045] In this embodiment, by using real-time self-verification and repeated error correction pattern recognition during a single decision-making reasoning process, high-frequency repetitive optimization steps can be captured instantly during session execution. Combined with repetition threshold determination and pre-fixation processing, stable error correction patterns can be quickly converted into temporary policy caches available for the current session. This not only significantly reduces the iteration rounds of subsequent decisions in the current session, shortens reasoning time, and reduces computing resource consumption through pre-verification and process acceleration, achieving real-time performance optimization within the session, but also avoids the risk of memory pollution caused by unverified experience from the source because the temporary policy only applies to the current session and is not directly written to long-term skill assets. At the same time, the synchronously updated structured session records also provide complete and accurate data support for subsequent attribution analysis and skill extraction.

[0046] S104: After the session ends, attribution analysis is performed on the structured session record, and the performance changes of the target task are attributed to preset attribution categories to generate attribution analysis results.

[0047] In this embodiment, after the session ends, attribution analysis is performed on the updated structured session record to attribute the performance improvement of this task to preset tool strategy type and constraint governance type attribution categories, and generate attribution analysis results that include attribution category, contribution ranking, and skill update guidance.

[0048] Furthermore, as one possible approach, after the session ends, attribution analysis is performed on the structured session record, and the performance changes in task execution are attributed to preset attribution categories. The specific implementation process for generating attribution analysis results includes the following steps: S1041: Based on a predefined checklist, the structured session records are automatically scanned and feature pattern matched to obtain a set of candidate attribution labels corresponding to the preset attribution categories.

[0049] In this embodiment, based on a predefined checklist, the structured session records updated after the session ends are automatically scanned and processed using feature pattern matching. This identifies instances where the solution fails to verify remaining resources, leading to constraint violations and task execution failures, thus obtaining a set of candidate attribution labels for the corresponding constraint governance type. The preset attribution categories include knowledge gap type, tool strategy type, constraint governance type, interactive model building type, and orchestration collaboration type.

[0050] S1042: For the attribution factors corresponding to each label in the candidate attribution label set, perform quantitative calculation processing on the contribution of performance change to obtain the attribution factor sequence.

[0051] The attribution factor sequence is initially sorted from highest to lowest contribution.

[0052] In this embodiment, the performance change impact contribution of the constraint governance attribution factors in the candidate attribution label set is quantitatively calculated. The evaluation shows that fixing the constraint check can eliminate 45% of the scheme execution errors, resulting in an initial sequence of attribution factors sorted from high to low contribution.

[0053] S1043: Perform priority calibration processing on the initially sorted attribution factor sequence, and adjust the safety-related attribution factors of the constraint governance type to the front of the sequence according to the risk level, so as to obtain the priority-calibrated attribution factor sequence.

[0054] In this embodiment, priority calibration processing is performed on the initially sorted attribution factor sequence, and the safety-related attribution factors of the constraint governance type are adjusted to the front of the sequence according to risk level, instead of being sorted by contribution, to obtain the priority-calibrated attribution factor sequence.

[0055] S1044: For each attribution factor in the priority-calibrated attribution factor sequence, perform regression testing and ablation testing in sequence to verify the causal relationship between each attribution factor and performance changes, eliminate invalid attribution factors, and obtain a set of valid attribution factors.

[0056] In this embodiment, for the constraint governance attribution factors in the priority-calibrated attribution factor sequence, regression testing and ablation testing are performed sequentially to verify the causal relationship between the factor and the failure of the solution, confirm that there are no other invalid attribution factors, and obtain a set of valid attribution factors.

[0057] S1045: Based on the set of valid attribution factors, integrate the corresponding attribution categories, contribution ranking, verification results and skill update guidance to generate attribution analysis results.

[0058] In this embodiment, based on the set of verified attribution factors, the corresponding constraint governance attribution categories, contribution ranking, verification results, and guardrail skill update guidance are integrated to generate the final attribution analysis results.

[0059] In this embodiment, automated and standardized attribution of task performance changes is achieved through automatic scanning and feature pattern matching of a predefined checklist. Performance issues are accurately mapped to five pre-defined categories: knowledge gap, tool strategy, constraint governance, etc., avoiding the subjectivity and inefficiency of manual attribution. By quantifying and initially ranking the contribution of attribution factors, the core issues with the greatest impact on performance can be clearly identified, enabling reasonable priority allocation for problem solving. By specifically prioritizing constraint governance-type security attribution factors, high-risk security issues are addressed in advance, effectively avoiding systemic risks caused by the expansion of security vulnerabilities. Through dual verification of regression testing and ablation testing, the causal relationship between attribution factors and performance changes is accurately confirmed, invalid attributions are eliminated, and the cascading of attribution errors is prevented from the root. Finally, standardized attribution analysis results are generated, including attribution category, contribution ranking, verification results, and skill update guidance. This provides precise targeted guidance for subsequent skill extraction and targeted updates, significantly improving the efficiency of skill updates, significantly shortening error repair time, and ensuring the accuracy and efficiency of continuous optimization of the intelligent agent system.

[0060] S105: Based on the attribution analysis results and multi-round structured conversation records, perform skill extraction processing to generate candidate skill definitions.

[0061] In this embodiment, based on the attribution analysis results and the structured session records of the accumulated three rounds of independent successful sessions, skill extraction processing is performed to align the common step sequences of the multi-round successful trajectories and generate candidate skill definitions that include the interception deployment process.

[0062] Furthermore, as an achievable approach, the specific implementation process of performing skill extraction processing and generating candidate skill definitions based on the attribution analysis results and multi-round structured conversation records includes the following steps: S1051: For the accumulated multi-round structured session records, perform session classification and effective fragment filtering processing based on the attribution analysis results of the corresponding sessions to obtain a set of positive example sessions where the task was successfully executed, a set of negative example sessions where the task was not executed, and effective execution trajectory fragments in the corresponding sessions.

[0063] In this embodiment, the accumulated 11 rounds of structured session records are combined with the attribution analysis results of the corresponding sessions to perform session classification and effective fragment screening, resulting in a set of positive example sessions with 8 rounds of successful task execution, a set of negative example sessions with 3 rounds of failed task execution, and effective execution trajectory fragments of interception scheme generation and firepower allocation in the corresponding sessions.

[0064] S1052: Perform multi-trajectory alignment and common step sequence extraction processing on multiple valid execution trajectory segments in the positive example session set to obtain a standardized positive example execution flow.

[0065] In this embodiment, the effective execution trajectory segments of the three rounds of dual-machine formation interception in the positive example session set are processed by multi-trajectory alignment and common step sequence extraction. The personalized parameters of a single session are removed to obtain a standardized positive example execution process of "dual-machine formation interception scheme generation" that can be stably reproduced in multiple rounds of sessions and has an interpretable structure.

[0066] S1053: For the set of counterexample sessions and the attribution analysis results of the corresponding sessions, perform high-frequency error links and constraint violation scenarios to locate and extract them, and obtain standardized guardrail rules.

[0067] In this embodiment, the set of negative example sessions and the constraint governance attribution analysis results of the corresponding sessions are used to locate and extract high-frequency error links and constraint violation scenarios, so as to obtain standardized guardrail rules for "ammunition reserve constraint verification" that can be verified in advance.

[0068] S1054: Perform minimum execution unit splitting, dependency mapping and hierarchical classification on the standardized positive example execution process and standardized guardrail rules to obtain atomic skills, composite skills and guardrail skills.

[0069] In this embodiment, the standardized positive example execution process and standardized guardrail rules are split into minimum execution units, and the dependency mapping and hierarchical classification processes are performed to obtain indivisible formation identification, interception geometry selection, firepower allocation, and constraint verification atomic skills, a dual-machine formation interception combination skill composed of multiple atomic skills, and a guardrail skill for pre-constraint verification.

[0070] S1055: Perform standardized definition supplementation processing on the atomic skill, the combined skill, and the guardrail skill respectively to generate candidate skill definitions.

[0071] In this embodiment, standardized definitions of input parameters, output results, pre-trigger conditions, and post-verification rules are added to the atomic skills, the combined skills, and the guardrail skills, respectively, and integrated to generate complete candidate skill definitions.

[0072] In this embodiment, by combining attribution analysis results with session classification and effective fragment screening, high-value experience data in task execution is accurately located, eliminating interference from invalid information and significantly improving the targeting and efficiency of skill extraction. By aligning multiple trajectories and extracting common step sequences from positive example sessions, standardized execution processes that are stably reproducible and interpretable across sessions can be screened, avoiding the solidification of single, accidental successes into general skills and ensuring the reliability and universality of positive example skills. By locating high-frequency error links and constraint violation scenarios from negative example sessions and extracting standardized guardrail rules, lessons learned from failures are transformed into a proactive defense mechanism. It can proactively intercept the recurrence of similar errors, significantly improving the robustness and security of the system. By breaking down the execution process and guardrail rules into a hierarchical system of atomic skills, composite skills, and guardrail skills, it enables fine-grained reuse of execution experience. New scenarios can be quickly constructed by combining existing atomic skills, greatly shortening the development cycle of new skills and reducing skill maintenance costs. Finally, through standardized definition supplements, it provides a unified specification for subsequent skill entry, version management, matching and calling, and verification and evaluation, ensuring the governance of skill assets and providing a high-quality candidate skill foundation for the continuous evolution of intelligent agents.

[0073] S106: Perform an upgrade gating verification and verification evaluation process on the candidate skill definition to obtain candidate skills that pass the verification evaluation.

[0074] In this embodiment, after generating candidate skill definitions for corresponding interception deployment schemes, the system first performs cross-session consistency checks on the candidate skill definitions to verify the number of times the stable and effective mode corresponding to the skill appears in different independent sessions. After confirming that the mode has been stably reproduced in at least 3 independent sessions and meets the preset cross-session consistency requirements, the system then replays the candidate skill on the bound 3 regression sample sets through regression testing. This verifies that after introducing the skill, the executability rate of the task execution scheme increases from 68% to 100%, and the average iteration rounds decrease from 3.8 rounds to 1.2 rounds, with significant improvement in core performance indicators, thus completing the performance contribution verification. Subsequently, the system performs security compliance checks and existing skill conflict detection on the candidate skill to confirm that it does not violate preset combat rules, security constraints, and compliance requirements, and that there are no execution logic conflicts or dependency conflicts with existing skills in the skill library. After completing the full upgrade gating verification, the candidate skill is then... Skills marked as upgrade candidates are first released in a pre-defined low-risk task set for a canary release period of 3 sessions. During the canary release, the system first performs a second-level standard evaluation on the candidate skill through a regression testing unit. Full regression testing is completed on 20 historical failure samples and critical boundary samples to confirm that the skill fix is ​​effective and does not introduce regression corruption. Then, an A / B control experiment is used to separate the independent performance gain of the skill. The task execution performance is compared between the two scenarios with and without the skill enabled, confirming that the performance improvement effect brought by the skill, such as the reduction of iteration rounds and token consumption, is significant. At the same time, the hit rate, the performance improvement after the hit, and abnormal fluctuations of the skill are continuously counted within a sliding time window of 10 sessions to confirm that the skill's running status is continuously stable. Finally, a graded evaluation and review is completed based on the business impact of the skill. After all dimensions of verification are passed, the candidate skill that has passed the verification evaluation is obtained.

[0075] S107: After assigning a unique version number to the candidate skills that have passed the verification and evaluation and binding them to the corresponding regression sample set and performance baseline, write them into the skill library to form reusable long-term skill assets.

[0076] In this embodiment, after assigning a unique version number v1.0 to the candidate skills that have passed the verification and evaluation, binding them with three corresponding regression sample sets and a performance baseline with a scheme executability rate of ≥90%, they are written into the skill library to form a reusable dual-machine formation interception long-term skill asset.

[0077] Furthermore, as an achievable approach, the specific implementation process of assigning a unique version number to the candidate skills that have passed the verification and evaluation, binding them to the corresponding regression sample set and performance baseline, and then writing them into the skill library to form reusable long-term skill assets includes the following steps: S1071: Perform baseline calibration of the benchmark performance and standardization and completion of skill meta-information on the candidate skills that have passed the verification and evaluation to obtain the skills to be added to the database.

[0078] Among them, the skills to be included in the database carry fixed performance benchmark indicators and standardized execution rules.

[0079] In this embodiment, taking the dual-aircraft formation interception candidate skill that has passed the verification and evaluation as an example, the dual-aircraft formation interception candidate skill that has passed the verification and evaluation is subjected to baseline calibration of benchmark performance and standardization of skill meta-information. The benchmark performance baseline is ≥90% scheme feasibility rate and ≤1.5 iteration rounds on average. Meta-information such as skill name, input and output, and execution steps are completed to obtain the skill to be added to the database carrying fixed performance benchmark indicators and standardized execution rules.

[0080] S1072: Perform effective regression sample screening and association binding processing on the multi-round verification session records and historical valid execution trajectories corresponding to the skills to be added to the database, and obtain the versioned skills to be bound to the corresponding regression sample set.

[0081] In this embodiment, the three rounds of verification session records and historical valid execution trajectories corresponding to the skills to be added to the database are subjected to effective regression sample screening and association binding processing. Three regression samples covering regular scenarios and boundary scenarios are selected to obtain the versioned skills to be bound to the corresponding regression sample set.

[0082] S1073: Perform unique version number assignment and compliance verification of the dependency relationship between skills on the skills to be versioned corresponding to the bound regression sample set to obtain versioned skills.

[0083] Among them, versioned skills carry a unique version number and have no dependency conflicts.

[0084] In this embodiment, for the skills to be versioned in the bound regression sample set, a unique version number v1.0 is assigned and compliance verification of the dependency relationship between skills is performed to confirm that the atomic skills they depend on have no dependency conflicts, thus obtaining versioned skills with unique version numbers and no dependency conflicts.

[0085] S1074: Perform secondary verification of existing asset conflicts and final review of security compliance on the versioned skill to obtain a compliant versioned skill.

[0086] Among them, compliant versioned skills are those that have passed full database verification.

[0087] In this embodiment, the versioned skill is subjected to a secondary verification of existing asset conflicts in the skill library and a final review of security compliance. This confirms that it has no execution conflicts with existing skills in the skill library and does not violate air combat rules, thus obtaining a compliant versioned skill that has passed the full database verification.

[0088] S1075: Write the compliant versioned skills into the skill library and update the skill index directory and version history ledger of the skill library simultaneously to obtain long-term skill assets.

[0089] In this embodiment, the compliant versioned skills are formally written into the skill library, and the skill index directory and version history ledger of the skill library are updated synchronously to obtain reusable, traceable, and rollbackable long-term skill assets.

[0090] In this embodiment, by benchmarking the performance baseline and standardizing and completing the metadata of the verified candidate skills, a unified performance measurement standard and standardized execution rules are established for the skills, ensuring that subsequent skill performance monitoring and comparison have quantifiable benchmark basis. By screening and binding effective regression sample sets, a reproducible testing basis is provided for the subsequent verification, iteration and rollback of skills, ensuring the consistency and traceability of skill performance. By assigning unique version numbers and verifying compliance of dependency relationships, version management of skill assets throughout their entire lifecycle is realized, effectively avoiding dependency conflicts and version chaos between skills. By conducting secondary verification of existing asset conflicts in the skill library and final review of security compliance, a dual security guarantee mechanism is constructed, eliminating the risk of conflicting and non-compliant skills entering the library from the source. Finally, by writing to the skill library and synchronously updating the index directory and version history ledger, the transformation from candidate skills to reusable, traceable and auditable long-term skill assets is completed. This not only improves the efficiency of skill matching and invocation in subsequent tasks, but also provides complete data support for subsequent skill iteration, fault location and accurate rollback, realizing the standardized and systematic governance of intelligent agent experience assets.

[0091] S108: Continuously monitor the operational status of skills in the skill library, and perform downgrading or obsolescence archiving on skills that meet the preset skill obsolescence requirements to achieve closed-loop continuous evolution.

[0092] In this embodiment, the skills in the skill library are continuously monitored for operational status. After the system upgrades the integrated resource management tool, it is detected that the independent contribution of the original ammunition constraint check guardrail skill drops to 0, which meets the preset skill obsolescence requirements. The skill is then frozen, verified, and obsolescence archived, forming a closed-loop continuous evolution of "record-attribution-update-verify-reuse-elimination".

[0093] Furthermore, as an achievable method, the specific implementation process of continuously monitoring the operational status of skills in the skill library and performing downgrading or obsolescence archiving on skills that meet preset obsolescence requirements to achieve closed-loop continuous evolution includes the following steps: S1081: Continuously collect and process the full-cycle operation data of each skill in the skill library using multi-dimensional statistical methods to obtain a real-time operation status sequence of the skills.

[0094] In this embodiment, taking the ammunition constraint verification guardrail skill that has been stored in the skill library as an example, the full-cycle operation data of the ammunition constraint verification guardrail skill in the skill library is continuously collected and processed with multi-dimensional indicators to obtain the real-time operation status sequence of the skill; wherein, the real-time operation status sequence of the skill includes the skill hit rate, performance gain after hit, abnormal fluctuations, and compliance execution records.

[0095] S1082: Perform downgrade trigger condition determination processing on the real-time running state sequence of the skills to obtain the set of skills to be reviewed and the set of abnormal skills.

[0096] Among them, the skill set to be reviewed is the skill set that triggers downgrade review; the abnormal skill set is the skill set that needs to be frozen immediately.

[0097] In this embodiment, the real-time running state sequence of the skill is processed to determine the downgrade trigger condition. After the system upgrades to a comprehensive resource management tool with built-in ammunition constraint checks, the ablation test shows that the independent contribution of the skill drops to 0 and the performance continues to degrade to the preset downgrade threshold, triggering passive downgrade review and obtaining a set of skills to be reviewed.

[0098] S1083: Perform validity verification, functional coverage verification, and compliance retrospective review on the set of skills to be reviewed and the set of abnormal skills to obtain the set of skills to be discarded and the set of skills to be retained and optimized.

[0099] In this embodiment, multiple rounds of validity verification and functional coverage verification are performed on the set of skills to be reviewed. After three rounds of session verification, it is confirmed that the new tool has fully covered the function of the skill, resulting in a set of skills to be discarded and a set of skills that can be retained and optimized.

[0100] S1084: Perform obsolescence status marking and full-link audit log retention processing on the set of skills to be obsolete to obtain archived skill files marked as obsolete.

[0101] In this embodiment, the set of skills to be deprecated is marked as deprecated and the full-link audit log is retained. The complete version history, full execution records and deprecation judgment criteria of the skills are archived and stored to obtain archived skill files marked as deprecated.

[0102] S1085: Perform skill library data synchronization update processing on the archived skill files and the retainable and optimized skill set to obtain an updated and optimized skill library, so as to complete the iterative update of closed-loop continuous evolution.

[0103] In this embodiment, the archived skill files and the retainable and optimized skill set are synchronized and updated with skill library data. Obsolete skills are removed from the skill library, and the skill index directory and version history ledger are updated to obtain a streamlined and optimized healthy skill library, thus completing the closed-loop continuous evolution iterative update.

[0104] In this embodiment, by continuously collecting and statistically analyzing the full-lifecycle operational data of each skill in the skill library, it is possible to comprehensively and in real-time grasp the hit rate, performance gain after a hit, abnormal fluctuations, and compliance execution status of each skill, providing precise quantitative data support for the full lifecycle management of skills. Through a tiered degradation trigger condition judgment mechanism, skills associated with security incidents or exhibiting violations are immediately frozen to quickly prevent the spread of risks. Special reviews are initiated for skills with continuously degrading performance, ensuring both the security of system operation and avoiding the impact of misjudgments on normal business operations. Through multi-dimensional evaluation including validity verification, functional coverage verification, and compliance retrospective review, accurate identification of… By accurately distinguishing between skills that are truly ineffective, functionally redundant, or violate compliance requirements, and those that can be retained and optimized, valuable skill assets can be prevented from being mistakenly deleted. Through obsolescence status marking and full-link audit log retention, the reasons for skill degradation, review process, and obsolescence time are fully recorded to meet the compliance audit and post-event review needs of enterprise-level systems. Finally, by synchronously updating the skill library data, ineffective skills are cleaned up in a timely manner to keep the skill library streamlined and efficient, avoiding the decline in matching efficiency and system performance degradation caused by the unlimited expansion of the skill library. This forms a complete closed loop of skill "entry-monitoring-evaluation-elimination", providing a solid asset governance foundation for the long-term stability, healthy and continuous evolution of intelligent agents.

[0105] Furthermore, as an alternative approach, the above method may also include: Upon receiving a new user task, the matching degree and confidence level are calculated based on the task characteristics of the user task and the skills in the skill library.

[0106] In this embodiment, taking the dual-aircraft formation interception skill already in the database as an example, the system receives a new user task input by the commander: "Enemy dual-aircraft formation intrusion, generate interception deployment plan". The system extracts the task characteristics such as formation type, task scenario, and constraints of the user task and performs matching degree calculation and confidence evaluation with the skills in the skill library.

[0107] When a suitable skill is matched from the skill library for the user task, the suitable skill is directly invoked to execute the task.

[0108] In this embodiment, the matching degree between the task and the dual-machine formation interception skill in the skill library is calculated to be 0.95, with a confidence level of 1.0, which meets the preset matching threshold. The applicable skill is matched from the skill library, and the applicable skill is directly called to execute the task, skipping the complete inference chain. The time for a single decision is reduced from 8 seconds to 1.5 seconds, and the token consumption is reduced from 2500 to 600, quickly generating an interception deployment plan that meets the requirements.

[0109] In this embodiment, by calculating the matching degree and evaluating the confidence level based on user task characteristics, skills suitable for the current task can be accurately identified, effectively avoiding execution errors caused by mismatches and ensuring the accuracy and reliability of skill invocation. When a suitable skill is matched, it is directly invoked and executed without having to build a complete inference chain from scratch. This can significantly shorten task response time, reduce token consumption and computational costs. At the same time, the standardized skill execution process can significantly improve the consistency and stability of task execution results, solving the core problems of low efficiency and large result fluctuations in high-frequency repetitive tasks in traditional static prompt word systems. It also enables the efficient reuse of historical successful experiences, providing core support for the long-term performance improvement of the intelligent agent system.

[0110] As a specific implementation, this embodiment is based on the application of campaign-level air combat decision-making scenarios. In this scenario, the intelligent agent system needs to provide the combat commander with decision-making suggestions on air combat force deployment and firepower allocation. In the initial state of system startup, there are no reusable skill reserves. All tasks need to be processed through a complete inference chain. With the long-term operation and use of the system, successful experiences in task execution can be gradually and automatically solidified into reusable standardized skills, thereby achieving a continuous and stable improvement in decision-making performance.

[0111] After the system completes initialization and startup, the skill library is empty. All tasks are processed through a complete inference chain, and then the system enters the first task session cycle. The commander inputs the core task as an enemy two-aircraft formation intrusion, requiring the generation of a corresponding interception deployment plan. Simultaneously, core input information such as enemy formation parameters, three available friendly aircraft, engagement rules constraints, and ammunition resource constraints are provided. After receiving the task, the system performs a full-chain structured record of the entire execution process, completely recording task information and the entire decision-making process. Because there are no matching available skills in the initial state, the system enters a complete inference mode to generate an interception deployment plan. During the plan generation process, the system performs real-time self-verification and error correction in each decision-making inference stage. It identifies that the "ammunition constraint check" stage is triggered three times consecutively in a single inference, and this type of error correction occurs repeatedly within the current session. Once the number of iterations reaches 3, meeting the preset threshold of 2-3 iterations, the system automatically moves the "ammunition constraint check" step to the default verification step within this session, and caches it as a temporary strategy that can be directly invoked within the session. In subsequent decision generation steps of this session, the system automatically executes this pre-verification step, reducing the time taken for a single decision by 20% compared to the initial state, and reducing the number of iteration rounds from 4 to 2. After this session ends, the system fully records the core performance indicators of this task: task execution successful, 2 iteration rounds, response time of 8 seconds, and no constraint violations. It also retains the interaction records of the commander adopting the solution and no manual correction. The processing flow of similar tasks is repeated in the 2nd to 5th sessions. The system completes a full structured record of 5 decision-making processes. During this stage, the system's average number of iteration rounds is 3.8, the solution executability rate is 68%, and the constraint violation rate is 22%.

[0112] In the fifth session, when handling a similar enemy dual-aircraft intrusion mission, the system established a stable successful execution path. The complete process involved retrieving the enemy formation database, invoking the threat assessment model, pre-verifying engagement rules, generating a deployment plan of "two forward aircraft + one rear support aircraft," allocating weapon loadouts, and outputting the plan and risk description. This plan was ultimately successfully executed. In the subsequent eighth and eleventh sessions, the same mission situation and successful execution path reappeared independently. The system automatically recognized that this successful pattern had been independently reproduced in three different sessions, meeting the minimum requirement for cross-session consistency. The system aligned the complete trajectories of these three successful executions, extracted a common sequence of steps, and identified enemy formation parameters, available friendly forces, and engagement rule constraints as variable adaptation parameters, thereby constructing a standardized skill definition. Subsequently, the system replayed and verified this skill on three bound regression samples, confirming that after introducing this skill, the executability rate of the plan increased from 68% in the no-skill state to 100%, and the average number of iterations decreased from 3.8 rounds to 1.2 rounds. The performance improvement was significant. Simultaneously, the system completed security and compliance verification for this skill, confirming that it did not violate existing operational rules and had no execution conflicts with the existing skill library. Since the initial skill library was empty, there were no conflict items in this verification. After verification, the system marked the skill as an upgrade candidate and initially enabled it in low-risk tasks, completing a three-session gray-scale run. During the gray-scale run, the system completed a second-level standard evaluation of the skill, performing full testing on 20 regression samples with a 100% pass rate and no regression violation issues. Simultaneously, the system conducted A / B testing, comparing the performance differences between scenarios with and without the skill enabled. It was confirmed that after enabling the skill, the average iteration rounds decreased from 3.8 rounds to 1.2 rounds, and the token consumption per task decreased from 2500 to 600, demonstrating a significant performance improvement. After all gray-scale verifications passed, the skill was officially written into the skill library, assigned a unique version number v1.0, and simultaneously bound to the corresponding regression sample set and performance baseline, completely preserving the skill's version history, dependencies, and performance benchmark data.

[0113] In the 12th session, the commander again input a similar mission involving an enemy two-aircraft formation intrusion, completely consistent with the mission situation in sessions 5, 8, and 11. The system calculated the matching degree and confidence level based on the mission characteristics and skills in the skill library, matching the corresponding skill named `intercept_2ship_standard` with a matching degree of 0.95 and a confidence level of 1.0. The system directly invoked this skill to complete the mission processing, skipping the complete inference chain and quickly generating a compliant interception deployment plan. The single decision-making time decreased from the initial 8 seconds to 1.5 seconds, and the single mission token consumption decreased from 2500 to 600. During the 12th to 20th session cycle, the skill achieved an 85% hit rate in applicable scenarios, with the average number of iterations after a hit remaining stable at 1.3 rounds, and the plan's executability rate reaching 92%. Within a sliding time window of 10 sessions, the system continuously monitored the skill's performance, confirming that its hit rate remained stable within the 80%-90% range, with a stable performance improvement after a hit, and no abnormal fluctuations.

[0114] In the 18th session, the interception plan generated by the system failed to check the remaining ammunition quantity, resulting in a violation of the ammunition constraint that the fire allocation exceeded the available ammunition. The plan was rejected by the commander. After the session, the system automatically scanned and analyzed the full structured record of this session, attributing the mission failure to a constraint governance failure. The core issue was a missed constraint verification step. A contribution assessment was also conducted, confirming that fixing this constraint verification step could eliminate 45% of the execution errors, indicating a high contribution level. Based on the execution trajectory and attribution results of this failure, the system extracted the corresponding pre-verification guardrail skill and generated a skill named check_ammunit. The pre-constraint hook rule for ion_constraint is set as a mandatory check that must be performed before all fire allocation schemes are generated. As this skill is a safety constraint verification skill with the highest priority, the system directly conducts a third-level in-depth evaluation after completing the safety compliance check. During the evaluation, the system completes testing on the full regression sample set and the boundary scenario sample set to confirm that the pre-constraint verification rule can effectively intercept ammunition constraint violations and will not introduce new execution errors. After the evaluation is passed, this guardrail skill is officially written into the skill library, assigned a unique version number v1.0, and set as a mandatory verification step that is automatically triggered before all fire allocation tasks are executed.

[0115] In the 22nd session, the commander input a new type of mission: an enemy four-aircraft formation intrusion, a significant difference from the previous two-aircraft formation scenario. The system matched the previous `intercept_2ship_standard` skill based on mission characteristics, but this skill's prerequisite was an enemy formation of two aircraft, which didn't match the four-aircraft formation scenario of this mission and couldn't be directly invoked. The system then decomposed the skills in its existing skill library into their smallest execution units, resulting in `assess_formation_type` (formation recognition), `select_intercept_geometry` (interception geometry selection), and `allocate_we` (firepower allocation). The system consists of four indivisible atomic skills: apons, engagement rule verification (check_engagement_rules), and interception combination skill (intercept_4ship_formation). Based on the mission requirements of this four-aircraft formation, the system creates a new combination skill adapted to the four-aircraft formation scenario. This combination skill directly reuses the two existing atomic skills, firepower allocation and engagement rule verification, and only replaces the two atomic skills adapted to the four-aircraft formation scenario, formation identification and interception geometry selection. The newly created combination skill completes full verification in three sessions. After confirming its effective execution, it is officially written into the skill library, assigned a unique version number v1.0, and the dependencies between this combination skill and the two atomic skills are recorded simultaneously.

[0116] During the 30th session, the system completed the upgrade to the new version of the integrated resource management tool, which now includes a complete ammunition constraint check function. The system continuously monitored the full-cycle performance of the check_ammunition_constraint guardrail skill previously written into the skill library. Through ablation testing, it was confirmed that removing the skill did not change the system's task execution performance, and the skill's independent performance contribution had dropped to 0. The system then triggered a passive downgrade review process, marking the skill as frozen and suspending its automatic triggering in tasks. During the subsequent three session verification cycles, the system confirmed that the new version of the integrated resource management tool fully covered all the functions of the guardrail skill. The system then marked the skill as obsolete, transferred it to the history archive for archiving, and fully preserved the skill's entire version history for auditing purposes. At the same time, it retained a full-link audit log containing the downgrade reason, execution time, and verification process.

[0117] By the 50th session cycle, the system's skill library had formed a complete system of 15 atomic skills and 8 combined skills, and 3 obsolete skills had been archived. In all tasks, 82% of the tasks could be matched with at least one applicable skill, and the average number of iterations per task decreased from the initial 3.8 rounds to 1.6 rounds, a reduction of 58%. The overall executability of the solutions increased from the initial 68% to 89%, an increase of 31%. The task constraint violation rate decreased from the initial 22% to 3%, a reduction of 86%. The computational cost of task execution (token consumption) decreased by 45%, and the average response time decreased from the initial 8 seconds to 3 seconds, a reduction of 62.5%. Throughout the 50 sessions, the system's decision-making performance maintained a steady upward trend, fully realizing verifiable and controllable continuous learning and autonomous evolution capabilities.

[0118] Through the above implementation method, the system demonstrated the complete bootstrapping evolution process of the skill base in 50 sessions: from no skill base to accumulating 23 skills (15 atoms + 8 combinations), from complete reasoning to skill reuse, from single skills to a hierarchical system, and from skill accumulation to skill downgrading and abandonment.

[0119] As another specific embodiment, this embodiment is based on the application of user service scenarios in e-commerce enterprises. In this scenario, e-commerce enterprises deploy intelligent customer service assistants to handle user inquiries across all channels. The core coverage includes high-frequency user service scenarios such as order status inquiry, product return and exchange processing, and consumer complaint handling. In the initial state of system startup, there are no reusable skill reserves. All user inquiries need to be responded to through a complete inference chain, which has the problems of long single response time, high inference token consumption, and high operating costs.

[0120] After the system completes initialization and startup, it enters the formal user consultation handling cycle. During the processing of user consultations from the 1st to the 20th, the system responds to each consultation through a complete inference chain. During this stage, the average response time for a single consultation is 12 seconds, and the average token consumption for a single consultation round reaches 3,000. After completing the full processing of 20 consultations, the system performs structured analysis and attribution processing on the accumulated full session records, identifying "order status query" as the core high-frequency task pattern. This type of task appeared 8 times in the 20 consultations, and the process of successful execution each time was highly stable. The standard execution path is that after the user provides the order number, the system calls the order system API to query the real-time status of the order, and then outputs the result in a format that is easy for the user to understand. Based on the identified stable high-frequency task patterns, the system extracts positive example skills from the successfully executed session trajectories and names them query_order_status. Simultaneously, the system performs multi-dimensional verification and auditing of this skill. First, it confirms that the task pattern has been stably reproduced in at least three independent sessions, meeting the cross-session consistency requirement. Then, regression testing verifies that the execution efficiency and response accuracy of the order status query task are significantly improved after the skill is introduced. Furthermore, the system completes security and compliance verification of the skill, confirming that it complies with the e-commerce platform's user information protection rules and order data query compliance requirements, and has no conflict with the existing rule system. After all verifications pass, the skill is officially added to the skill library.

[0121] During the 21st to 50th user consultation processing cycles, when a user initiates an order status query consultation, the system can match the order status query skill based on the features of the consultation content and directly call the skill to complete the consultation response without having to execute the complete inference chain. The average response time for this type of task has been reduced from the initial 12 seconds to 2 seconds, and the average token consumption per consultation round has been reduced from 3,000 to 400. The task processing efficiency has been improved by 83%, which has significantly reduced the system's inference calculation cost.

[0122] During the handling of the 35th user inquiry, a user filed a complaint about a delayed refund. The system's response violated the platform's "refund processing time constraint" rule. The platform promises that refunds will be credited within 3 business days after the application is approved, but this refund had been pending for more than 5 business days. The system failed to pre-verify this time constraint in its response, and the proposed solution did not comply with platform rules, causing user dissatisfaction. After the session ended, the system performed a full attribution analysis on the entire record of this failed session, attributing the task execution failure to a constraint governance failure. The core issue was the lack of a pre-verification step for the refund time constraint. Based on the execution trajectory and attribution results of this failure, the system extracted and generated a corresponding counterexample guardrail skill, named check_refund_timeliness, and set it as a mandatory pre-verification rule before processing all refund-related tasks, ensuring that compliance verification of the refund time constraint is completed before responding to any refund-related inquiry.

[0123] During the 50th to 100th user consultation processing cycles, the system continuously extracts, verifies, and adds new skills to the database from successful and failed user consultation sessions. Ultimately, the skill library accumulated 25 standardized skills covering various e-commerce customer service scenarios, achieving a task coverage rate of 75% for all user consultations. The average response time per consultation decreased from the initial 12 seconds to 4 seconds, and the overall computational cost decreased by 60% compared to the initial state. With the continuous improvement of the system's skill library and the stable enhancement of its response capabilities, user satisfaction with the intelligent customer service increased from 7.2 / 10 to 8.9 / 10, and the proportion of consultations requiring human intervention decreased from the initial 45% to 18%. This significantly reduced the operational costs of human customer service for e-commerce enterprises, fully validating the application value of this solution in enterprise-level commercial scenarios. Through the continuous evolution mechanism of the self-bootstrapping skill library, the system can quickly adapt to the business scenario requirements of e-commerce customer service, continuously optimize service performance, and achieve a significant reduction in operating costs and a steady improvement in service quality.

[0124] See Figure 2 This figure is a schematic diagram of the structure of an intelligent agent continuous evolution device provided in an embodiment of this application. Figure 2 As shown, the device 20 includes: an initialization module 21, a structured recording module 22, a multi-timescale growth module 23, an attribution analysis module 24, a skill extraction module 25, an upgrade gating and verification evaluation module 26, a skill asset management module 27, and a downgrade control module 28.

[0125] The initialization module 21 is used to construct an empty skill library; the skill library is used to store versioned, traceable, and reusable agent skill assets; the structured recording module 22 is used to respond to the target task input by the user, and to record the entire execution process of the target task in a structured manner to generate a structured session record; the structured session record includes task target information, decision trajectory information, human-computer interaction correction information, and performance index information; the multi-timescale growth module 23 is used to perform progressive growth processing on multiple timescales based on the structured session record, to obtain the session execution results after temporary policy caching and iterative optimization within the session, and to update the structured session record synchronously; the attribution analysis module 24 is used to perform attribution analysis processing on the structured session record after the session ends, and to record the performance changes of the target task execution. The attribution process is performed by assigning the attribution to a preset attribution category and generating attribution analysis results. The skill extraction module 25 is used to perform skill extraction processing based on the attribution analysis results and multi-round structured conversation records to generate candidate skill definitions. The upgrade gating and verification evaluation module 26 is used to perform upgrade gating verification and verification evaluation processes on the candidate skill definitions to obtain candidate skills that pass the verification evaluation. The skill asset management module 27 is used to assign a unique version number to the candidate skills that pass the verification evaluation and bind them to the corresponding regression sample set and performance baseline, and then write them into the skill library to form reusable long-term skill assets. The degradation control module 28 is used to continuously monitor the running status of skills in the skill library and perform degradation or obsolescence archiving processing on skills with performance degradation, functional redundancy, or violations of security compliance requirements to achieve closed-loop continuous evolution.

[0126] Furthermore, based on the above embodiments, the multi-timescale growth module 23 is specifically used to perform real-time self-verification and repetitive error correction pattern recognition processing on the execution process of a single decision reasoning in the structured session record, to obtain a set of repetitive error correction patterns within a single reasoning; to perform repetition threshold determination and pre-fixing processing on the set of repetitive error correction patterns; when the repetition count of the same error correction pattern reaches a preset first threshold, to obtain a session-in-session temporary policy cache available for the current session; to perform pre-verification and process acceleration processing on the subsequent decision execution of the current session based on the session-in-session temporary policy cache, to obtain the iteratively optimized session execution result, and to synchronously update the session execution result to the structured session record.

[0127] Furthermore, based on the above embodiments, the attribution analysis module 24 is specifically used to automatically scan and perform feature pattern matching processing on the structured session records based on a predefined checklist to obtain a set of candidate attribution tags corresponding to preset attribution categories; the preset attribution categories include knowledge gap type, tool strategy type, constraint governance type, interaction model building type, and orchestration collaboration type; for the attribution factors corresponding to each tag in the candidate attribution tag set, a quantitative calculation of the contribution of performance changes is performed to obtain an attribution factor sequence; the attribution factor sequence is initially sorted from high to low contribution; for the attribution factors corresponding to each tag in the candidate attribution tag set, a quantitative calculation of the contribution of performance changes is performed to obtain an attribution factor sequence; the attribution factor sequence is initially sorted from high to low contribution. The initial sequence of attribution factors is subjected to priority calibration, and the safety-related attribution factors with constraints and governance are adjusted to the top of the sequence according to their risk level, resulting in a priority-calibrated attribution factor sequence. For each attribution factor in the priority-calibrated attribution factor sequence, regression testing and ablation testing are performed sequentially to verify the causal relationship between each attribution factor and performance changes, and invalid attribution factors are eliminated to obtain a set of verified valid attribution factors. Based on the set of verified valid attribution factors, the corresponding attribution categories, contribution rankings, verification results, and skill update guidance directions are integrated to generate attribution analysis results.

[0128] Furthermore, based on the above embodiments, the skill extraction module 25 is specifically used to perform session classification and effective fragment filtering processing on the accumulated multi-round structured session records, combined with the attribution analysis results of the corresponding sessions, to obtain a set of positive example sessions with successful task execution, a set of negative example sessions with failed task execution, and effective execution trajectory fragments in the corresponding sessions; to perform multi-trajectory alignment and common step sequence extraction processing on multiple effective execution trajectory fragments in the positive example session set, to obtain a standardized positive example execution flow; to perform high-frequency error links and constraint violation scenarios location and extraction processing on the negative example session set and the attribution analysis results of the corresponding sessions, to obtain standardized guardrail rules; to perform minimum execution unit splitting, dependency mapping and hierarchical classification processing on the standardized positive example execution flow and standardized guardrail rules, to obtain atomic skills, combined skills and guardrail skills; the combined skills are composed of multiple sub-skills; and to perform standardized definition supplementation processing on the atomic skills, the combined skills and the guardrail skills respectively, to generate candidate skill definitions.

[0129] Furthermore, based on the above embodiments, the skill asset management module 27 is specifically used to perform baseline calibration of the performance of the candidate skills that have passed the verification and evaluation, and to standardize and complete the skill meta-information to obtain skills to be added to the database; the skills to be added to the database carry fixed performance benchmark indicators and standardized execution rules; to perform effective regression sample screening and association binding processing on the multi-round verification session records and historical effective execution trajectories corresponding to the skills to be added to the database, to obtain skills to be versioned that are bound to the corresponding regression sample sets; to perform unique version number assignment and compliance verification processing on the skills to be versioned that are bound to the corresponding regression sample sets, to obtain versioned skills; the versioned skills carry unique version numbers and have no dependency conflicts; to perform secondary verification of existing asset conflicts in the skill library and final review of security compliance processing on the versioned skills, to obtain compliant versioned skills; the compliant versioned skills are skills that have passed the full database entry verification; to write the compliant versioned skills into the skill library and to synchronously update the skill index directory and version history ledger of the skill library, to obtain long-term skill assets.

[0130] Furthermore, based on the above embodiments, the downgrade control module 28 is specifically used to continuously collect and process multi-dimensional indicator statistics of the full-cycle operation data of each skill in the skill library to obtain a real-time operation status sequence of skills; the real-time operation status sequence of skills includes skill hit rate, performance gain after hit, abnormal fluctuations, and compliance execution records; the downgrade trigger condition judgment processing is performed on the real-time operation status sequence of skills to obtain a set of skills to be reviewed and a set of abnormal skills; the set of skills to be reviewed is the set of skills that trigger downgrade review; the set of abnormal skills is the set of skills that need to be frozen immediately; the validity verification, functional coverage verification, and compliance retrospective review processing are performed on the set of skills to be reviewed and the set of abnormal skills to obtain a set of skills to be discarded and a set of skills that can be retained and optimized; the discarded skill set is marked as discarded and the full-link audit log is retained to obtain archived skill files marked as discarded; the archived skill files and the set of skills that can be retained and optimized are updated with skill library data synchronization processing to obtain an updated and optimized skill library, so as to complete the iterative update of closed-loop continuous evolution.

[0131] Furthermore, based on the above embodiments, the skill asset management module 27 can also be used to receive new user tasks, perform matching degree calculation and confidence evaluation based on the task characteristics of the user task and the skills in the skill library; when the user task matches an applicable skill from the skill library, directly call the applicable skill to execute the task.

[0132] The intelligent agent continuous evolution device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0133] See Figure 3 The figure is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, including: Memory 11 is used to store computer programs; The processor 12 is used to implement the steps of the intelligent agent continuous evolution method described in any of the above method embodiments when executing the computer program.

[0134] In this embodiment, the device can be an in-vehicle computer, a PC (Personal Computer), or a terminal device such as a smartphone, tablet computer, handheld computer, or portable computer.

[0135] The device may include a memory 11, a processor 12, and a bus 13.

[0136] The memory 11 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 may be an internal storage unit of the device, such as the hard disk of the device. In other embodiments, the memory 11 may be an external storage device of the device, such as a plug-in hard disk, SmartMediaCard (SMC), SecureDigital (SD) card, FlashCard, etc., all equipped on the device. Furthermore, the memory 11 may include both internal and external storage units of the device. The memory 11 can be used not only to store application software and various types of data installed on the device, such as program code executing the intelligent agent continuous evolution method, but also to temporarily store data that has been output or will be output. In some embodiments, the processor 12 may be a central processing unit (CPU).

[0137] In some embodiments, processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 11 or process data, such as program code for executing a method for continuous evolution of an intelligent agent.

[0138] This bus 13 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0139] Furthermore, the device may also include a network interface 14, which may optionally include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), typically used to establish communication connections between the device and other electronic devices.

[0140] Optionally, the device may further include a user interface 15, which may include a display, an input unit such as a keyboard, and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the device and to display a visual user interface.

[0141] Figure 3 Only devices with components 11-15 are shown; those skilled in the art will understand that... Figure 3 The structure shown does not constitute a limitation on the device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0142] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a computer-readable storage medium storing computer instructions for causing the computer to perform the methods described in any of the above embodiments.

[0143] The computer-readable media in this application embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0144] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the methods described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0145] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for methods, apparatuses, electronic devices, and media, since they are basically similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions of the method embodiments. The methods, apparatuses, electronic devices, and media described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0146] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for the continuous evolution of an intelligent agent, characterized in that, include: Build an empty skill library; The skill library is used to store versioned, traceable, and reusable agent skill assets; In response to the target task input by the user, the entire execution process of the target task is recorded in a structured manner to generate a structured session record; the structured session record includes task target information, decision trajectory information, human-computer interaction correction information, and performance indicator information. Based on the structured session records, a progressive growth process with multiple time scales is performed to obtain the session execution results after temporary policy caching and iterative optimization, and the structured session records are updated synchronously. After the session ends, attribution analysis is performed on the structured session record, and the performance changes of the target task are attributed to preset attribution categories to generate attribution analysis results. Based on the attribution analysis results and multi-round structured conversation records, skill extraction processing is performed to generate candidate skill definitions; An upgrade gating verification and validation evaluation process is performed on the candidate skill definition to obtain candidate skills that pass the validation evaluation; After assigning a unique version number to the candidate skills that have passed the verification and evaluation and binding them with the corresponding regression sample set and performance baseline, they are written into the skill library to form reusable long-term skill assets. The skills in the skill library are continuously monitored for operational status, and skills that meet the preset skill obsolescence requirements are downgraded or obsolescence archived to achieve closed-loop continuous evolution.

2. The method according to claim 1, characterized in that, The process of performing progressive growth processing across multiple time scales based on the structured session records, obtaining the session execution results after temporary policy caching and iterative optimization within the session, and synchronously updating the structured session records includes: The execution process of a single decision reasoning in the structured session record is subjected to real-time self-verification and repetition error correction pattern recognition to obtain a set of repetition error correction patterns within a single reasoning; The set of repeated error correction patterns is subjected to a threshold determination of the number of repetitions and pre-fixing processing. When the number of repetitions of the same error correction pattern reaches a preset first threshold, the in-session temporary policy cache available for the current session is obtained. Based on the in-session temporary policy cache, the subsequent decision-making of the current session is pre-verified and the process is accelerated to obtain the iteratively optimized session execution result, and the session execution result is synchronously updated to the structured session record.

3. The method according to claim 1, characterized in that, After the session ends, attribution analysis is performed on the structured session record, and the performance changes of the task execution are attributed to preset attribution categories, generating attribution analysis results, including: Based on a predefined checklist, the structured session records are automatically scanned and feature pattern matched to obtain a set of candidate attribution tags for the corresponding preset attribution categories; the preset attribution categories include knowledge gap type, tool strategy type, constraint governance type, interaction model building type, and orchestration collaboration type. For each attribution factor corresponding to a label in the candidate attribution label set, a quantitative calculation of the contribution of performance changes is performed to obtain an attribution factor sequence; the attribution factor sequence is initially sorted from high to low contribution. The initial sorted attribution factor sequence is subjected to priority calibration processing, and the safety-related attribution factors of constraint governance type are adjusted to the front of the sequence according to risk level, so as to obtain the priority-calibrated attribution factor sequence. For each attribution factor in the priority-calibrated attribution factor sequence, regression testing and ablation testing are performed sequentially to verify the causal relationship between each attribution factor and performance changes, eliminate invalid attribution factors, and obtain a set of valid attribution factors. Based on the verified set of attribution factors, the corresponding attribution categories, contribution rankings, verification results, and skill update guidance are integrated to generate attribution analysis results.

4. The method according to claim 1, characterized in that, Based on the attribution analysis results and multi-round structured conversation records, skill extraction processing is performed to generate candidate skill definitions, including: For the accumulated multi-round structured session records, combined with the attribution analysis results of the corresponding sessions, session classification and effective fragment filtering are performed to obtain a set of positive example sessions where the task was successfully executed, a set of negative example sessions where the task failed, and effective execution trajectory fragments in the corresponding sessions. Perform multi-trajector alignment and common step sequence extraction processing on multiple valid execution trajectory segments in the positive example session set to obtain a standardized positive example execution flow; For the set of counterexample sessions and the attribution analysis results of the corresponding sessions, perform high-frequency error links and constraint violation scenarios to locate and extract them, and obtain standardized guardrail rules. The standardized positive example execution process and standardized guardrail rules are subjected to minimum execution unit decomposition, dependency mapping and hierarchical classification to obtain atomic skills, composite skills and guardrail skills; the composite skills are composed of multiple sub-skills; The atomic skills, the combined skills, and the guardrail skills are standardized and supplemented with additional definitions to generate candidate skill definitions.

5. The method according to claim 1, characterized in that, The process of assigning a unique version number to the candidate skills that have passed the verification and evaluation, and binding them to the corresponding regression sample set and performance baseline, and then writing them into the skill library to form reusable long-term skill assets includes: The candidate skills that pass the verification and evaluation are subjected to baseline calibration of performance and standardization and completion of skill meta-information to obtain skills to be added to the database; the skills to be added to the database carry fixed performance benchmark indicators and standardized execution rules. Perform effective regression sample screening and association binding processing on the multi-round verification session records and historical valid execution trajectories corresponding to the skills to be added to the database, and obtain the versioned skills to be bound to the corresponding regression sample set; The versioned skills are obtained by performing unique version number assignment and compliance verification of dependency relationships between skills on the skills to be versioned corresponding to the bound regression sample set; each versioned skill carries a unique version number and has no dependency conflicts. The versioned skills undergo a secondary verification of existing asset conflicts in the skill library and a final review of security compliance to obtain compliant versioned skills; the compliant versioned skills are those that have passed the full database verification. The compliant versioned skills are written into the skill library and the skill index directory and version history ledger of the skill library are updated synchronously to obtain long-term skill assets.

6. The method according to claim 1, characterized in that, The continuous monitoring of the operational status of skills in the skill library, and the downgrading or archiving of skills that meet preset obsolescence requirements to achieve closed-loop continuous evolution, includes: The full-cycle operation data of each skill in the skill library is continuously collected and processed with multi-dimensional indicators to obtain a real-time operation status sequence of the skills; the real-time operation status sequence of the skills includes the skill hit rate, performance gain after hit, abnormal fluctuations, and compliance execution records; The real-time running status sequence of the skills is processed to determine the downgrade trigger conditions, resulting in a set of skills to be reviewed and a set of abnormal skills. The set of skills to be reviewed is the set of skills that will trigger downgrade review. The set of abnormal skills is the set of skills that need to be frozen immediately. The set of skills to be reviewed and the set of abnormal skills are subjected to validity verification, functional coverage verification and compliance retrospective review to obtain the set of skills to be discarded and the set of skills to be retained and optimized. The set of skills to be discarded is marked as discarded and the full-link audit log is retained to obtain archived skill files marked as discarded; The archived skill files and the set of skills that can be retained and optimized are synchronized and updated to obtain an updated and optimized skill library, so as to complete the iterative update of closed-loop continuous evolution.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Upon receiving a new user task, the matching degree and confidence level are calculated and evaluated based on the task characteristics of the user task and the skills in the skill library. When a suitable skill is matched from the skill library for the user task, the suitable skill is directly invoked to execute the task.

8. A device for the continuous evolution of an intelligent agent, characterized in that, include: The initialization module is used to build an empty skill library; The skill library is used to store versioned, traceable, and reusable agent skill assets; The structured recording module is used to respond to the target task input by the user, and to record the entire execution process of the target task in a structured manner to generate a structured session record; the structured session record includes task target information, decision trajectory information, human-computer interaction correction information, and performance index information. The multi-timescale growth module is used to perform progressive growth processing on multiple timescales based on structured session records, obtain the session execution results after temporary policy caching and iterative optimization within the session, and synchronously update the structured session records. The attribution analysis module is used to perform attribution analysis on the structured session record after the session ends, and attribute the performance changes of the target task to a preset attribution category, generating attribution analysis results; The skill extraction module is used to perform skill extraction processing based on the attribution analysis results and multi-round structured conversation records, and generate candidate skill definitions. The upgrade gating and verification evaluation module is used to perform upgrade gating verification and verification evaluation processes on the candidate skill definitions to obtain candidate skills that pass the verification evaluation; The skill asset management module is used to write the candidate skills that have passed the verification and evaluation into the skill library after assigning a unique version number and binding the corresponding regression sample set and performance baseline, so as to form reusable long-term skill assets. The degradation control module is used to continuously monitor the operational status of skills in the skill library, and to perform degradation or obsolescence archiving on skills that have degraded performance, have redundant functions, or violate security and compliance requirements, so as to achieve closed-loop continuous evolution.

9. An electronic device, characterized in that, The device includes: a processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

Images