A Method and System for Generating Intelligent Agent Task Skills Based on Heterogeneous Open Worlds

By using a heterogeneous open-world-based agent task skill generation method, candidate operation factors are identified and judged, and the generated skill packages can adapt to different situations. This solves the problems of incomplete demand discovery and insufficient skill reliability in existing technologies, and achieves efficient and reliable skill generation.

CN123114418APending Publication Date: 2026-11-17SHANDONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611276469.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for generating intelligent agent skills suffer from problems such as incomplete demand discovery, limited retrieval scope, generalization of local experience, and insufficient skill reliability, which prevent the generated skills from being reusable stably in different contexts.

Method used

A method for generating intelligent agent task skills based on heterogeneous open worlds is adopted. This method receives task skill requirements, identifies candidate operation factors, constructs paired evidence to obtain targets, obtains evidence and organizes it into structured source discovery, groups it according to program decisions, executes program admission judgment, and compiles and generates installable skill packages.

Benefits of technology

It improves the reliability and applicability of skill generation, and the generated skill packages can improve the overall task pass rate in multi-domain task testing, adapt to task changes within a certain range, avoid the injection of misleading content, and reduce creation time.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114418A_ABST
    Figure CN123114418A_ABST
Patent Text Reader

Abstract

This disclosure provides a method and system for generating intelligent agent task skills based on a heterogeneous open world, relating to the field of artificial intelligence technology. The method includes: receiving task skill requirements and identifying candidate operational factors; constructing paired evidence to obtain targets and acquiring evidence; organizing the acquired evidence into structured source discovery; grouping the evidence according to the procedural decisions supported by the structured source discovery, and summarizing each structured source discovery group to obtain candidate programs; constructing an admission record for each candidate program and performing program admission judgment to obtain admission results; introducing execution and encapsulation constraints based on general instructions and limited-scope instances in the admission results, and compiling to generate a skill package; and using skill syntax to perform grammatical guidance verification on the generated skill package. This disclosure, through implicit requirement discovery, structured evidence organization, and program admission judgment, distinguishes between general instructions and local examples, avoiding the direct inclusion of insufficient or conflicting evidence into skills.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, specifically to a method and system for generating intelligent agent task skills based on a heterogeneous open world. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Currently, intelligent agents based on artificial intelligence and large language models have been applied in scenarios such as information retrieval, code generation, data analysis, and automated task execution. To enable intelligent agents to complete specific tasks, corresponding skills can be configured for them. These skills typically include the operational steps required to complete the task, methods for using tools, and related knowledge.

[0004] Existing methods for creating intelligent agent skills generally include two approaches: direct generation and retrieval-enhanced generation. Direct generation involves generating corresponding skill content directly using a language model based on the user's provided technical requirements. Retrieval-enhanced generation, on the other hand, retrieves relevant information from sources such as technical documents, code repositories, and web pages, and then generates skill content based on the retrieval results.

[0005] However, the above two methods for creating and generating agent skills still have the following limitations: (1) Incomplete requirements discovery: The skill requirements provided by users are usually brief and do not clearly explain the applicable conditions, abnormal situations and operation boundaries that will affect the task execution method, resulting in the omission of necessary content in the generated skills.

[0006] (2) The search scope is limited: searching only based on the original skill requirements makes it easy to obtain information similar to the current task repeatedly, and it is difficult to find factors that are not explicitly described in the requirements but actually need to be considered.

[0007] (3) Local experience is easily generalized incorrectly: Operation methods in open-world materials are usually only applicable to specific objects, tools or environments. If they are written into skills without judgment, they may be mistakenly regarded as general operation rules.

[0008] (4) Insufficient skill reliability: Existing methods lack a systematic judgment on source evidence, scope of application and content conflicts. The generated skills may only be applicable to a small number of similar tasks and are difficult to reuse stably in different situations. Summary of the Invention

[0009] To address the aforementioned issues, this disclosure proposes a method and system for generating intelligent agent task skills based on heterogeneous open worlds. Taking insufficiently described skill requirements, source access specifications, and execution and encapsulation constraints as input, it obtains evidence from heterogeneous open world sources, discovers implicit requirements in the skill requirements that are missing but will change the behavior of the intelligent agent, determines whether candidate programs can enter the skill based on the applicable scope supported by the evidence, and compiles the content that passes the admission judgment into an installable skill package.

[0010] According to some embodiments, the present disclosure adopts the following technical solutions: A method for generating agent task skills based on heterogeneous open worlds includes: Receive task skill requirements and identify candidate operational factors; Construct paired evidence acquisition objectives, transform each paired evidence acquisition objective into a focused research question to identify implicit needs in candidate operational factors, and acquire evidence; organize the acquired evidence into structured source discovery; The source discovery groups are grouped according to the procedural decisions supported by the structured source discovery method. Candidate procedures are obtained by summarizing each group. An admission record is constructed for each candidate procedure, and the admission judgment is performed to obtain the admission result. Based on the general instructions and limited-scope instances in the admission results, execution and encapsulation constraints are introduced to compile and generate skill packs. The generated skill pack is grammatically guided and validated using skill grammar. Once validation is complete, the output is a skill pack that can be installed and loaded on demand by the agent during inference.

[0011] According to some embodiments, the present disclosure adopts the following technical solutions: A heterogeneous open-world-based intelligent agent task skill generation system includes: The initialization module is used to receive task skill requirements and identify candidate operation factors; The implicit demand discovery module is used to construct paired evidence acquisition targets, transform each paired evidence acquisition target into a focused research question to identify implicit demands in candidate operational factors, and acquire evidence; the acquired evidence is then organized into structured source discovery. The program admission module is used to group programs according to the program decisions supported by structured source discovery, summarize each structured source discovery group to obtain candidate programs, construct an admission record for each candidate program, perform program admission judgment, and obtain the admission result. The skill pack compilation module is used to compile and generate skill packs based on the general instructions and limited scope instances in the admission results, introduce execution and encapsulation constraints, and compile them. The syntax-guided verification module is used to perform syntax-guided verification on the generated skill pack using skill syntax. After verification, it outputs the skill pack that can be installed and loaded on demand by the agent during inference.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned method for generating intelligent agent task skills based on a heterogeneous open world.

[0013] According to some embodiments, the present disclosure adopts the following technical solutions: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned intelligent agent task skill generation method based on a heterogeneous open world.

[0014] According to some embodiments, the present disclosure adopts the following technical solutions: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the method for generating intelligent agent task skills based on a heterogeneous open world.

[0015] Compared with the prior art, the beneficial effects of this disclosure are as follows: This disclosure discloses a method for generating intelligent agent task skills based on heterogeneous open worlds. Taking insufficiently described skill requirements, source access specifications, and execution and encapsulation constraints as input, it obtains evidence from heterogeneous open world sources to discover implicit requirements in the skill requirements that are omitted but will alter the agent's behavior. Based on the applicability supported by the evidence, it determines whether candidate programs can be included in the skill set, and compiles the content that passes the admission judgment into an installable skill package. This disclosure discovers implicit requirements that affect the agent's execution behavior from incomplete skill requirements, obtains relevant materials around different task scenarios, and reduces the limitations of the original skill requirements on the retrieval scope; it judges the reliability and applicability of candidate operation methods based on evidence, avoiding the erroneous application of local experience as general rules; and it organizes the judged operation methods into a structurally complete skill package with clearly defined applicability boundaries that can be loaded and used by the agent.

[0016] The disclosed method for generating intelligent agent task skills based on heterogeneous open worlds can acquire procedural knowledge from open world materials such as technical documents, code repositories, and issue reports, reducing reliance on manually written skills and historical execution trajectories.

[0017] This disclosed method for generating agent task skills based on heterogeneous open worlds improves the reliability of skill content. By discovering implicit requirements, organizing structured evidence, and determining program access, it distinguishes between general instructions and local examples, avoiding the direct inclusion of content with insufficient evidence or conflicting information into skills.

[0018] The disclosed method for generating intelligent agent task skills based on heterogeneous open worlds achieves a comprehensive task pass rate of 55.8% in a multi-domain intelligent agent task test set, which is 19.9 percentage points higher than the no-skill method and 8.6 percentage points higher than the best automatic skill creation comparison method.

[0019] The method for generating intelligent agent task skills based on heterogeneous open worlds disclosed herein demonstrates through source perturbation experiments that it can avoid elevating injected misleading content to runtime instructions; and through requirement change experiments that the skills generated by this method can adapt to a certain range of task changes.

[0020] The disclosed method for generating intelligent agent task skills based on heterogeneous open worlds can automatically generate structurally standardized, installable, and callable skill packages with an average creation time of 23.21 minutes, which is lower than the 35.21 to 36.37 minutes of the comparative methods. Attached Figure Description

[0021] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0022] Figure 1 This is a diagram showing the results of a pilot experiment on the impact of open-world source access on the effectiveness of agent skill creation in this embodiment of the disclosure; in, Figure 1 (a) in the text refers to the Anthropic skill creator. Figure 1 (b) in the text refers to the OpenAI Skill Creator; Figure 2 This is an overall framework diagram of the agent task skill generation method based on heterogeneous open-world evidence in the embodiments of this disclosure; Figure 3 This is a task-level performance comparison chart between the skills generated in this embodiment and manually written skills; Figure 4 The diagram shows the ablation experiment results for implicit requirement discovery, structured source discovery, program admission, and syntax-guided verification in the embodiments of this disclosure. Figure 5 Box plot showing the actual number of lines read from the SKILL.md skill file by different agents and model configurations in the embodiments of this disclosure; in, Figure 5In the diagram, (a) represents the Claude Code + DeepSeek-V4-Pro configuration; (b) represents the ClaudeCode + Opus 4.8 configuration; (c) represents the Codex + DeepSeek-V4-Pro configuration; and (d) represents the Codex + GPT-5.5 configuration. Figure 6 This is a diagram showing the core file length and file composition distribution of the skill package under different agent and model configurations in the embodiments of this disclosure; in, Figure 6 In the table, (a), (b), (c), and (d) correspond to the core file length and file composition distribution of the skill package under different agent and model configurations, respectively. Detailed Implementation

[0023] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Terminology Explanation LLM agents: Intelligent systems built on large language models that can perform tasks, reason, and interact with users.

[0027] Agent skills: Reusable procedural knowledge carriers used to extend the capabilities of LLM agents, including task execution steps, tool usage specifications, applicable conditions, exception handling methods, and result verification methods, which can be loaded by the agent as needed during the inference or task execution phase.

[0028] Open-world materials: Heterogeneous information sources distributed in open environments that are not pre-organized into a complete skills knowledge base, including technical documents, code repositories, issue reports, web page materials, operation guides, and agent experience records, etc.

[0029] Implicit requirements: These are potential requirements that are not directly stated in the skill creation requirements but will affect the skill's applicability, execution process, or processing results.

[0030] Comparative evidence: Information obtained by comparing relevant materials under different task scenarios to determine whether operating conditions will cause changes in the processing method.

[0031] Operating procedure access: The process of determining whether a candidate operating method can be incorporated into an agent's skills based on the sufficiency, consistency, and scope of the evidence.

[0032] Skill package: A documented encapsulation of an agent's skills, including skill descriptions, execution steps, and necessary examples, scripts, templates, or references, which can be loaded and used by the target agent.

[0033] Example 1 One embodiment of this disclosure provides a method for generating intelligent agent task skills based on a heterogeneous open world, the method steps of which include: Step 1: Receive task skill requirements and identify candidate operational factors; Step 2: Construct paired evidence acquisition objectives, transform each paired evidence acquisition objective into a focused research question to identify implicit needs in candidate operational factors, and acquire evidence; organize the acquired evidence into structured source discovery; Step 3: Group the candidates according to the procedural decisions supported by the structured source discovery, summarize the candidate procedures for each structured source discovery group, construct an admission record for each candidate procedure, and perform the admission judgment to obtain the admission result; Step 4: Based on the general instructions and limited scope instances in the admission results, introduce execution and encapsulation constraints, and compile to generate a skill package; Step 5: Use skill syntax to perform grammatical guidance verification on the generated skill pack. After verification, output the skill pack that can be installed and loaded as needed by the agent during the inference process.

[0034] As one embodiment, this disclosure provides a method for generating intelligent agent task skills based on heterogeneous open worlds. The method takes insufficiently described skill requirements, source access specifications, and execution and encapsulation constraints as inputs, obtains evidence from heterogeneous open world sources, discovers implicit requirements that are missing in the skill requirements but will change the behavior of the intelligent agent, determines whether candidate programs can enter the skill based on the scope of applicability supported by the evidence, and compiles the content that passes the admission judgment into an installable skill package.

[0035] This disclosure models open-world skill creation as a procedural access problem based on source evidence. The skill creation process is not a simple compression or summary of fixed source material, nor is it a direct use of retrieved content as skill instructions. Instead, it obtains evidence material that conforms to source access specifications and determines that each candidate procedure should: (1) be included in the skill as a general instruction; (2) be retained as an example or illustration with a limited scope of application; or (3) be excluded due to insufficient evidence or conflict.

[0036] The overall method of this disclosure includes three stages: implicit requirement discovery, evidence-based program admission, and skill pack compilation. Each stage specifically includes four key components: implicit requirement discovery, structured discovery, program admission, and syntax-guided verification. The implicit requirement discovery component and the structured discovery component belong to the first stage, the program admission component belongs to the second stage, and the syntax-guided verification component belongs to the third stage.

[0037] The overall implementation process of this disclosure is in the skills creator. Φ The skill creator is implemented in [the context of the skill creator]. Φ It is a software functional module executed by a processor for implementing the intelligent agent task skill generation method of this disclosure, and is not limited to a specific intelligent agent platform or large language model platform. The skill creator can call the large language model, source retrieval interface and file reading and writing tools to sequentially perform implicit requirement discovery, evidence-based program admission and skill package compilation, and output an installable intelligent agent skill package.

[0038] Skill Creator Φ Receive input triplet ,in, This indicates an inadequate description of skill requirements, namely a brief natural language description of the tasks or abilities that the skill should support; The source access specification defines the permitted source types, retrieval channels, and exclusions. Sources can include technical documents, code repositories, web page materials, existing skills, or agent experience. This indicates the constraints on execution and skill package encapsulation, including the tools that can be used and the file structure requirements that the skill package needs to meet.

[0039] Ultimately, the skill creator Map the above inputs to installable skill packs:

[0040] in, This represents the final skill pack generated; This represents the core file of the skill pack; This indicates optional supporting resources, including references, examples, scripts, templates, or static resources.

[0041] As one example, in this skill creator The specific skill pack generation process is as follows: (a) Implicit needs discovery stage; The first phase of the process, Φ1, is used to discover implicit requirements that are overlooked in the skill requirements but affect procedural behavior, and to form a focused research question and structured source discovery.

[0042] The implicit requirement discovery phase includes implicit requirement discovery and structured requirement discovery. The specific components of the implicit requirement discovery component are as follows: (1) Constructing an operational framework. Skill requirements typically describe a specific task instance, while reusable skills need to be applicable to a set of related task instances. Therefore, the first stage begins with constructing an operational framework based on skill requirements. g Building the operational framework:

[0043] The operational framework identifies specific values ​​in skill requirements as candidate operational factors and analyzes decisions that are not yet clearly defined in the following procedural behaviors: A. How to generate the target output; B. How to handle failure scenarios; C. How to verify the generated results.

[0044] Candidate operating factors merely indicate areas requiring further investigation and do not directly constitute implicit requirements. A factor is only identified as an implicit operating factor when evidence shows that changing it would alter the program's behavior. (2) Obtaining comparative evidence. Directly using the original skill requirements for retrieval easily yields materials related to the same local task instance repeatedly; unconstrained expansion of requirements may introduce conditions unrelated to skill behavior. Therefore, the objective of this disclosure in constructing paired evidence is:

[0045] in, d Indicates a candidate operational factor. x and x′ This indicates consistency in all other aspects, focusing only on candidate operating factors. d Two changing task scenarios; Furthermore, three types of detection methods are used to compare whether two task scenarios require different procedural processing methods, and to determine whether candidate operational factors will change skill behavior. Specifically, these include: 1. Replacement Probe: Change the task subject, processing method, or tools used within the same capability category to test whether the program can be migrated to other task instances.

[0046] 2. Boundary detection: Introduce preconditions, failure scenarios, or operational constraints not specified in the original skill requirements to determine whether these factors will change the program's processing method.

[0047] 3. Proximity capability detection: Compare the target task with adjacent capabilities that have the same output interface to determine whether they require different procedural processing methods.

[0048] Each paired evidence acquisition objective was transformed into a focused research question to determine whether the two task scenarios required different treatment in either of the following procedural components: } in: Indicates applicable conditions; Indicates an operation action; Indicates the abnormal recovery method; Indicates the method for verifying the results.

[0049] Furthermore, regarding the goal of evidence acquisition Corresponding source discovery The set of program components that cause processing differences is defined as follows:

[0050] when satisfy:

[0051] Then the candidate operational factors d The record is an implicit requirement, and also records the procedural components affected by this factor and the applicable boundaries described in the source evidence.

[0052] Furthermore, the specific content of the structured discovery component includes: The first phase involves organizing the acquired evidence into structured source discovery, retaining the corresponding research questions, operational context, processing methods, affected procedural components, applicable boundaries, and source basis.

[0053] Evidence that uses the same processing method in different task scenarios is recorded as cross-scenario consistent evidence; incompatible processing methods from different sources are retained to prevent conflict; and cases with insufficient evidence are left unresolved without making a direct judgment.

[0054] The first phase output focuses on a set of research questions. and structured source discovery set :

[0055] Structured source discovery serves only as an observational finding with source evidence and is not directly used as an executable skill instruction. Whether it can be incorporated into a skill is determined by the second stage.

[0056] (ii) Evidence-based procedural access phase; Second-stage processing It is used to aggregate structured source findings related to the same procedural decision-making, summarize candidate procedures, and make admission decisions based on the scope of applicability supported by evidence.

[0057] The specific content of the admission component includes: (1) Aligning source discovery according to procedural decisions. Since the structured source discovery in the first stage is still related to its respective local source context, this invention first groups the source discovery according to the procedural decisions it supports, rather than according to the source topic or source document. During the grouping process, the original operating context and applicable conditions of each source discovery are preserved:

[0058] in, G j This represents a set of structured source findings related to the same procedural decision point. This is a structured set of source discovery methods; the program decision points may include route selection, accommodation selection, or failure handling, etc. During the grouping process, the original operating context and applicable conditions of each source discovery are preserved.

[0059] (2) Summarize candidate procedures. For each source discovery set G j Candidate programs are derived by induction: ←Induce ( G j ) The candidate program is represented as:

[0060] in: Indicates the applicable conditions for the candidate procedure; Indicates the operation actions of the candidate program; Indicates the exception recovery method of the candidate program; Indicates the method for verifying the results of the candidate program; This represents the mapping relationship between each component of the candidate program and the source discovery that supports these components.

[0061] As one embodiment, the candidate program only contains a source discovery set. Directly supported content. For source terms with the same meaning but different expressions, normalization can be performed; however, without cross-contextual evidence, specific names or fixed choices should not be extended beyond their original task context.

[0062] (3) Construct a range-aware admission record. For each candidate program Build an access record:

[0063] in, Indicates the scope of application Internally, it supports the discovery of the sources of each component of the candidate program; This indicates that, under overlapping operation conditions, it supports the discovery of sources of incompatible processing methods; It indicates the widest applicability that the current evidence can prove.

[0064] (4) Execute the admission judgment procedure. (The following is a list of steps / methods, not a direct translation of the , and Indicate candidate program Does it meet the requirements of support, consistency, and reusability?

[0065] The admission results for the candidate programs are as follows:

[0066] Ultimately, the admission results fall into the following three categories: 1. General Instructions: When a candidate program simultaneously meets the requirements of support, consistency, and reusability, it is admitted as a general instruction and added to the general instruction set. R g .

[0067] 2. Limited-scope instances: When a candidate program meets the support and consistency requirements, but existing evidence is insufficient to prove that it can be extended beyond a single-source local context, it is retained as a limited-scope instance and added to the limited-scope instance set. R e .

[0068] 3. Exclusion: Candidate procedures are excluded when they lack supporting evidence or when there are unresolved conflicts within their applicable scope. Excluded candidate procedures are retained in the audit log but are not passed on as skills content to the third phase.

[0069] Overall, the second-stage processing procedure is represented as follows:

[0070] (III) Skill package compilation stage; Third-stage processing This is used to compile already approved generic instructions and limited-scope examples into installable skill packs. The skill pack compilation phase does not create new candidate programs, nor does it expand the scope of evidence already identified in Phase 2. Given a set of generic instructions... Limited Scope Example Set and execution and encapsulation constraints Generate skill packs:

[0071] As one embodiment, the specific content includes: 1. Write the skill name, skill function description, and the appropriate time to apply the skill into the YAML pre-metadata of the SKILL.md file; 2. Organize the approved procedures, their applicable conditions, and security measures into the main text of SKILL.md. Executable instructions; 3. Organize the program into an executable sequential or conditional processing structure; 4. Store the support content that does not need to be fully loaded into the context when SKILL.md is loaded for the first time as optional. Set of resources X; 5. Store detailed descriptions and examples of defined scopes in the references / directory; 6. Store the executable program in the scripts / directory; 7. Store templates or static resources in the assets / directory; 8. Reference the accompanying resources in SKILL.md using a valid relative path for the skill pack.

[0072] Furthermore, after the skill package is compiled, the syntax-guided verification component checks the skill package using the skill syntax Gskill, which is derived from publicly available qualified skills. Skill syntax is an operational structure specification derived from the skill corpus, not a formal syntax at the character or lexical level. Syntax-guided verification includes checking: 1. Does the skill description clearly distinguish its intended use from adjacent tasks? 2. Whether the admitted procedure is organized as an executable sequence or conditional structure; 3. Whether to retain the applicable conditions and security measures for the procedure; 4. Whether optional supporting resources are referenced via a valid relative path from a skill pack; 5. Are the structure, reference relationships, and applicable boundaries of the skill pack valid?

[0073] Once the check is complete, the output is a skill pack that can be installed and loaded as needed by the agent during inference.

[0074] As one embodiment, this disclosure takes the generation of multi-day travel planning skills as an example. The actual input of this application task includes skill requirements. Source Access Guidelines and execution and encapsulation constraints Among them, skill requirements For: Develop a 7-day itinerary for a trip through Europe, using train travel between cities, and provide clear daily schedules and practical suggestions. (Source: Access Guidelines) Allow evidence from web search results, travel blogs or guides, forums or communities, and reviews such as Yelp, and set manually written skills as a prohibited source. Enforce and encapsulate constraints. The regulations stipulate that relevant information can be used for operational and result verification checks, and require the final skill package to meet the following requirements: As the core document, it also includes Table of Contents and Table of contents.

[0075] The skill creator first addresses the skill requirements. An operational framework is constructed that identifies specific values ​​such as "7 days," "Europe," and "train" as candidate operational factors such as travel duration, destination, and mode of transportation, rather than directly treating these specific values ​​as fixed rules in reusable skills.

[0076] For the aforementioned candidate operational factors, the skill creator constructs paired evidence acquisition objectives. For example, it compares train travel with air travel to determine whether different modes of transportation require different handling; it compares visa-free travel with visa-required travel to determine whether visa conditions change route planning; and it compares leisure travel with business travel to determine whether travel purpose changes itinerary arrangements. Each paired evidence acquisition objective is transformed into a corresponding focused research question to determine whether the candidate operational factors would change the applicability of the skill, the operational actions, the anomaly recovery methods, or the result verification methods.

[0077] Skill creator accesses based on source specifications Acquire evidence and organize it into structured source discovery. In this task scenario, structured source discovery includes: public transportation helps reduce travel costs; rail is suitable for shorter routes; air travel can shorten travel time for long distances; Schengen visa regulations may restrict route choices; rail operators' pet transport regulations may limit mode of transport choices; solo travelers need more itinerary flexibility; family travelers prioritize convenience; and some popular travel itineraries prioritize photo locations. Due to the limitations of manually crafting source discovery skills, the source access specifications are... Therefore, source discovery is not derived from manually written skills.

[0078] The skills creator grouped its sources according to the procedural decisions supported by the structured source discovery. Source discoveries related to solo and family travel were categorized into the Traveler Adaptation group; those related to public transportation, rail, and air travel were categorized into the Mode of Transportation Trade-off group; those related to visa regulations and pet transportation regulations were categorized into the Travel Constraints group; and those related to photo locations were categorized into the Photo Location Preference group.

[0079] Based on the above groupings, the skills creator identified the following candidate procedures: adjusting itinerary pace and activity arrangements according to traveler type; comparing modes of transportation based on cost, time, and applicability; handling travel constraints according to visa and pet transportation regulations; and always prioritizing popular photo locations.

[0080] The skill creator performs admission checks on the aforementioned candidate procedures. Since "adjust itinerary pace and activity arrangements according to traveler type" and "compare modes of transportation based on cost, time, and applicability" have supporting evidence and can be reused in different travel scenarios, they are admitted as general instructions. Because visa regulations and pet transportation regulations only apply to specific travel scenarios that meet the corresponding conditions, "handle travel constraints according to visa regulations and pet transportation regulations" is retained as a limited-scope example. Since existing evidence is insufficient to support consistently prioritizing popular photo locations across different travel tasks, "always prioritize popular photo locations" is excluded and not included in the runtime skill instructions.

[0081] Based on the above admission results, the skill creator generates the name as follows: An installable travel planning skills kit. The skills kit is designed to... As the core file, in The skill package includes the skill name, a description of the skill function, and operational steps for "clarifying needs, comparing solutions, developing a plan, and verifying results." The skill package also includes... Table of Contents and Table of contents, among which, The catalog includes and These are used to store mode of transport, traveler type, travel constraints, and evidence of origin separately. The catalog includes and It is used to store itinerary templates and checklists.

[0082] After the skill pack is compiled, a syntax check is performed on the skill description, executable processing flow, applicable conditions, security measures, and supporting resource references. Once the check is complete, the actual output of the application task is a travel planning skill pack that can be installed and loaded on demand by the agent.

[0083] Example 2 One embodiment of this disclosure provides an intelligent agent task skill generation system based on a heterogeneous open world, including: The initialization module is used to receive task skill requirements and identify candidate operation factors; The implicit demand discovery module is used to construct paired evidence acquisition targets, transform each paired evidence acquisition target into a focused research question to identify implicit demands in candidate operational factors, and acquire evidence; the acquired evidence is then organized into structured source discovery. The program admission module is used to group programs according to the program decisions supported by structured source discovery, summarize each structured source discovery group to obtain candidate programs, construct an admission record for each candidate program, perform program admission judgment, and obtain the admission result. The skill pack compilation module is used to compile and generate skill packs based on the general instructions and limited scope instances in the admission results, introduce execution and encapsulation constraints, and compile them. The syntax-guided verification module is used to perform syntax-guided verification on the generated skill pack using skill syntax. After verification, it outputs the skill pack that can be installed and loaded on demand by the agent during inference.

[0084] Example 3 One embodiment of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for generating intelligent agent task skills based on a heterogeneous open world.

[0085] Example 4 One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the intelligent agent task skill generation method based on a heterogeneous open world.

[0086] Example 5 One embodiment of this disclosure provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the method for generating intelligent agent task skills based on a heterogeneous open world.

[0087] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for generating intelligent agent task skills based on heterogeneous open worlds, characterized in that, include: Receive task skill requirements and identify candidate operational factors; Construct paired evidence acquisition objectives, transform each paired evidence acquisition objective into a focused research question to identify implicit needs in candidate operational factors, and acquire evidence; organize the acquired evidence into structured source discovery; The source discovery groups are grouped according to the procedural decisions supported by the structured source discovery method. Candidate procedures are obtained by summarizing each group. An admission record is constructed for each candidate procedure, and the admission judgment is performed to obtain the admission result. Based on the general instructions and limited-scope instances in the admission results, execution and encapsulation constraints are introduced to compile and generate skill packs. The generated skill pack is grammatically guided and validated using skill grammar. Once validation is complete, the output is a skill pack that can be installed and loaded on demand by the agent during inference.

2. The method for generating intelligent agent task skills based on heterogeneous open worlds as described in claim 1, characterized in that, The receiving task skill requirements, and the identification of candidate operation factors, include: Acquire task skill requirements and construct an operational framework; The operational framework identifies specific values ​​in skill requirements as candidate operational factors and analyzes decisions that are not yet clearly defined in procedural behavior.

3. The method for generating intelligent agent task skills based on heterogeneous open worlds as described in claim 1, characterized in that, The construction of paired evidence acquisition objectives involves transforming each paired evidence acquisition objective into a focused research question that identifies implicit needs among candidate operational factors, and acquiring evidence, including: Constructing paired evidence to obtain the target , d Indicates a candidate operational factor. x and x′ This indicates consistency in all other aspects, focusing only on candidate operating factors. d Two changing task scenarios; Three detection methods are used to determine whether two task scenarios require different procedural processing methods and whether candidate operational factors will change skill behavior. Each paired evidence acquisition objective is converted into a focused research question to determine whether two task scenarios need to take different processing in any procedural component, and the set of procedural components in which processing differences occur is defined. When the set of program components meets the set conditions, the candidate operational factors are recorded as implicit requirements, and the program components affected by the factors and the applicable boundaries described in the source evidence are also recorded.

4. The method for generating intelligent agent task skills based on heterogeneous open worlds as described in claim 1, characterized in that, The process of organizing the acquired evidence into structured source discovery includes: The obtained evidence is organized into structured source findings, retaining the corresponding research questions, operational context, processing methods, affected procedural components, applicable boundaries, and source basis; Evidence that uses the same processing method in different task scenarios is recorded as cross-scenario consistent evidence; incompatible processing methods from different sources are retained to prevent conflict; and cases with insufficient evidence are left unresolved without making a direct judgment.

5. The method for generating intelligent agent task skills based on heterogeneous open worlds as described in claim 1, characterized in that, The process involves grouping applicants according to the procedural decisions supported by structured source discovery, and then summarizing each structured source discovery group to obtain candidate programs. For each candidate program, an admission record is constructed, and program admission judgment is performed to obtain the admission result, including: Grouping is based on the procedural decisions supported by structured source discovery, while preserving the original operating context and applicable conditions of each source discovery during the grouping process: in, G j This represents a set of structured source findings related to the same procedural decision point. A set of structured source discoveries; For each source discovery set G j Candidate programs are derived by induction: ←Induce ( G j ) For each candidate program Build an access record: in, Indicates the scope of application Internally, it supports the discovery of the sources of each component of the candidate program; This indicates that, under overlapping operation conditions, it supports the discovery of sources of incompatible processing methods; Indicates the widest applicability that can be proven by the current evidence; Determine whether the candidate program meets the requirements of support, consistency and reusability to obtain the admission result, which includes general instructions, limited scope instances and excluded programs.

6. The method for generating intelligent agent task skills based on heterogeneous open worlds as described in claim 1, characterized in that, Based on the general instructions and limited scope instances in the admission results, execution and encapsulation constraints are introduced to compile and generate a skill package, including: The skill pack compilation phase does not create new candidate programs, nor does it expand the scope of already established evidence support; it is given a general set of instructions. Limited Scope Example Set and execution and encapsulation constraints Generate skill packs : in, These are optional supporting resources.

7. A system for generating intelligent agent task skills based on heterogeneous open worlds, characterized in that: include: The initialization module is used to receive task skill requirements and identify candidate operation factors; The implicit demand discovery module is used to construct paired evidence acquisition targets, transform each paired evidence acquisition target into a focused research question to identify implicit demands in candidate operational factors, and acquire evidence; the acquired evidence is then organized into structured source discovery. The program admission module is used to group programs according to the program decisions supported by structured source discovery, summarize each structured source discovery group to obtain candidate programs, construct an admission record for each candidate program, perform program admission judgment, and obtain the admission result. The skill pack compilation module is used to compile and generate skill packs based on the general instructions and limited scope instances in the admission results, introduce execution and encapsulation constraints, and compile them. The syntax-guided verification module is used to perform syntax-guided verification on the generated skill pack using skill syntax. After verification, it outputs the skill pack that can be installed and loaded on demand by the agent during inference.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent agent task skill generation method based on any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the intelligent agent task skill generation method based on any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the intelligent agent task skill generation method based on any one of claims 1-6.