Multi-agent collaborative task decomposition and execution system for industry evaluation

Abstract

This invention provides a multi-agent collaborative task decomposition and execution system for industry assessment, comprising: a task planning agent for receiving an industry assessment request, decomposing the request into sub-tasks, and generating a task execution graph using the sub-tasks as nodes; an execution agent group for executing the assigned sub-tasks in parallel or serially according to the sub-tasks in the task execution graph; a review and verification agent for verifying the data accuracy and consistency of the outputs of each execution agent, generating correction instructions for execution agents whose verification results do not meet the requirements, and re-verifying the data accuracy and consistency of the outputs of those execution agents, with the feedback loop iterating continuously until the verification results of each execution agent meet the requirements; and a report generation agent for integrating the outputs of each execution agent to generate a structured industry research report.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, multi-agent collaborative technology and large language model applications, specifically to a multi-agent collaborative task decomposition and execution system for industry evaluation. Background Technology

[0002] In today's rapidly changing economic environment, in-depth, comprehensive, and timely assessments of specific industries are crucial for corporate strategic planning, investment decisions, market analysis, and policy formulation. Traditional in-depth industry research typically relies on costly teams of human experts, involving extensive data collection, information filtering, cross-validation, data analysis, and report writing. This process is not only time-consuming and labor-intensive, but the quality and efficiency of the research results also heavily depend on the experience, expertise, and collaborative abilities of the expert team, presenting the following significant pain points: First, research is costly and inefficient. Human experts need to invest significant time and human resources in in-depth research. The process of sifting through massive amounts of literature, reports, and data to extract and integrate useful information, followed by complex analysis and logical reasoning, is lengthy and costly. This is especially true when dealing with emerging industries or interdisciplinary fields, where the breadth of expertise required makes it difficult for a single team to cover everything, further increasing the difficulty and cost of research.

[0003] Secondly, information fragmentation and data conflicts are difficult to resolve effectively. Industry research often requires obtaining information from multiple heterogeneous sources (such as academic papers, industry reports, news, company financial reports, policies and regulations, market research data, etc.). These sources may contain inconsistent data, conflicting viewpoints, or even direct contradictions. Traditional manual methods require significant effort in cross-validation and expert judgment when dealing with these conflicts, but they still struggle to completely avoid subjective biases or omissions of key information, thus affecting the accuracy and reliability of the report.

[0004] Furthermore, there is a lack of automated verification and feedback correction mechanisms. Existing automated or semi-automated research tools, such as simple search engines or text summarization systems, typically only provide preliminary information integration, lacking the ability to conduct in-depth automated review and verification of information quality, source authority, data accuracy, and logical consistency. When errors or contradictions are found, these systems cannot automatically trigger correction processes and can only rely on manual intervention, which greatly limits their application potential in generating high-quality, highly accurate reports. While large language models perform well in content generation, the potential for "illusion" phenomena during the generation process, as well as limitations in complex reasoning and fact-checking, make it difficult to directly use the output of large language models without rigorous verification in high-risk industry assessment scenarios.

[0005] Furthermore, the report generation process is not standardized, resulting in inconsistent quality. Human-written reports are easily influenced by the writer's style, experience, and area of ​​expertise, leading to variations in report structure, depth, and consistency. Therefore, the lack of standardized report generation processes and automated quality control mechanisms makes it difficult to consistently guarantee the quality of each research outcome.

[0006] Therefore, the current industry assessment field urgently needs a system that can fully utilize the advantages of large language models in information processing and logical reasoning, and deeply integrate automated verification and feedback correction mechanisms by simulating the professional division of labor and collaboration of human teams, so as to achieve end-to-end automation of in-depth research on complex industries, thereby significantly reducing costs, improving efficiency, and ensuring information accuracy and logical consistency. Summary of the Invention

[0007] To address the existing technical problems, this invention provides a highly reliable multi-agent collaborative task decomposition and execution system for industry evaluation, featuring intelligent verification and self-correction.

[0008] To achieve the above objectives, the technical solution of this invention is implemented as follows: A multi-agent collaborative task decomposition and execution system for industry evaluation includes: A task planning agent is used to receive industry evaluation requests, decompose the industry evaluation requests into sub-tasks, and generate a task execution graph with the sub-tasks as nodes. An execution agent group includes multiple execution agents, each of which executes the assigned sub-tasks in parallel or serially according to the sub-tasks contained in the task execution graph. The review and verification agent is used to verify the data accuracy and consistency of the output of each execution agent. Based on the verification results, for execution agents whose verification results do not meet the requirements, a correction instruction is generated and the data accuracy and consistency of the output of the execution agent is verified again. The feedback loop continues to iterate until the verification results of each execution agent meet the requirements. The report generation agent integrates the outputs of each of the aforementioned execution agents to generate a structured industry research report.

[0009] The multi-agent collaborative task decomposition and execution system for industry evaluation provided in the above embodiments includes a task planning agent, an execution agent group, a review and verification agent, and a report generation agent. Based on a large language model, it constructs an agent system that simulates the division of labor and collaboration of a human expert team, realizing full-process automation from macro research objectives to final report generation.

[0010] A task planning agent receives industry evaluation requests, decomposes these requests into subtasks, and generates a task execution graph using these subtasks as nodes. The task planning agent is responsible for intelligently decomposing high-level research objectives into a structured, executable flow of subtasks.

[0011] An executive agent swarm is formed, in which multiple executive agents are responsible for executing these sub-tasks in parallel or serially, including knowledge retrieval and data analysis, which facilitates the expansion of any required agent type according to actual application needs.

[0012] The review and verification agent verifies the accuracy and consistency of the outputs of each execution agent, iterating continuously based on the verification results until all execution agents' verification results meet the requirements. Therefore, the review and verification agent receives and automatically evaluates the outputs of the execution agents in real time, performing rigorous verification based on indicators such as source authority and logical consistency. Upon discovering conflicts or contradictions, it proactively triggers a feedback correction loop to guide the execution agents in making secondary corrections.

[0013] A report-generating agent integrates the outputs of the various executing agents to generate a structured industry research report. Ultimately, the report-generating agent is responsible for integrating all validated knowledge blocks to efficiently and automatically generate a high-quality, highly accurate in-depth industry research report. This closed-loop correction mechanism ensures the accuracy and robustness of the entire research process, significantly improving the efficiency and reliability of industry assessments. Attached Figure Description

[0014] Figure 1 This is a structural diagram of a multi-agent collaborative task decomposition and execution system for industry evaluation in one embodiment. Detailed Implementation

[0015] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the ways in which the invention may be implemented. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0017] In the following description, the expression “some embodiments” is used, which describes a subset of possible embodiments. However, it should be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0018] In response to the challenges of high costs, low efficiency, fragmented information, unresolved data conflicts, and a lack of automated verification and feedback mechanisms in the field of industry evaluation, the inventors of this application propose a multi-agent collaborative task decomposition and execution system for industry evaluation. The core idea of ​​this system is to construct an intelligent agent system that simulates the division of labor and collaboration within a human expert team. Based on a large language model, it automates the entire process from macro-level research objectives to final report generation, achieving end-to-end automation and intelligent verification and self-correction during the process, resulting in a highly reliable industry evaluation system. Based on a large language model, by mimicking the professional division of labor and collaboration processes of a human team, macro-level research objectives are transformed into structured, executable sub-tasks, which are then assigned to different specialized intelligent agents for parallel or sequential execution. A review and verification agent with feedback correction capabilities is introduced. This agent can evaluate the output of the execution agent in real time and automatically trigger a feedback correction loop when it detects data conflicts, suspicious sources, or logical contradictions. It sends correction instructions to the corresponding execution agent for secondary retrieval or re-analysis, thereby ensuring the accuracy and consistency of the intermediate process and the final result. Finally, the report generation agent integrates all the verified knowledge blocks to efficiently and automatically generate high-quality and highly accurate in-depth industry research reports.

[0019] Please see Figure 1 This application provides a multi-agent collaborative task decomposition and execution system for industry assessment, comprising: a task planning agent 11, used to receive an industry assessment request, decompose the request into sub-tasks, and generate a task execution graph using the sub-tasks as nodes; an execution agent group 12, comprising multiple execution agents, each of which executes the assigned sub-tasks in parallel or serially according to the sub-tasks contained in the task execution graph; a review and verification agent 13, used to verify the data accuracy and consistency of the outputs of each execution agent, and based on the verification results, generate correction instructions for execution agents whose verification results do not meet the requirements, and perform data accuracy and consistency verification on the outputs of those execution agents again, with the feedback loop continuously iterating until the verification results of each execution agent meet the requirements; and a report generation agent 14, used to integrate the outputs of each execution agent to generate a structured industry research report.

[0020] Among them, Task Planning Agent 11, acting as the "brain" of the multi-agent collaborative task decomposition and execution system for industry assessment, is responsible for deeply understanding and structurally decomposing the macro-level industry assessment goals input by the user, and generating an executable task execution graph. The implementation process of Task Planning Agent 11 generating a task execution graph based on industry assessment requests mainly includes the following steps: Step 1: Goal Understanding and Requirements Analysis. Task planning agent 11 receives industry evaluation requests from users in natural language and uses a built-in large language model to parse the core objectives, scope, time span, key indicators, and potential constraints. In an optional example, a qualified industry evaluation request meets three pre-defined criteria: complete core elements (including the evaluation object, scope, and core requirements), clear expression, and feasible requirements. If it fails, the system guides the user to supplement the correct information through four rule-based methods: intelligent missing information prompts, interactive guidance, template recommendations, and invalid request interception. Specifically, the parsing of the five elements—core objectives, scope, time span, key indicators, and potential constraints—must be based on the explicit information in the request and meet corresponding qualification criteria such as unambiguity, clear boundaries, and specific execution. Only if all five elements are parsed successfully is the requirement parsing considered successful; if any one fails, the parsing is deemed a failure.

[0021] Step 2, Task Decomposition Algorithm. Task planning agent 11, based on industry assessment process ontology knowledge and a large language model, recursively decomposes industry assessment requests into independently executable subtasks. The decomposition process can be formalized as a tree structure or a directed acyclic graph. The decomposition follows the corresponding generation rules for retrieval, analysis / modeling, and professional interpretation subtasks, corresponding to knowledge retrieval, data analysis, and scalable professional execution agents, respectively. After decomposition, subtasks must meet the conditions for successful decomposition, such as being independently executable and having clearly defined execution agents for each subtask. If decomposition fails, one of the following methods is executed: re-decompose the industry assessment request, simplify the industry assessment request using a dimensionality reduction strategy, or output requirement feedback for the current industry assessment request. All industry assessment requests must parse out core data retrieval subtasks; when specific requirements are included, corresponding exclusive subtasks are parsed out. The decomposition process can be formalized as a tree structure or a directed acyclic graph, as follows: Step 3: Task execution graph generation. The decomposed subtasks are constructed into a directed acyclic graph. .in: It is a set of subtask nodes, each node representing a minimum execution unit.

[0022] It is a set of directed edges, representing the dependencies between tasks.

[0023] It is the recommended type of execution agent assigned to each subtask (such as knowledge retrieval agent, data analysis agent).

[0024] It specifies the data type, expected input, and output format required for each subtask.

[0025] The Task Execution Graph (TEG) not only defines the order of task execution but also clarifies the collaborative relationships and data flow between agents.

[0026] Step 4, Resource and Priority Allocation. Task planning agent 11, based on the task dependencies and importance in the Task Execution Graph (TEG), allocates priorities to each subtask and estimates the required computing resources and execution time, providing a basis for subsequent execution agent scheduling.

[0027] Step 5: Output a structured task execution graph (TEG) with dependencies and agent allocation suggestions. Successful generation requires meeting five rules simultaneously: complete and non-redundant subtasks, independently executable subtasks, no circular dependencies, matching agent allocations, and structure conforming to specifications. If any rule is not met, generation is considered a failure, and the task decomposition algorithm needs to be retried to adjust subtasks, dependencies, or agent allocations until all rules are met.

[0028] The execution agent group 12 comprises multiple dedicated execution agents responsible for executing subtasks assigned to them in parallel or serially, based on the TEG generated by the task planning agent. The execution agent group 12 may include the following execution agents: First, the knowledge retrieval intelligent agent.

[0029] This knowledge retrieval agent focuses on retrieving information from various internal and external knowledge sources. Toolset: It possesses capabilities such as API calls, web crawling, database querying, and accessing internal knowledge bases. Functions: It receives sub-tasks (e.g., "retrieve the financial reports of XX company for the past five years," "collect policies and regulations for XX industry"), executes searches, filters relevant information, and extracts key data or text paragraphs. Output: Raw or preliminarily organized "knowledge blocks" (including source links, timestamps, and extracted raw content).

[0030] Second, data analysis intelligent agents.

[0031] This data analytics agent focuses on processing, analyzing, and modeling numerical and structured data. Its toolset includes the ability to access data processing libraries, statistical analysis tools, machine learning models, and visualization tools. Its functions include receiving sub-tasks (such as "calculate the compound annual growth rate of XX market," "predict trends in XX data," and "generate XX charts") and performing data cleaning, statistical analysis, modeling, and visualization. Output includes structured data, analysis results (such as predicted values ​​and statistical indicators), charts or visualization files, and references to source data and generated metadata.

[0032] Third, other scalable intelligent agents.

[0033] Other scalable intelligent agents, such as domain-specific specialized intelligent agents (e.g., legal and regulatory interpretation agents, supply chain analysis agents).

[0034] In the swarm of 12 execution agents, each agent executes and is scheduled in parallel / serial fashion. The agents are scheduled by a central scheduler based on dependencies in the Task Execution Graph (TEG). Independent subtasks can execute in parallel, while dependent subtasks are executed sequentially. The scheduler monitors the execution status and resource consumption of each agent. The output of each agent is a set of "knowledge blocks" generated for each subtask. Each It includes its content, source, agent ID, timestamp, and initial confidence estimate.

[0035] The review and verification agent 13 simulates the critical review process of human experts on research results. It is responsible for the automated, multi-dimensional, and rigorous evaluation of the knowledge blocks output by the executing agents, and triggers a feedback correction loop when problems are found. The review and verification agent 13 verifies the data accuracy and consistency of the outputs of each executing agent. Based on the verification results, for executing agents whose verification results do not meet the requirements, a correction instruction is generated, and the data accuracy and consistency of the outputs of those executing agents are verified again. The feedback loop iterates continuously until the verification results of all executing agents meet the requirements. The implementation process mainly includes the following steps: Step 1: Knowledge block reception and evaluation.

[0036] Real-time reception of knowledge blocks from each executing agent Automated evaluation is performed on each knowledge block or a group of related knowledge blocks, with evaluation dimensions including: Step two: Assess the authority of the information source.

[0037] Scoring is based on the type of knowledge source (such as government reports, academic journals, news media, personal blogs) and external knowledge from pre-defined authoritative databases or large language models. For example, reports from official statistical agencies score higher than unofficial analyses.

[0038] in, The ScoreMapper is a source authority score, representing the degree of trustworthiness of a knowledge source. Source refers to the origin of the knowledge block, such as a report, a news article, or a database. ScoreMapper is a mapping function used to calculate the authority score based on the source's type and reputation. SourceType specifies the type of knowledge source (e.g., government report, academic journal, news media, official company financial report, personal blog, etc.). Reputation represents the reputation or credibility of the knowledge source, which can be evaluated based on a pre-defined authoritative database, historical data, or external knowledge from a large language model.

[0039] Step 3: Data accuracy and consistency verification.

[0040] The data accuracy and consistency verification step aims to ensure that all data output by the executing agents is numerically accurate and informationally consistent, and to minimize the "illusions" or misinformation that large language models may introduce. In a specific example, data accuracy and consistency verification is implemented through the following multi-layered verification mechanism.

[0041] Cross-validation: The system automatically identifies and extracts key data points (such as company revenue, market share, technical parameters, policy release dates, etc.) from different sources that refer to the same entity or event for cross-validation. By comparing these similar data, it detects whether there are significant differences in their numerical values, timeframes, ranges, or qualitative descriptions, thereby identifying data conflicts (e.g., inconsistent company revenue figures in different news reports or industry reports, or obvious contradictions in the descriptions of the same event). When a difference is detected, a subsequent conflict measurement function is used. Conduct a quantitative assessment.

[0042] Numerical range validation: This function performs a preliminary check on the reasonableness of numerical data. The system uses preset industry benchmarks, historical data trends, statistical principles, or domain knowledge to conduct a preliminary reasonableness check on all numerical data. Numerical range validation can serve as a first-level filtering mechanism. For example, it checks whether market growth rates exceed common reasonable ranges (e.g., growth rates wouldn't be 1000%), whether demographic data conforms to common sense, and whether certain ratios in financial statements are abnormal. Any data exceeding a reasonable range will be marked as a potential error or outlier, requiring further verification.

[0043] Fact Checking and Conflict Quantification using the Large Language Model: The system verifies key factual statements output by the executing agent by accessing external knowledge bases. This step leverages the semantic understanding and reasoning capabilities of the large language model, combined with access to external knowledge bases (such as authoritative databases, encyclopedias, fact-checking websites, and official statistical data sources), to verify key factual statements output by the executing agent. Specifically, when potential data discrepancies exist between different knowledge blocks or between a knowledge block and external reference information, the system invokes a conflict measurement function. This function objectively quantifies the degree of conflict between different data points by calculating the normalized absolute or relative difference of numerical differences (for quantitative data), or semantic similarity distance (for qualitative descriptions or textual information). For example, two descriptions of the same event, even if worded differently, but highly similar semantically, have a low conflict metric; conversely, significant numerical discrepancies result in a high conflict metric. Ultimately, based on... The results were used to calculate the data accuracy and consistency scores. This score directly reflects the degree of consistency between the data. The closer the score is to 1, the more accurate and consistent the data is, and the closer it is to 0, the greater the conflict.

[0044] in, The score represents the accuracy and consistency of the data, reflecting the degree of agreement between the data. This refers to the knowledge block currently being validated (e.g., the output of an executing agent). It refers to other related knowledge blocks or reference data, which may come from different sources or different executive agents reporting the same fact. A conflict metric function is used to quantify the degree of data conflict between two knowledge blocks. This is a function of the degree of data conflict between data points. For numerical data, it can be normalized absolute or relative difference; for textual information, it can be semantic similarity distance. The closer the value is to 1, the greater the conflict; the closer it is to 0, the more consistent the data.

[0045] Step 4: Logical consistency check.

[0046] Causal chain analysis: Examines whether the causal relationships inferred by the agent are reasonable.

[0047] Context consistency: Check whether the content of the knowledge block is logically consistent with the current task objective or the results of previous tasks in the TEG.

[0048] Large Model Inference Verification: The logical rigor of the agent's inference process is independently verified using another large model, as follows: in, It is a logical consistency score, which assesses the logical rigor of the knowledge block content and its reasoning process. This is the knowledge block currently being validated. It is a specific large language model instance used for logical verification, designed to evaluate the logical relationships between reasoning chains and contexts. This refers to the reasoning chain that may be contained within a knowledge block, or the reasoning process provided by the executing agent that generated the knowledge block. Context: This refers to the overall context information of the current task, including the macro-level objectives of the industry assessment, the results of the previous tasks that have passed verification in the Task Execution Graph (TEG), and general knowledge in the relevant domain.

[0049] Step 5: Integrity and relevance verification.

[0050] The completeness and relevance checks primarily assess the precision with which a knowledge block covers the corresponding subtask. This involves verifying whether the knowledge block fully covers the information required by the subtask and is highly relevant to it. Specifically, the subtask requirements can be broken down into an information list, and each item in the list can be checked against the corresponding content in the knowledge block. If any item is missing, it indicates incompleteness, and the completeness fails. Furthermore, the knowledge block can be checked for irrelevant points. If concepts unrelated to the information list are included, it indicates low relevance, and the relevance fails.

[0051] Step Six: Comprehensive Confidence Assessment.

[0052] Generate a comprehensive confidence score for each knowledge block or subtask output: in These are the corresponding weighting coefficients used to balance the importance of different verification dimensions in the overall confidence score. These weights can be configured according to the application scenario and the degree of focus on different quality dimensions. It is the score for the authority of the information source. Data accuracy and consistency score. Logical consistency score. This is a score for completeness and relevance. It assesses whether the knowledge block fully covers the information required by the subtask and how relevant its content is to the current task objective. Additionally, the sum of all weight coefficients is 1.

[0053] Step 7: Feedback Correction Loop.

[0054] Triggering mechanism: When the overall confidence level of any knowledge block is lower than the preset threshold, or when a clear data conflict, logical contradiction or high-risk information is detected, the review and verification agent will trigger a feedback correction loop.

[0055] Correction Instruction Generation: The review and verification agent utilizes its large language model capabilities to generate precise and specific correction instructions for detected issues. For example, "Company A's revenue data comes from an unofficial blog; please retrieve it from the company's official website or SEC filings," or "The prediction model parameters are set incorrectly, resulting in an excessively high growth rate; please adjust the parameters and rerun the analysis."

[0056] Sending instructions and secondary execution: These corrective instructions are sent back to the corresponding execution agent (KRA or DAA). After receiving the instructions, the execution agent will prioritize secondary retrieval, re-analysis, or parameter adjustment to generate new knowledge blocks, and then resubmit them to the review and verification agent for evaluation.

[0057] Iteration and Convergence: This feedback loop iterates continuously until the confidence level of all knowledge blocks reaches a threshold, the maximum number of iterations is reached, or manual intervention is required. This ensures the accuracy and consistency of intermediate knowledge blocks.

[0058] Among them, report generation agent 14 is responsible for integrating, organizing, and formatting the knowledge blocks output by all executing agents that have passed review and verification, to generate a high-quality in-depth industry research report. The implementation scheme for generating a high-quality in-depth industry research report mainly includes the following: (1) Knowledge integration and structuring. Receive all knowledge blocks that have passed the review and verification of the intelligent agent. Based on the TEG generated by the planning agent and the preset report templates (such as industry analysis reports and market research reports), integrate the knowledge blocks into the corresponding chapters and paragraphs of the report.

[0059] (2) Large Model Report Writing. Utilizing its built-in large language model's powerful text generation capabilities, structured knowledge blocks are transformed into fluent, professional, and insightful report texts. This includes: Content Summary and Conclusion: Extracting and summarizing the knowledge from each part. Logical Connection and Narrative Flow: Ensuring clear logic between each chapter of the report, forming a coherent narrative. Chart Embedding and Explanation: Embedding charts generated by the data analysis agent into the report and providing professional explanations and analysis. Insight Extraction and Refinement: Based on the integrated information, extracting key viewpoints, trend insights, and recommendations.

[0060] (3) Report formatting and output. Based on user needs, the industry research report will be output in standard formats (such as PDF, Word, Markdown), supporting custom styles, layout and citation formats, and outputting high-quality, highly accurate and logically rigorous in-depth industry research reports.

[0061] To gain a more comprehensive understanding of the multi-agent collaborative task decomposition and execution system for industry assessment provided in this application embodiment, the following example uses an in-depth analysis of the global electric vehicle battery market and an assessment of the application potential of emerging AI technologies in the agricultural field to illustrate the multi-agent collaborative task decomposition and execution system for industry assessment.

[0062] Example 1: In-depth analysis of the global electric vehicle battery market Suppose a market research firm needs to conduct an in-depth analysis of the global electric vehicle (EV) battery market over the next ten years, including technology trends, supply chain challenges, major players, and market size forecasts.

[0063] The task planning agent receives an industry assessment request for "In-depth Analysis of the Global Electric Vehicle Battery Market," decomposes the request into subtasks, and generates a task execution graph using these subtasks as nodes. The main steps are as follows: 1) User Input: Analyze the global electric vehicle battery market for the next ten years, including technology, supply chain, major manufacturers, and market size forecasts. 2) Planning the intelligent agent: Decompose this macro-objective into: Subtask 1.1: Search for the current development status and roadmap of mainstream EV battery technologies (LFP, NCM, Solid-state).

[0064] Subtask 1.2: Collect information on global reserves, production, and price trends of key raw materials such as lithium, cobalt, and nickel.

[0065] Subtask 1.3: Identify the major global EV battery manufacturers and their market share.

[0066] Subtask 1.4: Analyze the EV policies and regulations of various countries and their impact on battery technology and supply chain.

[0067] Subtask 1.5: Forecast future EV sales and battery demand.

[0068] Subtask 1.6: Establish a forecasting model for the EV battery market size.

[0069] 3) TEG generation: Construct a task execution graph (TEG) that includes the above sub-tasks and their interdependencies (e.g., sub-task 1.5 depends on the policy analysis results of 1.4, and 1.6 depends on the prediction results of 1.5).

[0070] An execution agent swarm comprises multiple execution agents, each of which executes a correspondingly assigned subtask in parallel or sequentially, based on the subtasks contained in the task execution graph. The main steps are as follows: 1) Subtasks are executed in parallel or serially.

[0071] The knowledge retrieval agent is responsible for performing subtasks 1.1 (retrieving papers from IEEE and Elsevier databases), 1.2 (retrieving papers from USGS and Bloomberg Commodity Data), 1.3 (retrieving reports from SNE Research and Benchmark Mineral Intelligence), and 1.4 (retrieving reports from government websites and IEA reports). It outputs the various knowledge blocks.

[0072] Data analytics agent: Responsible for executing sub-tasks 1.5 (EV sales forecasting based on historical sales and policy analysis results) and 1.6 (building a market size model based on raw material prices, technology routes, and demand forecasts). Outputs forecast data and charts.

[0073] The review and verification agent is used to verify the data accuracy and consistency of the outputs of each execution agent. Based on the verification results, for execution agents whose verification results do not meet the requirements, a correction instruction is generated, and the data accuracy and consistency of the outputs of those execution agents are verified again. This feedback loop continues to iterate until the verification results of all execution agents meet the requirements. The main steps are as follows: 1) Knowledge Block Reception: The review and verification agent receives knowledge blocks from the execution agent.

[0074] 2) Confidence assessment: For example, when assessing the "Global Lithium Resource Reserves" knowledge block in subtask 1.2: 3) Data Conflict Detection: KRA may retrieve different data on lithium resource reserves from two different sources (e.g., an investment bank report and an academic study). The review and verification agent detects numerical differences exceeding a preset threshold.

[0075] 4) Source authority assessment: The agent assessment found that one of the sources was a niche investment blog, which had a low authority score (S_{auth}).

[0076] 5) Trigger feedback loop: \(C_{output}\) If the output is below the threshold, the review and verification agent immediately generates a correction instruction: "Please ask KRA to re-retrieve global lithium resource reserves data, preferably from the USGS (United States Geological Survey) or similar official geological survey agencies, and provide at least two authoritative sources for cross-validation." 6) Secondary execution: KRA receives the instruction, re-executes the retrieval task, and submits a new knowledge block from a more authoritative source.

[0077] 7) Logical Consistency Verification: In the output of Subtask 1.6, "Market Size Forecasting Model," DAA may have predicted that a certain technology (such as solid-state batteries) will gain a significant market share in 2028. The review and verification agent found a logical contradiction between this prediction and the conclusion in the "Solid-State Battery Technology Roadmap" knowledge block of Subtask 1.1, which states that "mass production is expected after 2030." The review agent generated the instruction: "Please have DAA re-examine the solid-state battery market penetration forecast and adjust the model parameters based on the technology maturity report." The report generation agent integrates the outputs of each of the aforementioned execution agents to generate a structured industry research report. The main steps are as follows: 1) Once all knowledge blocks have passed verification, the report generates an agent that integrates all the information.

[0078] 2) Report Generation: Automatically generate reports, including executive summaries, market size forecasts (with charts and graphs), technology trend analysis, supply chain risk assessment, and competitive landscape of major manufacturers. For example, the report will clearly state that "global lithium reserves, according to the latest USGS report, are XX tons, and have been cross-validated by multiple sources," rather than simply citing a potentially controversial figure.

[0079] 3) Output: Generate a high-quality, in-depth analysis report on the global electric vehicle battery market.

[0080] Example 2: Assessment of the application potential of emerging AI technologies in agriculture Suppose a high-tech investment company wants to assess the application potential of an emerging AI technology (e.g., a vision-based intelligent identification and precision application system for crop diseases and pests) in the agricultural field and identify potential investment opportunities.

[0081] The task planning agent receives an industry assessment request for "In-depth Analysis of the Global Electric Vehicle Battery Market," decomposes the request into subtasks, and generates a task execution graph using these subtasks as nodes. The main steps are as follows: 1) User input: "Evaluate the application potential of vision-based intelligent identification and precision application systems for crop diseases and pests in the agricultural field, identify investment opportunities, and analyze market entry barriers." 2) Planning the intelligent agent: Decompose this macro-objective into: Subtask 2.1: Search for existing pest and disease identification technologies and market pain points.

[0082] Subtask 2.2: Evaluate the core innovations and technological maturity of this AI technology.

[0083] Subtask 2.3: Analyze the demand for precision spraying systems in major global agricultural markets.

[0084] Subtask 2.4: Identify the main competitors, startups, and funding in this field.

[0085] Subtask 2.5: Assess the potential impact of policies and regulations on the application of smart agriculture technologies.

[0086] Subtask 2.6: Analyze the business model and profit potential of this technology.

[0087] 3) TEG generation: Construct the task execution graph.

[0088] An execution agent swarm comprises multiple execution agents, each of which executes a correspondingly assigned subtask in parallel or sequentially, based on the subtasks contained in the task execution graph. The main steps are as follows: 1) Parallel / serial execution of subtasks Knowledge retrieval agent: Retrieves agricultural science and technology journals, patent databases, agricultural market reports, and venture capital press releases.

[0089] Data analytics agent: Analyzes patent trends and financing data to assess market size potential.

[0090] The review and verification agent is used to verify the data accuracy and consistency of the outputs of each execution agent. Based on the verification results, for execution agents whose verification results do not meet the requirements, a correction instruction is generated, and the data accuracy and consistency of the outputs of those execution agents are verified again. This feedback loop continues to iterate until the verification results of all execution agents meet the requirements. The main steps are as follows: 1) Knowledge block reception: Receives the output of the executing agent.

[0091] 2) Confidence assessment and feedback: When KRA retrieved "the core innovations of this AI technology," it may have extracted key information from a company's press conference presentation. The review and verification agent detected that this information lacked independent third-party verification, and the large language model gave it a low S_{logic} score for objectivity evaluation.

[0092] 3) Review and Verification of Intelligent Agent Generation Instructions: "Please have KRA re-examine the core innovations of this AI technology, prioritizing verification from peer-reviewed academic papers or independent third-party technical evaluation reports, and provide a detailed explanation of its differences from existing technologies." 4) KRA receives the instruction, performs a secondary search, and submits a more authoritative technical evaluation report.

[0093] 5) Data Conflict Resolution: When analyzing funding information, DAA may obtain different funding amounts for a startup from two different VC databases. The review agent detects the conflict and requires DAA to prioritize verifying the data with the company's official announcements or credible news sources.

[0094] 6) Report generation: The report-generating agent integrates all validated knowledge blocks and automatically generates in-depth assessment reports that include technical analysis, market size forecasts, competitive landscape, SWOT analysis, and investment recommendations.

[0095] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a computer, server, network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-agent collaborative task decomposition and execution system for industry evaluation, characterized in that, include: A task planning agent is used to receive industry evaluation requests, decompose the industry evaluation requests into sub-tasks, and generate a task execution graph with the sub-tasks as nodes. An execution agent group includes multiple execution agents, each of which executes the assigned sub-tasks in parallel or serially according to the sub-tasks contained in the task execution graph. The review and verification agent is used to verify the data accuracy and consistency of the output of each execution agent. Based on the verification results, for execution agents whose verification results do not meet the requirements, a correction instruction is generated and the data accuracy and consistency of the output of the execution agent is verified again. The feedback loop continues to iterate until the verification results of each execution agent meet the requirements. The report generation agent integrates the outputs of each of the aforementioned execution agents to generate a structured industry research report.

2. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 1, characterized in that, The review and verification intelligent agent is specifically used to verify the accuracy and consistency of data through a multi-level verification mechanism; The multi-layered verification mechanism includes: identifying and extracting key data points from different sources that refer to the same entity or event for cross-validation; conducting preliminary rationality checks on numerical data; and accessing external knowledge bases to verify key factual statements output by the executing agent.

3. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 1, characterized in that, The review and verification agent is also used to perform logical consistency verification on the output of each execution agent, and to perform integrity and relevance verification on the output of each execution agent. The logical consistency check formula is as follows: It is the logical consistency score. For the knowledge block currently being validated, For large language model instances used for logic verification, It is the reasoning chain contained in the knowledge block, and Context is the context information of the current task; The integrity and relevance checks are performed to verify the accuracy with which the knowledge block covers the corresponding subtask.

4. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 1, characterized in that, Before decomposing the industry assessment request, the task planning agent also parses the industry assessment request. If the parsing is unsatisfactory, it outputs supplementary prompts through set rule-based methods. The established rule-based methods include: intelligent missing item prompts, interactive guidance, template recommendations, and invalid request interception.

5. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 4, characterized in that, If the parsing result of the industry assessment request does not meet the condition of including the five items of core objectives, scope, time span, key indicators, and potential constraints, then the parsing is unqualified.

6. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 4, characterized in that, The task planning agent recursively decomposes the industry evaluation request into subtasks. If the decomposed subtasks are independently executable and the execution agent corresponding to each subtask is determined, then the decomposition is successful. Conversely, the decomposition will fail.

7. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 6, characterized in that, If the decomposition fails, one of the following actions will be taken: re-decompose the industry assessment request, simplify the industry assessment request by using a dimensionality reduction strategy, or output a requirement feedback for the current industry assessment request.

8. The multi-agent collaborative task decomposition and execution system for industry assessment as described in claim 6, characterized in that, After generating the task execution graph, the task planning agent determines whether the following rules are met: subtasks are complete and without redundancy, subtasks can be executed independently, dependencies are not cyclical, agent allocation is matched, and the structure conforms to the specifications. If any rule is not met, the generation is determined to have failed, and the process is retried to receive the industry evaluation request, decompose the industry evaluation request to obtain subtasks, and generate the task execution graph with the subtasks as nodes.

Images