Large language model multi-agent cooperative work method and system

Abstract

The invention relates to the technical field of large language models, in particular to a large language model multi-agent cooperative working method, which comprises the following steps: S1, designing a plurality of agents to simulate social division and cooperation, so that each agent executes different professional abilities, the plurality of agents including a planning agent, a tool agent and a reflection agent; s2, when task input is obtained, the planning agent filters retrieved tools, and a complex task is decomposed into subtasks easy to understand in a code and annotation mode; s3, performing targeted tool enhancement on the filtered tool set by the tool agent, and executing specific operation planned by the planning agent to obtain a subtask result; s4, evaluating a subtask result by the reflection agent, and planning the task from different angles by combining static planning and dynamic planning; and S5, the planning agent feeds back a summary result according to execution of the subtasks.

Description

Technical Field

[0001] The present invention relates to the technical field of large language models, and in particular to a large language model multi-agent collaborative working method and system. Background Art

[0002] Existing research on LLM-driven agent planning strategies has achieved significant success and can be categorized into two types: static planning and dynamic planning. Static planning methods, such as thought chains and thought trees, enable LLMs to decompose reasoning tasks into multiple intermediate steps or subproblems. In dynamic planning methods, each plan is derived through reasoning based on action and feedback, resulting in strong single-step invocation and planning capabilities. In 2022, the Princeton University and Google Brain team proposed "ReAct: Synergizing Reasoning and Acting in Language Models." This paper uses LLMs to generate reasoning trajectories and task-specific actions in an interleaved manner, achieving greater synergy between the two. By accessing external resource APIs to acquire additional business knowledge and generating human-like task-solving trajectories, it overcomes the common hallucination and error propagation issues in thought chain reasoning. While it possesses tool invocation and execution capabilities, because it plans the next step during execution, when handling multi-step reasoning tasks, the model may forget or confuse the current task and fail to integrate historical computational results to output the final answer.

[0003] In recent years, some research has explored solving complex problems by enabling a single agent to play multiple roles within a task. Driven by a single LLM, a single agent independently performs all reasoning, planning, and tool execution. In this model, there is no feedback mechanism from other agents, a limitation that hinders their potential to learn and improve their performance through feedback from other agents. It is worth noting that humans are often limited by their own abilities and knowledge when handling complex tasks, and LLM-based agents with human-like thinking also face the same problem. A refined social division of labor can improve the accuracy and quality of task processing. Solving complex problems often requires cooperation and negotiation among team members. However, individual agents lack the ability to collaboratively execute tasks, and frequent switching of task roles can reduce efficiency. Therefore, a multi-agent architecture comprises two or more agents, each with distinct roles and responsibilities. Through communication and negotiation mechanisms among multiple agents, the overall performance of the solution task can be effectively improved. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for multi-agent collaborative work of a large language model. When processing complex reasoning tasks, it is solved by calling external resource APIs, thereby improving the accuracy and quality of task execution and solving the problems raised in the above background technology.

[0005] To achieve the above object, one aspect of the present invention provides the following technical solutions: A multi-agent collaborative working method for a large language model includes the following steps: S1: Design multiple intelligent agents to simulate social division of labor and collaborative cooperation, so that each intelligent agent can perform different professional capabilities. Multiple intelligent agents include planning agents, tool agents, and reflective agents. S2: When task input is obtained, the planning agent filters the retrieved tools and decomposes the complex task into easy-to-understand subtasks using code and comments; S3: The tool agent performs targeted tool enhancement for the filtered toolset and executes the specific operations planned by the planning agent to obtain subtask results; S4: The reflective agent evaluates the subtask results and combines static and dynamic planning to plan tasks from different perspectives. S5: The planning agent summarizes the results based on the execution feedback of the subtasks.

[0006] Preferably, filtering the retrieved tools in step S2 specifically includes: Given a toolset , contains all tools and a brief description of each tool, related parameters and functions; Before task planning, the planning agent will Perform tool filtering to select a subset of tools that are highly relevant to the task , RTS is expressed as: ; in, represents the screening function of the planning agent, Indicates the tip of the filtering tool, and I indicates complex task information.

[0007] Preferably, in step S2, decomposing the complex task into easily understandable subtasks by means of code and comments specifically includes: The planning agent analyzes the given complex task Q and performs task planning to obtain a set of logically related subtasks S. The subtask set S is expressed as: ; ; in, Indicates task planning prompts, x indicates the number of subtasks, Represents the decomposed subtasks, Represents the subtask goal, Indicates the tool that needs to be called. Indicates parameter configuration, Represents the result of a subtask.

[0008] Preferably, the tool agent in step S3 performs targeted tool enhancement on the filtered toolset, specifically including: The tool agent passes the subtask goal Required tools and enhanced tool target descriptions To determine the parameter configuration of the calling tool , parameter configuration Expressed as: ; in, represents the enhancement function of the tool agent, Represents a prompt for a generated tool parameter.

[0009] Preferably, the step S3 of executing the specific operation planned by the planning agent to obtain the feedback result specifically includes: The tool agent uses the results obtained from using external resources to Perform task summary and obtain subtask results , subtask results Expressed as: ; in, A prompt indicating a task summary. Indicates parameter configuration Use the configured interpreter to call the required tool .

[0010] Preferably, the reflective agent evaluation subtask results in step S4 specifically include: After the tool agent completes the subtask Afterwards, the reflective agent describes and subtask goals Determine whether the tool selected for the current subtask is correct. If not, generate a mark <re-select>, indicating that the tool to be called for this subtask needs to be reselected. The tool agent needs to reselect the tool based on the feedback and subtask goal. Reselect the tool and execute; Reflective agent determines subtask outcomes Whether to include exception error information. If it does, generate a tag. <re-call>, indicating that the tool call failed and the parameter configuration of the calling tool needs to be reconsidered , the tool agent needs to analyze and select parameter configuration again based on the feedback implement; The reflective agent further determines Whether If not, generate a mark <re-sum>, indicating that the tool agent needs to perform the subtask again and summarize based on the feedback.

[0011] Preferably, the step S4 of combining static planning with dynamic planning to plan tasks from different perspectives specifically includes: Reflective Agent Combined with Completed Subtasks information to plan the next subtask , specifying subtasks with comments and code the goals and actions to be performed; Subtask The subtask corresponding to the subtask set S If the two mission objectives are compared or calling tools If not, a mark is generated <re-plan>, indicating that the subsequent tasks should be re-planned, and the reflection agent summarizes the information and feeds it back to the planning agent.

[0012] Preferably, the planning agent in step S5 summarizes the results based on the execution feedback of the subtasks, specifically including: After the tool agent completes all subtasks When , and after the reflection agent gives the feedback that there is no error, the planning agent calculates the execution results R of all subtasks , summarize the summary result of the complex task Q, and the summary result is expressed as: ; in, , A prompt indicating the summary results.

[0013] Preferably, step S5 further includes: Visualize the summary results.

[0014] Another aspect of the present invention provides a large language model multi-agent collaborative working system, the large language model multi-agent collaborative working system comprising: A planning agent, which is used to control the overall direction of the task, filter the retrieved tools, and decompose complex tasks into easy-to-understand subtasks using code and comments; A tool agent, which is used to perform targeted tool enhancement on the filtered tool set and execute the specific operations planned by the planning agent to obtain subtask results; A reflective agent is used to evaluate subtask results and plan tasks from different perspectives by combining static planning with dynamic planning.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a large language model multi-agent collaborative working method, which improves the ability of each agent to handle tasks by simulating social division of labor and cooperation. The planning agent will filter the retrieved tools and decompose complex tasks into easy-to-understand subtasks by means of code and annotations. The tool agent is used to perform targeted tool enhancement on the filtered tool set and execute the specific operations planned by the planning agent to obtain subtask results. The reflection agent evaluates the subtask results and task planning, and combines static planning with dynamic planning to plan tasks from different perspectives. Through the multi-agent task allocation mechanism, complex tasks are decomposed into independent subtasks. Each agent performs the corresponding subtask based on its domain expertise, which significantly improves the LLM's ability to solve complex reasoning tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the process of a large-scale language model multi-agent collaborative working method provided by an embodiment of the present invention; Figure 2 A schematic diagram of the framework of a large-scale language model multi-agent collaborative working method provided by an embodiment of the present invention; Figure 3 A schematic diagram of a planning agent decomposing a complex task into easily understandable subtasks in a large-scale language model multi-agent collaborative working method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] Figure 1 A flowchart of a multi-agent collaborative working method for a large-scale language model provided by an embodiment of the present invention is provided. Figure 2 The framework diagram of the large-scale language model multi-agent collaborative working method provided by the embodiment of the present invention. The embodiment of the present invention provides a large-scale language model multi-agent collaborative working method, such as Figure 1 and Figure 2 As shown, the following steps are included: S1: Design multiple intelligent agents to simulate social division of labor and collaborative cooperation, so that each intelligent agent can perform different professional capabilities. Multiple intelligent agents include planning agents, tool agents, and reflective agents. S2: When task input is obtained, the planning agent filters the retrieved tools and decomposes the complex task into easy-to-understand subtasks using code and comments; S3: The tool agent performs targeted tool enhancement for the filtered toolset and executes the specific operations planned by the planning agent to obtain subtask results; S4: The reflective agent evaluates the subtask results and combines static and dynamic planning to plan tasks from different perspectives. S5: The planning agent summarizes the results based on the execution feedback of the subtasks.

[0019] Specifically, in one embodiment of the present invention, filtering the retrieved tools in step S2 specifically includes: Given a toolset , contains all tools and a brief description of each tool, related parameters and functions; Before task planning, the planning agent will Perform tool filtering to select a subset of tools that are highly relevant to the task , RTS is expressed as: ; in, represents the screening function of the planning agent, Indicates the tip of the filtering tool, and I indicates complex task information.

[0020] For example, in one specific example, for: You are an agent who helps users select tools from the tool library to complete user instructions; Your task is to select the toolset that can be used to complete the user's instructions based on the user's instructions; You should understand the User's instructions and determine which given tools are appropriate to complete the instructions based on each tool's description; You need to pay attention to the following rules: 1. When selecting a toolset to complete a user instruction, the toolset start marker is first generated. <toollist>", and then select one or more tools from a given tool library to complete the user's instructions; 2. Toolset start mark" <toollist> " and the toolset end marker "< / toollist> "The data type between them is a list; 3. Each tool in the TOOLLIST tool set is separated by a comma. Please strictly follow the example format to output, otherwise errors may occur.

[0021] Figure 3 This is a schematic diagram of a planning agent decomposing a complex task into easily understandable subtasks in a large language model multi-agent collaborative working method provided by an embodiment of the present invention. Figure 3 As shown, in one embodiment of the present invention, step S2 uses codes and comments to decompose complex tasks into easy-to-understand subtasks, specifically including: The planning agent analyzes the given complex task Q and performs task planning to obtain a set of logically related subtasks S. The subtask set S is expressed as: ; ; in, Indicates task planning prompts, x indicates the number of subtasks, Represents the decomposed subtasks, Represents the subtask goal, Indicates the tool that needs to be called. Indicates parameter configuration, Represents the result of a subtask.

[0022] For example, in one specific example, for: You are an agent who plans solutions to user queries; You should give your plan in natural language in the comments and generate the corresponding API calls in code. Another model will receive your API calls, execute them and provide you with the results in natural language; You need to pay attention to the following rules: 1. When you need to call the API, first generate a call start marker" <api>”, then select an API from the given API list and generate the exact API name and parameters. After the system identifies the API to be executed, it extracts the exact function call in the launch token and execution token, calls the API, and returns the required results. After the end token, you only need to continue with the task planning; 2. Please strictly follow the example output, which is a necessary condition for generating a complete plan.

[0023] For example, the planning agent plans the following instruction Q: I'm planning a trip to Paris next week and would like to generate a QR code for my ticket. Could you please generate a QR code for the following URL: "https: / / www.airline.com / ticket / ABC123"? Additionally, I need to convert the date 2022-12-25 to the Islamic calendar to check if it falls on a holiday. Finally, I'd like to generate a random UUID for my hotel reservation. Thank you! Mission Planning: #Understand what the user means: 1. Generate a QR code for the provided ticket URL; 2. Convert the date "2022-12-25" to the Islamic calendar to check whether it falls on a holiday; 3. Generate a random UUID for hotel booking.

[0024] #Step 1: Generate a QR code for the provided ticket URL, #To generate a QR code, you need to call Generate QR, from tools import Generate QR, QR_code= <api> Generate, QR(url="https: / / www.airline.com / ticket / ABC123")< / api> , #Step 1: End; #Step 2: Convert the date "2022-12-25" to the Islamic calendar. #To convert the date "2022-12-25" to the Islamic calendar, you need to call Convert Date, from tools import Convert Date, Calendar_date= <api> Convert Date(to="Islamic",date="2022-12-25")< / api> , #Step 2: End; #Step 3: Generate a random UUID for hotel booking, #To generate a random UUID, you need to call Generate UUID, from tools import Generate UUID, Random_UUID= <api> Generate UUID()< / api> , #Step 3: End; #The last step: summarize the above results, #Summarize the content and provide feedback to the user on the task completion status, #Final step: End.

[0025] In one embodiment of the present invention, the tool agent in step S3 performs targeted tool enhancement on the filtered toolset, specifically including: The tool agent passes the subtask goal Required tools and enhanced tool target descriptions To determine the parameter configuration of the calling tool , parameter configuration Expressed as: ; in, represents the enhancement function of the tool agent, Represents a prompt for generating a tool parameter.

[0026] For example, in one specific example, for: You are an agent responsible for generating the parameters for calling the tool under relevant instructions according to the description of the tool documentation; You will get the objectives and information of the relevant instructions as well as the selected tool and its description.

[0027] You need to pay attention to the following rules: 1. Please recall the tool parameters in the specified format according to the given tool description, subtask target and selected tool; 2. You only need to confirm the parameters and types required to execute the tool. No additional operations are required.

[0028] For example, the original description of the tool Convert Date: Name: Convert Date Description: Convert the date to the specified calendar; Parameter Description: Parameter 1: "to", Description: "Convert the calendar of the specified date", Required: true, Parameter 2: "date", Description: "a date format", Required: false, Parameter 3: "from", describe:"", Required: false.

[0029] The Convert Date tool has been enhanced specifically for the user instructions shown. The enhanced tool is described as follows: Name: Convert Date, Description: This tool allows users to quickly convert a Gregorian date to the Islamic calendar (or other supported calendars) and accurately determine the date's significance in the target calendar. Parameter 1: "to", Description: "Specifies the target calendar to convert to. In user scenarios, this parameter should be "Islamic" to convert calendar dates to the Islamic calendar." Required: true, Parameter 2: "date", Description: "Specifies the date to convert. In user instructions, this parameter should be set to the specific calendar date the user wishes to convert." Required: false, Parameter 3: "from", Description: "Specifies the calendar to which the input date belongs. In user commands, this parameter should be the Gregorian calendar." Required: false.

[0030] Furthermore, in one embodiment of the present invention, the step S3 of executing the specific operation planned by the planning agent to obtain the feedback result specifically includes: The tool agent uses the results obtained from using external resources to Perform task summary and obtain subtask results , subtask results Expressed as: ; in, A prompt indicating a task summary. Indicates parameter configuration Use the configured interpreter to call the required tool .

[0031] For example, in one specific example, for: You are an agent responsible for executing the results of each subtask and summarizing it after the task planning; Your task is to generate the parameters for executing each subtask call tool based on the user's instructions and the task plan.

[0032] You need to pay attention to the following rules: 1. You just need to call the tool, fill in the correct parameters, and the execution results will be returned by another model; 2. Summarize the generated results based on the returned results and the target of the current instruction.

[0033] In one embodiment of the present invention, the reflective agent evaluation subtask result in step S4 specifically includes: After the tool agent completes the subtask Afterwards, the reflective agent describes and subtask goals Determine whether the tool selected for the current subtask is correct. If not, generate a mark <re-select>, indicating that the tool to be called for this subtask needs to be reselected. The tool agent needs to reselect the tool based on the feedback and subtask goal. Reselect the tool and execute; Reflective agent determines subtask outcomes Whether to include exception error information. If it does, generate a tag. <re-call>, indicating that the tool call failed and the parameter configuration of the calling tool needs to be reconsidered , the tool agent needs to analyze and select parameter configuration again based on the feedback implement; Reflective agent further determines subtask results Is it consistent with the subtask goal? If not, generate a mark <re-sum>, indicating that the tool agent needs to perform the subtask again and summarize based on the feedback.

[0034] Furthermore, in one embodiment of the present invention, the step S4 of combining static planning with dynamic planning to plan tasks from different perspectives specifically includes: Reflective Agent Combined with Completed Subtasks information to plan the next subtask , specifying subtasks with comments and code the goals and actions to be performed; Subtask The subtask corresponding to the subtask set S If the two mission objectives are compared or calling tools If not, a mark is generated <re-plan>, indicating that the subsequent tasks should be re-planned, and the reflection agent summarizes the information and feeds it back to the planning agent.

[0035] The planning agent is responsible for static planning, and the reflection agent is responsible for dynamic planning and adjusting task planning. The main purpose of dynamic planning is to correct errors in task planning and tool calling in a timely manner. It mainly includes: when the tool agent completes each subtask, the reflection agent checks whether the tool is selected correctly and the subtask results are correct. Whether there is an exception, subtask result With subtask goals Are they consistent? Dynamically adjust and reselect tools based on the above situations, re-call tools, and re-execute subtasks. Subtask results With subtask goals If they are consistent, the next subtask is executed.

[0036] In one embodiment of the present invention, the planning agent in step S5 summarizes the results based on the execution feedback of the subtasks, specifically including: After the tool agent completes all subtasks When , and after the reflection agent gives the feedback that there is no error, the planning agent calculates the execution results R of all subtasks , summarize the summary result of the complex task Q, and the summary result is expressed as: ; in, , A prompt indicating the summary results.

[0037] For example, in one specific example, for: You are an agent responsible for summarizing the results of all subtasks. Your task is to summarize the results of the complex problem based on the subtasks planned for the complex problem and their corresponding execution results.

[0038] You need to pay attention to the following rules: 1. The results you summarize must all come from the execution results of each subtask; 2. Only summarize the execution results and do not modify the original intention.

[0039] Furthermore, step S5 further includes: Visualize the summary results.

[0040] The present invention also provides a large language model multi-agent collaborative working system, the large language model multi-agent collaborative working system comprising: A planning agent, which is used to control the overall direction of the task, filter the retrieved tools, and decompose complex tasks into easy-to-understand subtasks using code and comments; A tool agent, which is used to perform targeted tool enhancement on the filtered tool set and execute the specific operations planned by the planning agent to obtain subtask results; A reflective agent is used to evaluate subtask results and plan tasks from different perspectives by combining static planning with dynamic planning.

[0041] The large language model multi-agent collaborative work system provided by the present invention improves the ability of each agent to handle tasks by simulating social division of labor and cooperation. The planning agent will filter the retrieved tools and decompose complex tasks into easy-to-understand sub-tasks in the form of code and annotations. The tool agent is used to perform targeted tool enhancement on the filtered tool set and execute the specific operations planned by the planning agent to obtain sub-task results. The reflection agent evaluates the sub-task results and task planning, and combines static planning with dynamic planning to plan tasks from different perspectives. Through the collaborative work of multiple agents, the full-process automated processing and dynamic optimization of complex tasks are realized, multi-level semantic analysis of user instructions is performed, domain tasks are decomposed into executable sub-task flows, and available tools are intelligently arranged to obtain external resource information to assist in solving tasks. On this basis, a dynamic optimization mechanism is introduced to reflect and evaluate whether the tasks are completed efficiently.

[0042] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents. < / api> < / toollist>

Claims

1. A large language model multi-agent collaborative working method, characterized by: The following steps are involved: S1: Design multiple intelligent agents to simulate social division of labor and collaborative cooperation, so that each intelligent agent can perform different professional capabilities. Multiple intelligent agents include planning agents, tool agents, and reflective agents. S2: When task input is obtained, the planning agent filters the retrieved tools and decomposes the complex task into easy-to-understand subtasks using code and comments; S3: The tool agent performs targeted tool enhancement for the filtered toolset and executes the specific operations planned by the planning agent to obtain subtask results; S4: The reflective agent evaluates the subtask results and combines static and dynamic planning to plan tasks from different perspectives. S5: The planning agent summarizes the results based on the execution feedback of the subtasks.

2. The large language model multi-agent collaborative working method according to claim 1 is characterized in that: The filtering of the retrieved tools in step S2 specifically includes: Given a toolset , contains all tools and a brief description of each tool, related parameters and functions; Before task planning, the planning agent will analyze the toolset Perform tool filtering to select a subset of tools that are highly relevant to the task , RTS is expressed as: ; in, represents the screening function of the planning agent, Indicates the tip of the filtering tool, and I indicates complex task information.

3. The large language model multi-agent collaborative working method according to claim 2 is characterized in that: In step S2, the complex task is decomposed into easy-to-understand subtasks using code and comments. Specifically, the following subtasks are included: The planning agent analyzes the given complex task Q and performs task planning to obtain a set of logically related subtasks S. The subtask set S is expressed as: ; ； in, Indicates task planning prompts, x indicates the number of subtasks, Represents the decomposed subtasks, Represents the subtask goal, Indicates the tool that needs to be called. Indicates parameter configuration, Represents the result of a subtask.

4. The large language model multi-agent collaborative working method according to claim 3 is characterized in that: In step S3, the tool agent performs targeted tool enhancement for the filtered toolset, specifically including: The tool agent passes the subtask goal Required tools and enhanced tool target descriptions To determine the parameter configuration of the calling tool , parameter configuration Expressed as: ； in, represents the enhancement function of the tool agent, Represents a prompt for a generated tool parameter.

5. The large language model multi-agent collaborative working method according to claim 4 is characterized in that: The specific operation planned by the planning agent and the feedback results obtained by executing it in step S3 specifically include: The tool agent uses the results obtained from using external resources to Perform task summary and obtain subtask results , subtask results Expressed as: ； in, A prompt indicating a task summary. Indicates parameter configuration Use the configured interpreter to call the required tool .

6. The large language model multi-agent collaborative working method according to claim 1, characterized in that: The results of the reflective agent evaluation subtask in step S4 specifically include: After the tool agent completes the subtask Afterwards, the reflective agent describes and subtask goals Determine whether the tool selected for the current subtask is correct. If not, generate a mark <re-select>, indicating that the tool to be called for this subtask needs to be reselected. The tool agent needs to reselect the tool based on the feedback and subtask goal. Reselect the tool and execute; Reflective agent determines subtask outcomes Whether to include exception error information. If it does, generate a tag. <re-call>, indicating that the tool call failed and the parameter configuration of the calling tool needs to be reconsidered , the tool agent needs to analyze and select parameter configuration again based on the feedback implement; The reflective agent further determines Whether If not, generate a mark <re-sum> , indicating that the tool agent needs to perform the subtask again and summarize based on the feedback.< / re-sum> 7. The large language model multi-agent collaborative working method according to claim 6, characterized in that: The step S4 combines static planning with dynamic planning to plan tasks from different perspectives, specifically including: Reflective Agent Combined with Completed Subtasks information to plan the next subtask , specifying subtasks with comments and code the goals and actions to be performed; Subtask The subtask corresponding to the subtask set S If the two mission objectives are compared or calling tools If not, a mark is generated <re-plan> , indicating that the subsequent tasks should be re-planned, and the reflection agent summarizes the information and feeds it back to the planning agent.< / re-plan> 8. The large language model multi-agent collaborative working method according to claim 1 is characterized in that: The planning agent in step S5 summarizes the results based on the execution feedback of the subtasks, specifically including: After the tool agent completes all subtasks When the reflection agent gives the correct feedback, the planning agent calculates the execution results of all subtasks. , summarize the summary result of the complex task Q, and the summary result is expressed as: ； in, , A prompt indicating the summary results.

9. The large language model multi-agent collaborative working method according to claim 8, characterized in that: Step S5 further includes: Visualize the summary results.

10. A large language model multi-agent collaborative working system, characterized by: The large language model multi-agent collaborative working system includes: A planning agent, which is used to control the overall direction of the task, filter the retrieved tools, and decompose complex tasks into easy-to-understand subtasks using code and comments; A tool agent, which is used to perform targeted tool enhancement on the filtered tool set and execute the specific operations planned by the planning agent to obtain subtask results; A reflective agent is used to evaluate subtask results and plan tasks from different perspectives by combining static planning with dynamic planning.

Images