A penetration testing system and method based on multi-agent collaboration

Abstract

The present invention provides a penetration testing system and method based on multi-agent collaboration, comprising an interaction module, a plan generation agent using a large language model (LLM), a plan planning agent, and a plan execution agent. The interaction module receives a problem to be solved input by a user; the plan generation agent generates a penetration testing task plan based on the problem to be solved; the plan planning agent splits the penetration testing task plan into multiple sub-plans; and utilizes reinforcement learning technology to select the best-matching API for the optimal tool described in the current plan, extracts the parameters required by the current API from the sub-tasks, and generates the final penetration testing command; the plan execution agent receives the penetration testing command delivered by the plan planning agent, executes the command, and records the resulting execution results. This invention allows users to simply provide a penetration testing goal and objective, and the system automatically generates a penetration testing plan, performs the penetration testing task, and ultimately delivers the results.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer security detection, and in particular relates to a penetration testing system and method based on multi-agent collaboration. Background Art

[0002] Penetration testing is a security assessment method that simulates attacker behavior to evaluate computer systems, networks, or applications to check their security. Its primary purpose is to simulate hacker attacks to identify vulnerabilities, weaknesses, and security risks within the system. In traditional penetration testing, penetration testers typically utilize automated tools to analyze the target system. The penetration testing process is generally divided into seven steps: initial interaction, intelligence gathering, threat modeling, vulnerability analysis, penetration attack, post-penetration attack, and report generation. Traditional penetration testing requires significant expertise and experience from the tester, requiring them to be able to use a variety of penetration testing tools and effectively combine them.

[0003] Current automated penetration testing tools generally integrate the functionality of multiple penetration testing tools, making them easier for testers to use. However, these tools still require strong technical expertise and familiarity with the tools. Despite the availability of numerous excellent tools and methods in this field, fully automated penetration testing remains elusive. The difficulty in automating penetration testing lies in the comprehensive knowledge required to understand and discover various vulnerabilities. In practice, penetration testers typically employ a combination of depth-first and breadth-first search. Breadth-first search is used to gain a comprehensive understanding of the target, followed by depth-first search to test specific services and vulnerabilities. However, this approach requires extensive experience and expertise. Summary of the Invention

[0004] In order to solve the problem that existing automated penetration testing methods still rely on the experience and professional knowledge of testers, the present invention provides a penetration testing system based on multi-agent collaboration, aiming to provide an intelligent and automated penetration testing system through the collaborative work of multiple agents, so as to significantly improve the efficiency, accuracy and intelligence level of penetration testing.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A penetration testing system based on multi-agent collaboration includes an interaction module and a plan generation agent including a large language model (LLM), a plan planning agent, and a plan execution agent, wherein:

[0007] The interaction module is used to receive the problem to be solved input by the user;

[0008] The plan generation agent is used to generate a penetration test task plan based on the problem to be solved;

[0009] The planning agent is used to split the penetration test task plan into multiple sub-plans; and uses reinforcement learning technology to screen the best-fitting API corresponding to the optimal tool described in the current plan, and extracts the parameters required by the current API from the sub-tasks to generate the final penetration test command;

[0010] The plan execution agent is used to receive the penetration test command transmitted by the plan planning agent, execute the command, and record the generated execution result.

[0011] Preferably, it also includes an analysis and summary module, which is used to generate a penetration test and risk assessment report for the problem to be solved input by the current user based on the execution results recorded by the planned execution agent and the general summary generated by the LLM in each agent, and discover vulnerabilities based on the penetration test and risk assessment report.

[0012] Preferably, the problem to be solved input by the user includes the target's IP or address information, and if the information is incomplete, the system will exit in time; the penetration test task plan generated by the planning agent includes the target information given by the user.

[0013] Preferably, the method further includes: if the planning agent is not in the initial state, the planning agent receives the execution result of the plan execution agent, updates and improves the current task plan according to the current task list and the execution result of the previous round of plan, and generates a new subtask list;

[0014] For each subtask, LLM alignment technology is applied to generate the corresponding plan description.

[0015] Preferably, the method of using reinforcement learning technology to screen out the best-fitting API corresponding to the optimal tool described in the current plan, extracting the parameters required by the current API from the subtasks, and generating the final penetration test command specifically includes:

[0016] Integrate penetration testing tools into APIs based on their functions, build plug-in libraries, and perform modular management;

[0017] Using the secondary screening algorithm and RAG technology, the best-matching API corresponding to the optimal tool described in the current plan is selected from the plug-in library;

[0018] Using the SymbCoT algorithm, the LLM is enabled with a standardized and precise description paradigm to extract the core parameters required by the current API from the subtask and generate the final penetration test command.

[0019] Preferably, the plan execution agent is used to receive the penetration test command delivered by the plan planning agent, execute the command, and record the generated execution results, specifically including:

[0020] The plan execution agent receives the specific penetration test command transmitted by the plan planning agent, executes the command, and obtains the execution result of the penetration test tool;

[0021] Send the execution results of the penetration testing tool to the LLM of the plan execution agent to generate a summary of the current penetration testing commands;

[0022] Determine the execution results of the penetration testing tool. If all subtasks are not completed, go to the step of generating multiple sub-plans and feed back the execution results of the penetration testing tool to the planning agent; if all subtasks are completed, the user's penetration testing requirements have been met.

[0023] Preferably, when the plan generation agent receives the problem to be solved input by the user and generates a task plan for the automated penetration test, the system exits in a timely manner for input whose decision result is that it is not executable.

[0024] The present invention also provides a penetration testing method based on multi-agent collaboration, comprising:

[0025] The plan generation agent generates a penetration test task plan based on the issues to be resolved;

[0026] The planning agent splits the penetration test task plan into multiple sub-plans. It then uses reinforcement learning technology to select the best-matching API for the optimal tool described in the current plan, extracts the parameters required by the current API from the sub-tasks, and generates the final penetration test command.

[0027] The plan execution agent receives the penetration test command delivered by the plan planning agent, executes the command, and records the generated execution results;

[0028] Among them, the plan generation agent, plan planning agent and plan execution agent all contain the large language model LLM.

[0029] The present invention also provides a computer device, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the penetration testing method based on multi-agent collaboration.

[0030] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is loaded by a processor, it can execute the steps described in the penetration testing method based on multi-agent collaboration.

[0031] The multi-agent collaborative penetration testing method provided by the present invention has the following beneficial effects:

[0032] The present invention uses a multi-intelligent agent approach to set up a plan generation agent, a plan planning agent, and a plan execution agent that include a large language model (LLM), so that the agent can act as a decision maker, executor, and other testers in traditional penetration testing. The user only needs to provide a penetration test goal and purpose, and the system can automatically provide a penetration test plan and perform penetration test tasks in a collaborative manner, and finally provide the results.

[0033] The automated task plan generation by the plan generation agent avoids the tediousness and errors of manual operations and enables timely exit for unexecutable input, ensuring the rationality and feasibility of the task and avoiding unnecessary waste of resources. The plan planning agent breaks down complex tasks into multiple sub-plans and, using reinforcement learning, selects the optimal tool and its optimal API adaptation solution, ensuring efficient and accurate processing of each sub-task and improving the overall performance of the penetration testing process. This solution allows users to accurately and efficiently complete penetration testing tasks without requiring specific tool usage or in-depth penetration testing knowledge, reducing labor costs and improving penetration efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0035] Figure 1 Flowchart of a penetration testing method based on multi-agent collaboration according to Example 1 of the present invention;

[0036] Figure 2 To generate a flowchart of penetration testing tasks;

[0037] Figure 3 A flowchart for splitting penetration testing tasks;

[0038] Figure 4 A flowchart for generating penetration testing commands;

[0039] Figure 5 A flowchart for performing penetration testing tasks;

[0040] Figure 6 A block diagram of a penetration testing system based on multi-agent collaboration provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0042] Example 1

[0043] The present invention provides a penetration testing method based on multi-agent collaboration, specifically Figure 1 As shown, the method includes:

[0044] S1: The planner agent receives user input of a problem to be solved and generates a task plan for the automated penetration test. If the input is deemed unworkable, the system exits immediately.

[0045] In one embodiment, S1 includes:

[0046] S11: The user input is natural language and needs to include information such as the target IP or address. If the information is incomplete, the system will exit in time.

[0047] S12: Generate a penetration test plan based on the user input, where the plan includes the target information provided by the user.

[0048] Specifically, such as Figure 2 As shown, step S11 specifically includes: the address information of the target host or system, such as address information in the form of IP:PORT; and the user's penetration testing requirements, such as performing a vulnerability scan on the target address or performing port probing on the target. If the plan generation agent determines that the user-entered information is incomplete or the user's penetration testing requirements cannot be met, it will safely exit and output a corresponding prompt message, prompting the user to modify the requirements or provide additional information. Step S12 specifically includes: the plan generation agent generates a rough list of penetration testing task plans based on the user's input.

[0049] S2: The planning agent (Refiner agent) uses the Automatic multi-step reasoning (ART) algorithm (or HTN decomposition algorithm, HRL algorithm, behavior tree algorithm) to split or refine the plan based on the task plan and the execution results of the previous round of tasks (excluding the first round) to generate accurate sub-plans.

[0050] like Figure 3 As shown, in one embodiment, S2 includes:

[0051] S21: If the planning agent is in the initial state (i.e., the first cycle), the planning agent receives the task plan generated by the plan generation agent and uses the Automatic multi-step reasoning (ART) algorithm to split the plan into subtask plans in a more scientific and efficient manner.

[0052] S22: If the planning agent is not in the initial state (i.e., after the first cycle), the planning agent receives the execution result of the plan execution agent, updates and improves the current task plan based on the current task list and the execution result of the previous round of plan, and generates a new subtask list.

[0053] S23: For each subtask, use LLM alignment technology to generate the corresponding plan description.

[0054] Specifically, for example, step S21 specifically includes: judging the current Agent state. If it is the first round of loop, that is, the plan list generated by the plan generation agent after the user input is passed to the plan planning agent, and no task has been executed, the plan planning agent uses the Automatic multi-step reasoning (ART) algorithm to split the overall plan and formulate sub-task plans. The requirements of each sub-task correspond to an API in the plug-in library. Step S22 specifically includes: if it is after the first round of loop, that is, a sub-task has been completed and the plan planning agent has received the execution result from the plan execution agent. The plan planning agent updates the current task plan list based on the current task plan list and the execution results of the previous round of tasks.

[0055] S3: Use reinforcement learning technology to screen out the best-fitting API for the optimal tool corresponding to the current plan description, and use the SymbCoT algorithm (or Zero-Shot and Few-Shot Reasoning method, Prompt Chaining algorithm, Constraint-Based Task Planning algorithm) to extract the core parameters required by the current API from the subtask and generate the final penetration test command.

[0056] like Figure 4 As shown, in one embodiment, S3 includes:

[0057] S31: Integrate penetration testing tools into APIs based on their functions, build a plug-in library, and perform modular management.

[0058] S32: Using the secondary screening algorithm and RAG technology, the best-fitting API corresponding to the optimal tool described in the current plan is selected from the plug-in library built in S31.

[0059] S33: Using the SymbCoT algorithm, the LLM is enabled to extract the core parameters required by the current API from the subtasks with a standardized and precise description paradigm, and generate the final penetration test command.

[0060] Specifically, for example, step S31 includes: integrating various penetration testing tools and abstracting the functions of the penetration testing tools into APIs. Each API requires description information of the corresponding function and description information of the parameter meaning and format. Step S32 includes: using RAG technology to find multiple APIs whose functional description information matches the user's needs from a large number of APIs, and then using LLM to select the most suitable API from the multiple APIs. Step S33 includes: after selecting the corresponding API, using LLM to generate the parameters required by the API based on the API description information and requirements, and generate the final penetration testing command.

[0061] S4: The Executor Agent receives the specific command information from the Planning Agent and executes it. The results are summarized and recorded. If all subtasks are not completed, the process returns to step S2 until all subtasks are completed.

[0062] like Figure 5 As shown, in one embodiment, S4 includes:

[0063] S41: The plan execution agent receives the specific penetration test command delivered by the plan planning agent, executes the command, and obtains the execution result of the penetration test tool.

[0064] S42: Obtain the execution results of the penetration test tool from S41. The execution results will be sent to the LLM to generate a high-level summary of the current penetration test command.

[0065] S43: Obtain the execution results of the penetration testing tool from S41. If all subtasks are not completed, proceed to step S2 and feed the execution results of the penetration testing tool back to the planning agent to help the planning agent more flexibly adjust and optimize the subtasks. If all subtasks are completed, it means that the user's penetration testing requirements have been met.

[0066] Specifically, for example, step S41 specifically includes: the plan execution agent calls the corresponding function of the corresponding penetration testing tool in the plug-in library according to the penetration testing command, and obtains the result after the penetration testing tool function is executed. Step S42 specifically includes: inputting the penetration testing command and the function execution result of the corresponding penetration testing tool to LLM, and letting LLM make a highly generalized summary of it. Step S43 specifically includes: obtaining the execution result of the penetration testing tool from S41, and judging whether all subtasks have been completed according to the subtask list and the execution result. If it is judged that the subtasks have not been completed, the penetration test execution result will be fed back to the planning agent, so that it can improve and optimize the subtask list according to the current task list and the penetration test execution result, and give the subtasks to be completed in the next round; if it is judged that the subtasks have been completed, it means that the user's penetration testing requirements have been completed, and there is no need to continue to generate or modify the subtask list.

[0067] S5: Generate an automated penetration test and risk assessment report for the current user input based on the core metadata of the execution results and the summary generated by the LLM.

[0068] In one embodiment, S5 includes generating an automated penetration test and risk assessment report for the current user input based on the core data of the execution results of all subtasks and the summary generated by the LLM. Specifically, the penetration test commands used by all subtasks and the execution results of all subtasks are recorded. Based on the penetration test commands and execution results of all subtasks, as well as all the summary content generated in S42, a risk assessment report for the current user input is generated.

[0069] The multi-agent collaborative penetration testing method provided by the present invention has the following beneficial effects:

[0070] (1) First, the present invention provides a penetration testing method based on multi-agent collaboration, which can significantly improve the efficiency and intelligence level of penetration testing. Through the automated task plan generation of the plan generation agent, the tediousness and errors of manual operation are avoided, and unexecutable input can be exited in time, thus ensuring the rationality and feasibility of the task and avoiding unnecessary waste of resources.

[0071] (2) Secondly, the planning agent uses the Automatic multi-step reasoning (ART) algorithm to scientifically decompose complex tasks into multiple sub-plans and uses LLM alignment technology to accurately describe each sub-task. Combined with reinforcement learning, it selects the optimal tool and its best API adaptation solution. Through the modular management plug-in library, flexible tool calling is achieved, ensuring the efficient and accurate processing of each sub-task, improving the execution effect of the entire penetration testing process.

[0072] (3) In addition, the SymbCoT algorithm accurately extracts API core parameters to generate the final penetration test command, making the task description and tool call more closely aligned and adaptable, ensuring the effectiveness and accuracy of the test command. The ExecutorAgent receives and executes the penetration test command, and forms a closed-loop optimization through feedback from the execution results, helping the planning agent dynamically adjust the task strategy. In this way, the system can flexibly adjust and optimize according to actual conditions during execution, making it more adaptable.

[0073] (4) Finally, the system can automatically generate penetration test and risk assessment reports, providing a highly summarized summary based on the core metadata of the task and LLM, thereby simplifying the report generation process and improving the visualization and insight of test information. This multi-agent collaborative penetration testing solution not only enhances the system's intelligent perception and rapid decision-making capabilities, but also significantly improves the quality and effectiveness of penetration testing, providing users with a more reliable network security assessment tool.

[0074] Example 2

[0075] Based on the same inventive concept, the present invention also provides a penetration testing system based on multi-agent collaboration, such as Figure 6 As shown in the figure, the system includes an interaction module, a plug-in library module, a management and error handling module, an analysis and summary module, and a plan generation agent (Planner agent), a plan planning agent (Refiner agent) and a plan execution agent (Executor Agent) containing a large language model (LLM). It aims to automatically generate penetration test task plans through the plan generation agent to reduce human errors, and scientifically decompose and refine tasks through the plan planning agent, automatically select the most appropriate penetration tools and APIs, and improve execution efficiency.

[0076] Specifically, the interaction module is used to receive the problems to be solved input by the user and the system return information; the management and error handling module is used to manage the interaction and operation between the scheduling modules, and to handle the scenarios when system errors occur; the plug-in library module is used to store the functional APIs integrated by the penetration testing tools; the plan generation agent is used to judge whether the user requirements can be completed based on the user input and to generate a penetration testing task plan; the plan planning agent is used to plan the penetration testing task plan, decompose the plan into multiple subtasks, and screen out the appropriate API for each subtask to generate a penetration testing command; the plan execution agent is used to execute the penetration testing command and record the execution results of the penetration testing tool; the analysis and summary module is used to generate an automated penetration testing and risk assessment report for the current user query.

[0077] Each module in the multi-agent collaborative penetration testing system can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0078] The present invention also provides a computer device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of an embodiment of a penetration testing method based on multi-agent collaboration. The specific implementation method can be found in the method embodiment and will not be repeated here.

[0079] Furthermore, the present invention provides a non-transitory computer-readable storage medium containing instructions, wherein the storage medium stores a computer program. For example, the storage medium may be a memory containing instructions, wherein the instructions are executable by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device. When executed by the processor, the computer program can implement the steps of the embodiment of the penetration testing method based on multi-agent collaboration. The specific implementation method can be found in the method embodiment and will not be further described here.

[0080] Those skilled in the art will appreciate that embodiments of the present invention may provide methods, systems, or computer program products. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0081] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0082] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0084] It should be pointed out that the specific implementation methods described above can enable those skilled in the art to understand the invention more comprehensively, but do not limit the invention in any way. Therefore, although the present specification and examples have described the invention in detail, those skilled in the art should understand that the invention can still be modified or replaced by equivalents; and all technical solutions and improvements that do not deviate from the spirit and scope of the invention are included in the scope of protection of the patent for the invention. Any figure mark in the claims should not be regarded as limiting the claims involved. Any simple change or equivalent replacement of the technical solution that can be obviously obtained by any person familiar with the art within the technical scope disclosed in the present invention falls within the scope of protection of the present invention.

Claims

1. A penetration testing system based on multi-agent collaboration, characterized in that: It includes an interactive module and a plan generation agent, a plan planning agent, and a plan execution agent including a large language model (LLM), among which: The interaction module is used to receive the problem to be solved input by the user; The plan generation agent is used to generate a penetration test task plan based on the problem to be solved; The planning agent is used to split the penetration test task plan into multiple sub-plans; and uses reinforcement learning technology to screen the best-fitting API corresponding to the optimal tool described in the current plan, and extracts the parameters required by the current API from the sub-tasks to generate the final penetration test command; The plan execution agent is used to receive the penetration test command delivered by the plan planning agent, execute the command, and record the generated execution result; The reinforcement learning technology is used to screen out the best-fitting API corresponding to the optimal tool described in the current plan, and the parameters required by the current API are extracted from the subtasks to generate the final penetration test command, which specifically includes: Integrate penetration testing tools into APIs based on their functions, build plug-in libraries, and perform modular management; Using the secondary screening algorithm and RAG technology, the best-matching API corresponding to the optimal tool described in the current plan is selected from the plug-in library; Using the SymbCoT algorithm, the LLM is enabled with a standardized and precise description paradigm to extract the core parameters required by the current API from the subtask and generate the final penetration test command.

2. The penetration testing system based on multi-agent collaboration according to claim 1 is characterized in that: It also includes an analysis and summary module, which is used to generate a penetration test and risk assessment report for the problem to be solved input by the current user based on the execution results recorded by the planned execution agent and the summary generated by the LLM in each agent, and discover vulnerabilities based on the penetration test and risk assessment report.

3. The penetration testing system based on multi-agent collaboration according to claim 1 is characterized in that: The problem to be solved input by the user includes the target's IP or address information. If the information is incomplete, the system will exit in time; the penetration test task plan generated by the planning agent includes the target information given by the user.

4. The penetration testing system based on multi-agent collaboration according to claim 1 is characterized in that: The process also includes: if the plan planning agent is not in the initial state, the plan planning agent receives the execution result of the plan execution agent, updates and improves the current task plan according to the current task list and the execution result of the previous round of plan, and generates a new subtask list; For each subtask, LLM alignment technology is applied to generate the corresponding plan description.

5. The penetration testing system based on multi-agent collaboration according to claim 1 is characterized in that: The plan execution agent is used to receive the penetration test command delivered by the plan planning agent, execute the command, and record the generated execution results, specifically including: The plan execution agent receives the specific penetration test command transmitted by the plan planning agent, executes the command, and obtains the execution result of the penetration test tool; Send the execution results of the penetration testing tool to the LLM of the plan execution agent to generate a summary of the current penetration testing commands; Determine the execution results of the penetration testing tool. If all subtasks are not completed, go to the step of generating multiple sub-plans and feed back the execution results of the penetration testing tool to the planning agent; if all subtasks are completed, the user's penetration testing requirements have been met.

6. The penetration testing system based on multi-agent collaboration according to claim 1 is characterized in that: When the plan generation agent receives the problem to be solved input by the user and generates a task plan for the automated penetration test, the system exits in a timely manner if the decision result is an input that is not executable.

7. A penetration testing method based on multi-agent collaboration, characterized in that: include: The interactive module receives the problem to be solved input by the user; The plan generation agent generates a penetration test task plan based on the issues to be resolved; The planning agent splits the penetration test task plan into multiple sub-plans; It also uses reinforcement learning technology to screen out the best-fitting API corresponding to the optimal tool described in the current plan, extracts the parameters required by the current API from the subtasks, and generates the final penetration test command; The plan execution agent receives the penetration test command delivered by the plan planning agent, executes the command, and records the generated execution results; Among them, the plan generation agent, plan planning agent and plan execution agent all contain the large language model LLM.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is loaded into a processor, it is capable of executing the steps of the method according to claim 7.

Images