Large-model multi-agent task scheduling method with memory and retrieval capabilities

Abstract

The invention discloses a large-model multi-agent task scheduling method with memory and retrieval capabilities, and the method comprises the steps: carrying out the cognitive analysis of a complex task based on task scheduling agents, and carrying out the distribution of subtasks through combining the functional attributes of all task execution agents and an integrated tool; each task execution agent completes execution work of a specific task through an integration tool, a task result is returned to the task scheduling agent for integration, and execution of a next link subtask or completion of the task is determined according to a task condition; maintaining task context information and a task entity information base by adopting a dynamic memory pool; and cross-agent and cross-task node multi-dimensional information sharing is realized by using cross-agent search. According to the method, the problem of state loss of a multi-agent system is solved by introducing a memory enhancement mechanism, and the context sensing capability in the task execution process is remarkably improved; cross-task and cross-agent knowledge reuse is realized through a cross-agent search mechanism, and the decision quality of the system is improved; the method is compatible with multi-source heterogeneous task requirements, breaks through the limitation of a traditional multi-agent system on task complexity, cross-agent collaboration and knowledge reuse capability, and can be widely applied to the field of intelligent scheduling of complex scenes such as emergency rescue, industrial automation and urban public management.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of large model intelligent agents, and particularly relates to a large model multi-agent task scheduling method with memory and retrieval capabilities. BACKGROUND

[0002] With the rapid development of social economy and the continuous improvement of technological level, complex scenarios such as emergency command, industrial automation, and urban public management increasingly rely on intelligent systems. These fields often face challenges such as complex tasks, large amounts of information, high real-time requirements, and multi-objective collaborative optimization. Traditional single-agent systems or rule-based task scheduling methods often show insufficient flexibility, slow response, limited global coordination capabilities, and other problems when facing the above complex scenarios, making it difficult to meet actual needs.

[0003] Currently, task scheduling systems mainly rely on preset heuristic algorithms or optimization algorithms based on static environment modeling, such as genetic algorithms, ant colony algorithms, and integer programming. These methods can achieve certain results in specific scenarios, but when faced with dynamic changes in the environment, sudden tasks, and multi-objective collaboration, they often show poor adaptability, slow response, and unstable scheduling results. In addition, existing scheduling systems often lack effective use of historical experience, and cannot extract knowledge from past scheduling processes to guide current or future task allocation.

[0004] In recent years, with the development of artificial intelligence technology, especially the rise of large models, some research attempts to introduce agents into task scheduling systems to achieve autonomous matching of tasks and resources. However, existing multi-agent scheduling methods are mostly based on local information for decision-making, lacking global perspective and long-term memory capabilities, making it difficult for the system to form effective scheduling strategies in complex and changing environments. In addition, the collaboration mechanism between multiple agents is not perfect, with limited information sharing and conflict coordination capabilities, affecting the overall scheduling efficiency and task completion quality of the system.

[0005] In summary, there is an urgent need for a task scheduling method with memory and retrieval capabilities, which can combine historical information and support multi-agent collaborative decision-making, to improve the intelligent level, adaptability, and overall performance of the task scheduling system. SUMMARY

[0006] The purpose of the present application is to provide a large model multi-agent task scheduling method with memory and retrieval capabilities, which realizes efficient decomposition, dynamic scheduling, and intelligent execution of complex tasks by constructing an agent collaboration architecture with enhanced cognitive and memory capabilities.

[0007] The technical solution for achieving the object of the present application is: on the one hand, a large model multi-agent task scheduling method with memory and retrieval capabilities is provided, which realizes efficient decomposition, dynamic scheduling and intelligent execution of complex tasks by constructing an agent collaborative architecture with enhanced cognitive and memory capabilities.

[0008] Further, the method comprises:

[0009] The task scheduling agent analyzes the complex task based on cognition, and assigns subtasks in combination with the function attributes of each task execution agent and integrated tools;

[0010] Each task execution agent completes the execution of specific tasks through integrated tools, returns the task results to the task scheduling agent for integration, and decides to execute the next subtask or end the task according to the task situation;

[0011] A dynamic memory pool is used to maintain the task context information and task entity information library;

[0012] Cross-agent retrieval is used to realize multi-dimensional information sharing across agents and task nodes.

[0013] On the other hand, a large model multi-agent task scheduling system with memory and retrieval capabilities is provided, which comprises a multi-agent task scheduling framework, a memory enhancement module and a cross-agent retrieval module;

[0014] The multi-agent task scheduling framework, as a core module for receiving user requests and completing task decomposition, task scheduling and task execution, adopts a cognitive architecture based on a large model, and includes a task decomposer, a dynamic scheduler and a state coordinator;

[0015] The memory enhancement module is used to store task entity parameters during task execution, adopts a key-value storage structure, and saves specified parameters in the form of a dynamic memory pool in the task scheduling and execution thread, and can perform local persistent storage according to requirements;

[0016] The cross-agent retrieval module is used for multi-agent to retrieve task parameters from the short-term memory pool, designs and integrates a key-value based retrieval tool, and realizes cross-agent state and information sharing.

[0017] Further, the retrieval module can be called by the agent as an external tool.

[0018] Furthermore, the multi-agent task scheduling framework includes a task scheduling agent and a task execution agent; and the multi-agent task scheduling framework adopts a two-layer architecture, wherein the upper layer is the task scheduling agent, which is used to receive task requests, decompose task logic and dynamically allocate sub-tasks, and the lower layer consists of several task execution agents, including several functionally specialized agents, each agent integrating cognitive reasoning functions and tool calling interfaces, and is responsible for executing specific tasks.

[0019] Furthermore, the task scheduling agent is used for task analysis and scheduling, driven by a large model, and has the ability to decompose and route tasks. In the task decomposition and scheduling phase, after receiving user input, the task scheduling agent decomposes complex tasks into multiple sub-tasks based on Prompt and thought chains, and dynamically allocates them according to the professional capabilities of each task execution agent. During task execution, it maintains the global state, tracks the execution progress of each task execution agent, and handles the dependencies between tasks. After the task execution is completed, it integrates the output results of each task execution agent and returns them to the user.

[0020] The task execution agent is used to execute specific tasks. It adopts a reasoning-then-action approach, where a large model generates logical judgments based on the current state, formulates action strategies, and then executes specific operations by calling integrated tools.

[0021] Furthermore, the task scheduling agent adopts a two-layer architecture, in which the upper-layer task scheduling agent is responsible for task decomposition, sub-task allocation and global state maintenance; the lower-layer task execution agent is configured according to functional specialization and integrates cognitive reasoning functions and tool call interfaces.

[0022] Furthermore, the task execution agent adopts a three-stage execution method: the reasoning stage, which generates an action strategy based on the current task description, prompts, context and historical information; the action stage, which calls the business tools defined in the tool registry or the cross-agent retrieval module; and the feedback stage, which updates the dynamic memory pool according to the tools or retrieval results and determines whether the sub-task has been completed.

[0023] Furthermore, the cross-agent retrieval module supports preset retrieval modes. Based on task information requirements and task dependencies, different retrieval modes are matched by keywords, supporting multi-information type retrieval of current task information and historical task information, as well as multi-scenario retrieval across tasks and across agents.

[0024] On the other hand, a task scheduling method for a large-scale multi-agent task scheduling system with memory and retrieval capabilities is provided, the task scheduling method comprising:

[0025] The task scheduling agent of the multi-agent task scheduling framework receives user instructions, decomposes complex tasks into multiple sub-tasks through large model semantic understanding, and dynamically allocates sub-tasks to the corresponding task execution agents according to the professional capabilities of each task execution agent.

[0026] The task execution agent of the multi-agent task scheduling framework performs reasoning analysis through an integrated large model, formulates action strategies, and calls business tools or retrieval tools to execute tasks.

[0027] The memory enhancement module records entity parameters in real time during task execution using a key-value storage structure, forming a dynamic memory pool;

[0028] The cross-agent retrieval module retrieves historical information across tasks or across agents from a dynamic memory pool based on a predefined retrieval pattern, supporting decision-making by the task-executing agent.

[0029] The task execution agent returns the execution results to the task scheduling agent, which then integrates the results of all subtasks and generates user feedback.

[0030] Compared with the prior art, the significant advantages of this invention are:

[0031] (1) The present invention maintains the context information during the task execution process through a dynamic memory module, which can effectively solve the problem of state loss of traditional intelligent agents.

[0032] (2) This invention utilizes a cross-agent retrieval module to achieve cross-task and cross-agent knowledge transfer and reuse, thereby improving the overall decision-making quality of the system.

[0033] (3) Based on the feedback mechanism of memory and retrieval, this invention supports online optimization of scheduling strategies, continuous execution of tasks, and continuous evolution of agent capabilities.

[0034] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0035] Figure 1 This is a flowchart of a large-scale multi-agent task scheduling method with memory and retrieval capabilities in one embodiment.

[0036] Figure 2 This is a schematic diagram of the multi-agent task scheduling framework in one embodiment.

[0037] Figure 3 This is a schematic diagram of the task execution agent composition in one embodiment.

[0038] Figure 4 This is a flowchart of the subtask execution process of a task execution intelligent agent in one embodiment.

[0039] Figure 5 This is a business relationship diagram of the memory enhancement module, the agent, and the cross-agent retrieval module in one embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0041] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0042] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0043] In one embodiment, a large-scale multi-agent task scheduling method with memory and retrieval capabilities is provided. The method achieves efficient decomposition, dynamic scheduling, and intelligent execution of complex tasks by constructing an agent collaborative architecture with enhanced cognitive and memory capabilities.

[0044] Furthermore, in one embodiment, combined with Figure 1 The method includes:

[0045] Based on the task scheduling agent, complex tasks are cognitively analyzed, and sub-tasks are allocated by combining the functional attributes of each task execution agent and integration tools.

[0046] Each task execution agent completes the execution of specific tasks through the integration tool, returns the task results to the task scheduling agent for integration, and decides to execute the next sub-task or end the task based on the task situation.

[0047] A dynamic memory pool is used to maintain the task context information and task entity information database;

[0048] Cross-agent retrieval enables multi-dimensional information sharing across agents and task nodes.

[0049] In one embodiment, a large model multi-agent task scheduling system with memory and retrieval capabilities is provided to implement a large model multi-agent task scheduling method with memory and retrieval capabilities. The system includes a multi-agent task scheduling framework, a memory enhancement module, and a cross-agent retrieval module.

[0050] The multi-agent task scheduling framework, as the core module for receiving user requests and completing task decomposition, task scheduling, and task execution, adopts a cognitive architecture based on a large model and includes a task decomposer, a dynamic scheduler, and a state coordinator.

[0051] The memory enhancement module is used to store task entity parameters during task execution. It adopts a key-value storage structure and uses a dynamic memory pool to save the specified parameters in the task scheduling and execution threads, and can perform local persistent storage as needed. Here, the task entity parameters include task status, tool call results, and cross-agent shared information.

[0052] The cross-agent retrieval module is used for multiple agents to retrieve task parameters from the short-term memory pool, and designs and integrates key-value-based retrieval tools to achieve cross-agent state and information sharing.

[0053] Preferably, in some embodiments, the retrieval module can be invoked by the intelligent agent as an external tool.

[0054] Preferably, in some embodiments, the multi-agent task scheduling framework includes a task scheduling agent and a task execution agent; and the multi-agent task scheduling framework adopts a two-layer architecture, wherein the upper layer is a task scheduling agent, which is used to receive task requests, decompose task logic and dynamically allocate sub-tasks, and the lower layer consists of several task execution agents, including several functionally specialized agents, each agent integrating cognitive reasoning functions and tool calling interfaces, and is responsible for executing specific tasks.

[0055] Preferably, in some embodiments, the task scheduling agent is used for task analysis and scheduling, driven by a large model, and has the ability to decompose and route tasks. In the task decomposition and scheduling phase, after receiving user input, the task scheduling agent decomposes complex tasks into multiple sub-tasks based on Prompt and thought chains, and dynamically allocates them according to the professional capabilities of each task execution agent. During task execution, it maintains the global state, tracks the execution progress of each task execution agent, and handles the dependencies between tasks. After the task execution is completed, it integrates the output results of each task execution agent and returns them to the user.

[0056] The task execution agent is used to execute specific tasks. It adopts a reasoning-then-action approach. The large model generates logical judgments based on the current state, formulates action strategies, and then executes specific operations by calling integrated tools (algorithm model, memory module, retrieval module, database).

[0057] Here, the task execution agent is highly configurable and can be added or removed as needed.

[0058] Preferably, in some embodiments, the task scheduling agent adopts a two-layer architecture, wherein the upper-layer task scheduling agent is responsible for task decomposition, sub-task allocation and global state maintenance; the lower-layer task execution agent is configured according to functional specialization and integrates cognitive reasoning functions and tool calling interfaces.

[0059] Preferably, in some embodiments, the task execution agent adopts a three-stage execution method: inference stage, generating an action strategy based on the current task description, prompt words, context and historical information; action stage, calling the business tools defined in the tool registry or the cross-agent retrieval module; feedback stage, updating the dynamic memory pool according to the tools or retrieval results, and determining whether the sub-task has been completed.

[0060] Preferably, in some embodiments, the cross-agent retrieval module supports preset retrieval modes. Based on task information requirements and task dependencies, it matches different retrieval modes using keywords, supporting multi-information type retrieval of current task information and historical task information, as well as multi-scenario retrieval across tasks and across agents.

[0061] Preferably, in some embodiments, the task scheduling agent communicates with the task execution agent through a structured message passing protocol, and the message content includes subtask instructions, execution status, and result data.

[0062] In one embodiment, a task scheduling method is provided for implementing the large-scale multi-agent task scheduling system with memory and retrieval capabilities, the task scheduling method comprising:

[0063] The task scheduling agent of the multi-agent task scheduling framework receives user instructions, decomposes complex tasks into multiple sub-tasks through large model semantic understanding, and dynamically allocates sub-tasks to the corresponding task execution agents according to the professional capabilities of each task execution agent.

[0064] The task execution agent of the multi-agent task scheduling framework performs reasoning analysis through an integrated large model, formulates action strategies, and calls business tools or retrieval tools to execute tasks.

[0065] The memory enhancement module records entity parameters in real time during task execution using a key-value storage structure, forming a dynamic memory pool;

[0066] The cross-agent retrieval module retrieves historical information across tasks or across agents from a dynamic memory pool based on a predefined retrieval pattern, supporting decision-making by the task-executing agent.

[0067] The task execution agent returns the execution results to the task scheduling agent, which then integrates the results of all subtasks and generates user feedback.

[0068] The business relationship diagram between the memory enhancement module and the agent and cross-agent retrieval modules is as follows: Figure 5 As shown.

[0069] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the task scheduling method described above.

[0070] For specific limitations on each step, please refer to the limitations on task scheduling methods mentioned above, which will not be repeated here.

[0071] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the task scheduling method described above.

[0072] For specific limitations on each step, please refer to the limitations on task scheduling methods mentioned above, which will not be repeated here.

[0073] The specific steps of the large-scale multi-agent task scheduling method with memory and retrieval capabilities of the present invention will be described in detail below.

[0074] like Figure 1 As shown, the task scheduling and execution in the embodiment scenario includes the following steps:

[0075] Step 1: The task scheduling agent receives user instructions.

[0076] The large-scale multi-agent task scheduling framework adopts a two-layer architecture, with the upper layer being a task scheduling agent, such as... Figure 2 As shown, the task scheduling agent handles user interaction functions, feeding user input into a large model, which then performs natural language semantic understanding.

[0077] Step 2: The task scheduling agent analyzes the task content and decomposes it into subtasks.

[0078] The task scheduling agent understands user instructions through large-scale model semantic understanding. It uses prompt word engineering to guide the agent in understanding task requirements, breaking down complex tasks into multi-step, multi-stage sub-tasks, and extracting the necessary instruction information for each sub-task from user input.

[0079] Step 3: Distribute subtasks to the task execution agent.

[0080] Based on the subtask decomposition results of the task scheduling agent and the applicable task scenarios of each task execution agent in the multi-agent framework, the subtasks are distributed to the corresponding task execution agents.

[0081] Step 4: The task execution agent formulates an execution strategy based on the task content and prompts.

[0082] The task-executing agent adopts a "reasoning-action-feedback" execution method, such as... Figure 3 As shown. The current step is the inference phase of the agent. The large model performs inference analysis based on the currently received task description, prompts, existing context, and historical information, and formulates an action strategy. A set of condition-action rules is defined in the prompts of the large model agent to guide the agent in making action plans. The rules can be formalized as follows:

[0083] R i :if p t ∈C i then a t =A i

[0084] Where R i For the i-th rule, p t For the instruction input to the agent at time t, C i The set of conditions for the rule, i.e., the conditions that the instruction input must satisfy, a t For the action of the agent at time t, A i The action corresponding to the rule. All rules form a rule set R = {R1, R2, ..., R...} n The agent matches the current instruction input p. t The action is determined by a specific rule in the rule set.

[0085] Step 5: Based on the action plan made by the agent's inference, determine whether the agent's current action should invoke business tools or retrieval tools. Combined with... Figure 4 Specifically, it includes:

[0086] Step 5.1: When executing a specific subtask, invoke the integrated business tools to perform the task.

[0087] Leveraging the semantic understanding and logical reasoning capabilities of the large model, the intelligent agent enables tool invocation. The integrated business tools support algorithm models, third-party APIs, GIS tools, and visualization tools. By defining a tool registry to describe tool functions and invocation methods, and combining this with prompts from the large model, the agent is guided to generate the correct tool invocation request, completing the entire tool invocation process. An example of a tool registry is shown below:

[0088]

[0089] Step 5.2: After the business tool returns the execution result, update the dynamic memory pool.

[0090] The tool returns the execution results in structured text format. The memory enhancement module updates the corresponding task entity information in the dynamic memory pool based on the execution results for subsequent retrieval.

[0091] Step 5.3: The tool returns the results to the task execution agent.

[0092] The tool encapsulates the results into an unstandardized structured text format and returns it to the agent. The task execution agent parses the results returned by the tool and extracts key information.

[0093] Step 5.4: The task execution agent determines whether the subtask has been completed.

[0094] If the subtask has been completed, proceed to step 6; otherwise, jump to step 5.1, where the agent performs replanning until the task is completed.

[0095] Step 5.5: If context or cross-agent information needs to be queried when performing a task, the retrieval tool is invoked to perform a dynamic memory pool retrieval task.

[0096] When executing subtasks, it is necessary to retrieve task context information or execution result information from other task-executing agents. This information serves as input parameters for the current agent to execute tasks or invoke tools. By invoking a retrieval tool to search the dynamic memory pool, efficient cross-task and cross-agent information retrieval and sharing can be achieved. The retrieval tool's search is implemented through keyword matching of search patterns. By predefining different search patterns, the retrieval tool is specified to search for specific content in the dynamic memory pool. The search pattern is specified in the agent's prompt in structured text form. An example of a search pattern is shown below:

[0097]

[0098]

[0099] Step 5.6: The search results are returned to the task execution agent.

[0100] The retrieval tool returns the entity information retrieved from the dynamic memory pool to the agent in the form of structured text. The task execution agent parses the retrieval results, extracts information, and uses it for subsequent tasks.

[0101] Step 6: The task execution agent reports the subtask execution results to the task scheduling agent.

[0102] Each task execution agent returns the results of its task execution to the upper-level task scheduling agent in the form of structured text, for subsequent parsing.

[0103] Step 7: Determine whether the task has been completed.

[0104] If the task has been completed, proceed to step 8; if not, jump to step 2, and the task planning agent performs replanning and re-execution until the task is completed.

[0105] Step 8: Return the integrated task results to the user.

[0106] The task scheduling agent integrates and optimizes the results received from all task execution agents, and returns the integrated results to the user in a user-friendly natural language format.

[0107] As a specific example, the present invention will be further described below with reference to embodiments:

[0108] Taking flood disaster emergency rescue scenario task scheduling as an example: The user inputs the command through the interactive interface: "Initiate flood disaster emergency response, which requires the completion of disaster area analysis, rescue material transportation route planning and equipment scheduling." Based on the multi-agent task scheduling framework of this invention, the system coordinates the meteorological data analysis agent, the transportation route planning agent, and the rescue equipment deployment agent to realize dynamic task decomposition and execution.

[0109] After the user inputs a command, the task scheduling agent, through large-scale model semantic understanding, identifies the task objective as "flood emergency relief" and extracts key requirements: analyze the weather and hydrological conditions of the disaster area; plan the transportation routes for relief supplies; and dispatch relief equipment and supplies to designated locations. Based on prompt word engineering, the task scheduling agent decomposes the complex task into the following sub-tasks: ① Weather analysis sub-task: analyze the rainfall and water level trends for the next 24 hours; ② Transportation route planning sub-task: plan the optimal transportation route based on real-time traffic and flooding conditions; ③ Equipment and material dispatch sub-task: dispatch relief equipment (such as water pumps and lifeboats) and supplies (such as food and medicine) to the disaster area. The sub-tasks are then assigned based on the professional capabilities of the task execution agent.

[0110] The meteorological data analysis agent calls meteorological APIs (such as the China Meteorological Administration's real-time rainfall data interface) and hydrological monitoring databases to obtain 24-hour rainfall forecasts and river water level data for the affected area. Rainfall amounts (e.g., "rainfall of 150mm in the next 6 hours") and water level thresholds (e.g., "river warning level 5.8m") are stored in a dynamic memory pool with the key "flood_risk_level". Historical flood data for the same period (e.g., "average rainfall for the same period over the past 5 years") is retrieved to help determine the current risk level.

[0111] The transportation route planning agent calls GIS tools (such as the Gaode Map API) to obtain real-time traffic data, and combines it with the `flood_risk_level` parameter in a dynamic memory pool (to avoid high-risk areas). Multiple candidate routes are generated using the A* algorithm.

[0112] The system retrieves historical transport path success rates from the dynamic memory pool (e.g., "Path A has a historical success rate of 92%") and selects the optimal path. It then returns a path solution (e.g., "Use Path A first, Path B as an alternative") and updates the path_status (real-time path status) in the dynamic memory pool.

[0113] The rescue equipment deployment agent calls the material inventory management system API to obtain available rescue equipment (e.g., "50 water pumps, 20 lifeboats") and material distribution information. Based on the path_status returned by the transportation route planning agent, the material allocation strategy is dynamically adjusted (e.g., "prioritize deploying water pumps to the endpoint of path A"). According to real-time demand (e.g., "water levels in a certain area continue to rise"), an emergency replenishment process is triggered, utilizing drone delivery tools to replenish materials.

[0114] The task scheduling agent will integrate all task execution results and return them to the user interface in natural language and visual charts (such as route maps and material distribution maps), and prompt: "It is recommended to immediately activate the red alert response and prioritize ensuring the smooth flow of the rescue channel along route A."

[0115] In summary, this invention addresses the state loss problem in multi-agent systems by introducing a memory enhancement mechanism, significantly improving context awareness during task execution. It also enhances system decision-making quality by enabling cross-task and cross-agent knowledge reuse through a cross-agent retrieval mechanism. Furthermore, the method is compatible with diverse and heterogeneous task requirements, overcoming the limitations of traditional multi-agent systems in terms of task complexity, cross-agent collaboration, and knowledge reuse capabilities. It can be widely applied in intelligent scheduling for complex scenarios such as emergency rescue, industrial automation, and urban public management.

[0116] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for scheduling large-scale multi-agent tasks with memory and retrieval capabilities, characterized in that, The method constructs an intelligent agent collaborative architecture with enhanced cognitive and memory capabilities to achieve efficient decomposition, dynamic scheduling, and intelligent execution of complex tasks.

2. The large-scale multi-agent task scheduling method with memory and retrieval capabilities according to claim 1, characterized in that, The method includes: Based on the task scheduling agent, complex tasks are cognitively analyzed, and sub-tasks are allocated by combining the functional attributes of each task execution agent and integration tools. Each task execution agent completes the execution of specific tasks through the integration tool, returns the task results to the task scheduling agent for integration, and decides to execute the next sub-task or end the task based on the task situation. A dynamic memory pool is used to maintain the task context information and task entity information database; Cross-agent retrieval enables multi-dimensional information sharing across agents and task nodes.

3. A large-scale multi-agent task scheduling system with memory and retrieval capabilities, implementing the method of any one of claims 1 to 2, characterized in that, The system includes a multi-agent task scheduling framework, a memory enhancement module, and a cross-agent retrieval module; The multi-agent task scheduling framework, as the core module for receiving user requests and completing task decomposition, task scheduling, and task execution, adopts a cognitive architecture based on a large model and includes a task decomposer, a dynamic scheduler, and a state coordinator. The memory enhancement module is used to store task entity parameters during task execution. It adopts a key-value storage structure and uses a dynamic memory pool to save the specified parameters in the task scheduling and execution threads, and can perform local persistent storage as needed. The cross-agent retrieval module is used for multiple agents to retrieve task parameters from the short-term memory pool, and designs and integrates key-value-based retrieval tools to achieve cross-agent state and information sharing.

4. The large-scale multi-agent task scheduling system with memory and retrieval capabilities according to claim 3, characterized in that, The retrieval module can be invoked by the intelligent agent as an external tool.

5. The large-scale multi-agent task scheduling system with memory and retrieval capabilities according to claim 4, characterized in that, The multi-agent task scheduling framework includes a task scheduling agent and a task execution agent; and the multi-agent task scheduling framework adopts a two-layer architecture, wherein the upper layer is the task scheduling agent, which is used to receive task requests, decompose task logic and dynamically allocate sub-tasks, and the lower layer consists of several task execution agents, including several functionally specialized agents. Each agent integrates cognitive reasoning functions and tool calling interfaces and is responsible for executing specific tasks.

6. The large-scale multi-agent task scheduling system with memory and retrieval capabilities according to claim 5, characterized in that, The task scheduling agent is used for task analysis and scheduling, driven by a large model, and has the ability to decompose and route tasks. In the task decomposition and scheduling phase, after receiving user input, the task scheduling agent decomposes complex tasks into multiple sub-tasks based on Prompt and thought chains, and dynamically allocates them according to the professional capabilities of each task execution agent. During task execution, it maintains the global state, tracks the execution progress of each task execution agent, and handles the dependencies between tasks. After the task execution is completed, it integrates the output results of each task execution agent and returns them to the user. The task execution agent is used to execute specific tasks. It adopts a reasoning-then-action approach, where a large model generates logical judgments based on the current state, formulates action strategies, and then executes specific operations by calling integrated tools.

7. The large-scale multi-agent task scheduling system with memory and retrieval capabilities according to claim 6, characterized in that, The task scheduling agent adopts a two-layer architecture. The upper-layer task scheduling agent is responsible for task decomposition, sub-task allocation, and global state maintenance. The lower-layer task execution agent is configured according to functional specialization and integrates cognitive reasoning functions and tool call interfaces.

8. The large-scale multi-agent task scheduling system with memory and retrieval capabilities according to claim 7, characterized in that, The task execution agent adopts a three-stage execution method: the reasoning stage, which generates an action strategy based on the current task description, prompts, context and historical information; and the action stage, which calls the business tools defined in the tool registry or the cross-agent retrieval module. During the feedback phase, the dynamic memory pool is updated based on the tool or search results, and it is determined whether the subtask has been completed.

9. The large-scale multi-agent task scheduling system with memory and retrieval capabilities according to claim 8, characterized in that, The cross-agent retrieval module supports preset retrieval modes. Based on task information requirements and task dependencies, it matches different retrieval modes through keyword matching. It supports multi-information type retrieval of current task information and historical task information, as well as multi-scenario retrieval across tasks and across agents.

10. A task scheduling method for implementing the system of claim 9, characterized in that, The task scheduling method includes: The task scheduling agent of the multi-agent task scheduling framework receives user instructions, decomposes complex tasks into multiple sub-tasks through large model semantic understanding, and dynamically allocates sub-tasks to the corresponding task execution agents according to the professional capabilities of each task execution agent. The task execution agent of the multi-agent task scheduling framework performs reasoning analysis through an integrated large model, formulates action strategies, and calls business tools or retrieval tools to execute tasks. The memory enhancement module records entity parameters in real time during task execution using a key-value storage structure, forming a dynamic memory pool; The cross-agent retrieval module retrieves historical information across tasks or across agents from a dynamic memory pool based on a predefined retrieval pattern, supporting decision-making by the task-executing agent. The task execution agent returns the execution results to the task scheduling agent, which then integrates the results of all subtasks and generates user feedback.

Images