Task processing system and method based on distributed multi-agent

Abstract

The invention provides a distributed multi-agent-based task processing system and method, and relates to the technical field of artificial intelligence. A task processing system based on distributed multi-agent comprises a plurality of agents used for processing agent scheduling information; the planning reasoning module is used for selecting at least one agent from a plurality of agents as a target agent based on a pre-stored agent information dictionary according to a first task instruction input by a user, and generating first agent scheduling information according to the first task instruction and the target agent, the agent information dictionary comprises names and corresponding functions of a plurality of agents; and the message management module is used for sending the first agent scheduling information to the target agent, so that the target agent executes the first task contained in the first agent scheduling information. The agents are scheduled based on the agent scheduling information, so that multi-agent collaboration is realized, and the maintenance cost is low and reliability is realized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a task processing system and method based on distributed multi-agent systems. Background Technology

[0002] Currently, large-scale model-driven multi-agent systems are a popular research area, with numerous publicly available multi-agent frameworks and systems proposed. Multi-agent systems can accomplish specific tasks through the cooperation of multiple agents.

[0003] Typically, task-oriented multi-agent systems construct workflows for specific tasks, assigning different agents to collaborate. They are highly specialized and refined, but lack versatility and scalability. To address this issue, some multi-agent systems are designed to leverage the reasoning capabilities of large models, improving versatility and scalability, thus facilitating the definition and integration of various agents. However, these systems often rely heavily on the reasoning capabilities of large models, resulting in high maintenance costs and low reliability. Summary of the Invention

[0004] Based on this, this application provides a distributed multi-agent task processing system and method to achieve low-maintenance and reliable multi-agent task processing.

[0005] According to one aspect of this application, a task processing system based on distributed multi-agent systems is proposed, comprising: multiple agents for processing agent scheduling information; a planning and reasoning module for selecting at least one agent as a target agent from the multiple agents based on a pre-stored agent information dictionary, according to a first task instruction input by a user, and generating first agent scheduling information based on the first task instruction and the target agent, wherein the agent information dictionary includes the names of the multiple agents and their corresponding functions; and a message management module for sending the first agent scheduling information to the target agent, thereby enabling the target agent to execute the first task contained in the first agent scheduling information.

[0006] According to some embodiments, the target intelligent agent is used to: call a preset large model to analyze a first task and obtain analysis results; if the analysis results indicate that the first task cannot be directly executed, generate knowledge keywords based on the first task; search in a pre-stored knowledge rule base based on the knowledge keywords to obtain a target knowledge dictionary; and concatenate the target knowledge dictionary with the first task and update the first task using the concatenation result.

[0007] According to some embodiments, the target agent is also used to: directly execute the first task if the analysis result indicates that the first task can be directly executed.

[0008] According to some embodiments, the target intelligent agent is further configured to: match the function of the target intelligent agent with the first task; if the matching result indicates that the target intelligent agent cannot perform the first task independently, select at least one intelligent agent as an auxiliary intelligent agent from multiple intelligent agents based on an intelligent agent information dictionary; obtain multiple first task steps based on the auxiliary intelligent agent, its function, and the first task, thereby obtaining a first task execution plan; and store the first task execution plan in a memory bank, wherein the memory bank is a local memory bank unique to the target intelligent agent and / or a public memory bank shared by the auxiliary intelligent agent and the target intelligent agent.

[0009] The first task steps are executed sequentially by the target agent and the auxiliary agent until the first task execution plan is completed.

[0010] According to some embodiments, the target intelligent agent is further configured to: S1: take the first step in the first task execution scheme as the current first task step; S2: store the current task execution state in a memory bank, wherein the current task execution state includes at least one of the following: the execution progress of the current first task step, the execution progress of the first task execution scheme, the execution details of the current first task step, and the current instruction reception status; S3: when the executing entity of the current first task step is an auxiliary intelligent agent, based on the target intelligent agent, generate second intelligent agent scheduling information according to the current first task step and the auxiliary intelligent agent, send the second intelligent agent scheduling information to the corresponding auxiliary intelligent agent, and jump to step [step name missing]. S5; S4: If the target agent is the executing entity of the current first task step, execute the current first task step, obtain the processing result, and jump to step S7; S5: Obtain the third agent scheduling information sent by the auxiliary agent, wherein the third agent scheduling information is generated by the auxiliary agent after processing the second agent scheduling information; S6: Obtain the processing result according to the third agent scheduling information; S7: Update the current task execution status and / or the next step of the current first task step using the processing result; S8: Take the next step of the current first task step as the current first task step, and repeat steps S2-S8 until the first task execution plan is completed.

[0011] According to some embodiments, the target agent is also used to: extract memory keywords based on the processing results; search the memory bank based on the memory keywords to obtain target memory information; concatenate the target memory information with the processing results and update the processing results using the concatenated results; and update the current task execution status and / or the next first task step using the updated processing results.

[0012] According to some embodiments, the auxiliary intelligent agent is used to: obtain a corresponding second task based on the second intelligent agent scheduling information; execute the second task and obtain a processing result; generate third intelligent agent scheduling information based on the processing result of the auxiliary intelligent agent; and send the third intelligent agent scheduling information to the target intelligent agent.

[0013] According to some embodiments, the auxiliary agent is also used to: match the function of the auxiliary agent with the second task; if the matching result is that the auxiliary agent cannot perform the second task alone, select at least one agent from multiple agents as the second auxiliary agent based on the agent information dictionary; obtain multiple second task steps based on the second auxiliary agent, its function, and the second task, thereby obtaining a second task execution plan; store the second task execution plan in a memory bank; and execute the second task steps sequentially based on the auxiliary agent and the second auxiliary agent until the second task execution plan is completed.

[0014] According to one aspect of this application, a task processing method based on distributed multi-agent systems includes: using a planning and reasoning module, selecting at least one agent as a target agent from multiple agents based on a pre-stored agent information dictionary, according to a first task instruction input by a user; generating first agent scheduling information based on the first task instruction and the target agent; wherein the agent information dictionary includes the names of multiple agents and their corresponding functions, and the multiple agents are used to process the agent scheduling information; using a message management module, sending the first agent scheduling information to the target agent; and using the target agent to execute the task contained in the first agent scheduling information.

[0015] According to one aspect of this application, an electronic device is provided, comprising: one or more processors; a storage device for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the method as described above.

[0016] According to one aspect of this application, a computer-readable medium is provided having a computer program stored thereon that, when executed by a processor, implements the method described above.

[0017] Through the embodiments provided in this application, the planning and inference module selects a target intelligent agent to execute the instruction based on the user's input, generates first intelligent agent scheduling information, and sends the first intelligent agent scheduling information to the target intelligent agent through the message management module, causing the target intelligent agent to execute the first task contained in the first intelligent agent scheduling information. This application schedules intelligent agents based on intelligent agent scheduling information, thereby achieving multi-agent collaboration. It has low maintenance costs and high reliability, eliminates the need for pre-construction of intelligent agent workflows, and solves the problem in existing technologies where high dependence on the inference capabilities of large models leads to high maintenance costs and low reliability in task processing systems. Attached Figure Description

[0018] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings, without exceeding the scope of protection claimed by this application.

[0020] Figure 1 A block diagram of a distributed multi-agent task processing system provided in an embodiment of this application;

[0021] Figure 2 This is one of the flowcharts for the steps performed by the target intelligent agent provided in the embodiments of this application;

[0022] Figure 3 A second flowchart illustrating the steps performed by the target intelligent agent in the embodiments of this application;

[0023] Figure 4 The third flowchart illustrates the steps performed by the target intelligent agent in the embodiments of this application.

[0024] Figure 5 The fourth flowchart illustrates the steps performed by the target intelligent agent in the embodiments of this application.

[0025] Figure 6 One of the flowcharts for the steps performed by the auxiliary intelligent agent provided in the embodiments of this application;

[0026] Figure 7 A second flowchart illustrating the steps performed by the auxiliary intelligent agent provided in this application embodiment;

[0027] Figure 8 A flowchart of a task processing method based on distributed multi-agent provided in an embodiment of this application;

[0028] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0031] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0032] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0033] It should be understood that although the terms first, second, third, etc., may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of this application. As used herein, the term "and / or" includes all combinations of any one and more of the associated listed items.

[0034] For specific implementation details, please refer to the following examples.

[0035] Figure 1 This is a block diagram of a distributed multi-agent task processing system provided in an embodiment of this application. Figure 1 As shown, the system includes multiple intelligent agents 110, a planning and reasoning module 120, and a message management module 130.

[0036] Multiple intelligent agents 110 are used to process intelligent agent scheduling information. The multiple intelligent agents 110 can be understood as an intelligent agent library, including multiple intelligent agents stored in a distributed manner.

[0037] It should be explained that the various intelligent agents in the multiple intelligent agents 110 are connected through a distributed network, and each intelligent agent is completely autonomous and can communicate and cooperate directly.

[0038] According to the example embodiment, the agents in the multiple agents 110 implement distributed agent collaboration based on Ray. Ray is an open-source distributed computing framework that mainly implements stateless remote methods and stateful remote classes. Remote methods and remote classes refer to methods and classes that run outside the main process. This application utilizes Ray's distributed computing capabilities to implement each module and each agent of the multi-agent framework as a remote class, accessing each module and agent through references to the remote classes. This achieves concurrent collaboration of agents, with each agent running in an independent process.

[0039] The planning and reasoning module 120 is used to select at least one intelligent agent as the target intelligent agent from multiple intelligent agents 110 based on the first task instruction input by the user and a pre-stored intelligent agent information dictionary, and to generate first intelligent agent scheduling information based on the first task instruction and the target intelligent agent. The intelligent agent information dictionary includes the names of multiple intelligent agents 110 and their corresponding functions.

[0040] The planning and reasoning module 120 can acquire a first task instruction input by the user. In some embodiments, an interactive interface is provided for the user to input the first task instruction. Furthermore, to enable the planning and reasoning module 120 to acquire the first task instruction, the interactive interface is communicatively connected to the planning and reasoning module 120. Alternatively, the interactive interface can be communicatively connected to the message management module 130, which then sends the acquired first task instruction to the planning and reasoning module 120.

[0041] The planning and reasoning module 120 selects the most suitable agent to execute the task based on the acquired first task instruction. During the selection process, based on the agent information dictionary stored in the planning and reasoning module 120, and referring to the names and corresponding functions of multiple agents 110 stored therein, the module matches the content of the first task instruction and selects at least one agent among the multiple agents 110 that best matches the first task instruction as the target agent.

[0042] Furthermore, the planning and reasoning module 120 can also call a pre-set large model and combine it with the agent information dictionary to select the target agent.

[0043] It should be noted that, in addition to being stored in the planning and reasoning module 120, the agent information dictionary is also stored in each of the multiple agents 110.

[0044] After selecting the target agent, the planning and reasoning module 120 generates first agent scheduling information based on the first task instruction and the target agent. According to the example embodiment, the first agent scheduling information includes the sender (i.e., the planning and reasoning module 120), the message content (i.e., the first task instruction), and the receiver (i.e., the target agent).

[0045] The message management module 130 is used to send the scheduling information of the first intelligent agent to the target intelligent agent, so that the target intelligent agent executes the first task contained in the scheduling information of the first intelligent agent.

[0046] The scheduling information of the first intelligent agent is sent to the target intelligent agent through the message management module 130. The target intelligent agent parses the received scheduling information of the first intelligent agent and obtains the first task to be executed.

[0047] The target agent internally determines the execution method for the first task, which can be either independent execution or collaborative execution with other agents. If independent execution is required, the target agent directly executes the task, and the process ends upon completion. If collaboration with other agents is needed, the target agent communicates directly with the agents to be invoked, calling them for collaboration. The invocation method involves generating new agent scheduling information and sending it to other agents. This process is a chained call, where agents communicate with each other via messages. Since each agent runs as an independent process, parallel scheduling, i.e., distributed collaboration, is possible.

[0048] This application uses a planning and reasoning module 120 to select a target agent to execute a user-input instruction, generates first agent scheduling information, and sends this information to the target agent via a message management module 130, causing the target agent to execute the first task contained in the scheduling information. This application schedules agents based on agent scheduling information, thereby achieving multi-agent collaboration. It boasts low maintenance costs and high reliability, eliminates the need for pre-constructing agent workflows, and solves the problem of high maintenance costs and low reliability in existing technologies due to high reliance on the reasoning capabilities of large models.

[0049] Based on the above embodiments, in order to provide a more detailed description of the task processing system based on distributed multi-agent provided in this application, this application provides a specific embodiment to implement the above-mentioned task processing system based on distributed multi-agent, but this does not represent a limitation on this application.

[0050] In this embodiment, the agents in the multiple agents 110 are defined through Python classes. The defined agents need to include the following attributes: name, introduction, agent_info (agent information dictionary), and message_handle (message handle).

[0051] Here, name is a globally unique name for the agent, corresponding one-to-one with the agent, and is used to identify the agent.

[0052] The introduction is a functional description of the agent, using natural language to describe the agent's functions. It is used to define the agent's functions and to retrieve the corresponding agent based on those functions.

[0053] `agent_info` is an agent information dictionary, containing information about all agents, with `name` as the key and `introduction` as the value. `agent_info` stores the introduction of all agents; when an agent needs to invoke a function from another agent, it can retrieve the corresponding agent through `agent_info`.

[0054] message_handle is a reference to an instantiated object of the message management module 130, used to call its process_message method to send messages.

[0055] Furthermore, to enable communication between agents, the agent also includes a `send_message` method. The `send_message` method is used to send messages to other agents. Parameters include `sender_name` (sender's name), `message_content` (message content), and `target_name` (receiver's name), where `name` is the agent's attribute `name`. Through this method, an agent can send a message to another agent. Agents send messages by calling methods of the instantiated object of the message management module 130.

[0056] Since all agents are scheduled by sending messages, a message management module 130 is constructed to implement communication with the agents. The message management module 130 is implemented by constructing a message class. During agent registration, the information of each agent and its remote reference object are saved in the message class. The message management module 130 includes the `agent_ref_info` attribute and the `process_message` method.

[0057] The `agent_ref_info` object stores references to all agents in a dictionary format, where the key is the agent's name and the value is a reference object of the agent. The `process_message` method defines the message format and forwards the message to the agent. The message format is the method's parameter, consisting of three parameters: `sender_name` (sender's name), `message_content` (message content), and `target_name` (receiver's name). When agent A wants to send a message to agent B, the actual process is as follows: agent A calls the `process_message` method of `message_handle` (the method described here), given the three parameters. Then, the message class finds the target agent in `agent_ref_info` based on the `target_name` parameter, obtains a reference to the target agent, and calls the agent's `process_message` method through the reference. This completes the process of agent A sending a message to agent B.

[0058] Furthermore, to enable the agent to process agent scheduling information, the agent also includes a `process_message` method. `process_message` is the method for the agent to process messages. When the agent receives agent scheduling information, it will execute the `process_message` method, which defines the response logic to the message. The agent scheduling information processed by the agent includes not only the first agent scheduling information generated by the planning and inference module 120, but also the second and third agent scheduling information used for direct communication between agents.

[0059] After defining the agents, they need to be manually registered to form multiple agents 110. The registration process is as follows: instantiate all agents, then save the name and introduction of each agent to the agent_info of all other agents. At the same time, save the name and introduction of all agents to the planning and reasoning module 120. After registration, the planning and reasoning module 120 and the agents can obtain the information of all agents and select the appropriate agent for scheduling based on this information.

[0060] According to some embodiments, refer to Figure 2 The target agent is used to execute steps S210-S240.

[0061] In step S210, a preset large model is invoked to analyze the first task and obtain the analysis results.

[0062] When the target agent receives the scheduling information from the first agent, it extracts the first task from it, then calls the preset large model and uses its reasoning ability to analyze the first task to determine whether there is enough information to execute the task. If so, the analysis result is that the first task can be executed directly; if not, the analysis result is that the first task cannot be executed directly.

[0063] In step S220, if the analysis result indicates that the first task cannot be directly executed, knowledge keywords are generated based on the first task.

[0064] When the analysis results indicate that the first task cannot be executed directly, the pre-set large model is enhanced based on knowledge. The main method is to realize the reasoning process of the large model agent combined with the knowledge rule base.

[0065] It is important to emphasize that the knowledge rule base needs to have knowledge query functionality, and can be in formats such as knowledge graphs, databases, or JSON dictionaries; this application does not impose any restrictions on this. The method for constructing the knowledge rule base can be chosen based on the actual situation; this application does not impose any restrictions on this either.

[0066] In order to perform knowledge queries from the knowledge rule base, it is first necessary to generate knowledge keywords based on the first task as the keywords for the query.

[0067] In step S230, a target knowledge dictionary is obtained by searching a pre-stored knowledge rule base based on knowledge keywords.

[0068] Based on knowledge keywords, similar information is queried in the knowledge rule base, the retrieved information is recorded as the target knowledge dictionary, and returned to the target intelligent agent.

[0069] In some embodiments, the first task is represented by s, and the target knowledge dictionary is represented by dictionary d. The specific structure of s and d can be a key-value pair representation, that is: {key1:value1,key2:value2,...keyN,valueN}.

[0070] In step S240, the target knowledge dictionary is concatenated with the first task, and the first task is updated using the concatenation result.

[0071] The target knowledge dictionary is concatenated with the first task, and the concatenation result is used as the new first task.

[0072] Furthermore, after step S240, the process also includes repeating steps S210-S240, that is, after the target agent obtains the new first task, it continues to determine whether additional information is needed based on the new first task, until the analysis result indicates that the first task can be executed directly.

[0073] In this way, the knowledge rule base and large model-driven multi-agent system are combined to solve the problems of agent reasoning illusion and unreliability.

[0074] According to some embodiments, the target agent is also used to perform step S221.

[0075] In step S221, if the analysis result indicates that the first task can be executed directly, the first task is executed directly.

[0076] If the analysis results indicate that the first task can be executed directly, then the target agent is used to execute the first task.

[0077] During execution, the large model is invoked to reason about the first task, which yields the task steps for executing the first task. Based on the target agent, the task execution result of the first task is obtained by executing the task steps step by step.

[0078] According to some embodiments, refer to Figure 3 In the process of the target intelligent agent executing step S221, it can be specifically implemented through steps S310-S350.

[0079] In step S310, the function of the target agent is matched with the first task.

[0080] Based on the agent information dictionary stored in the target agent, the functions of the target agent are matched with the first task, and the matching result is used to determine whether the target agent can execute the first task independently.

[0081] According to the example embodiment, during the matching process, if the target agent's functionality completely covers the keywords of the first task, the matching result is that the target agent can execute the first task independently; otherwise, the matching result is that the target agent cannot execute the first task independently.

[0082] In step S320, if the matching result indicates that the target agent cannot perform the first task independently, at least one agent is selected from multiple agents 110 as an auxiliary agent based on the agent information dictionary.

[0083] When the matching result indicates that the target agent cannot perform the first task alone, the target agent calls other agents among multiple agents 110 to assist it in performing the first task.

[0084] The process of the target agent selecting and calling an agent is achieved by the target agent using an agent information dictionary stored therein, selecting at least one agent from multiple agents 110 as an auxiliary agent.

[0085] The process of the target agent selecting the auxiliary agent is similar to the process of the planning and reasoning module 120 selecting the target agent.

[0086] The target agent can also invoke a large model to reason about the first task and obtain the task steps to execute the first task. For each task step, it uses the agent information dictionary stored in the target agent to select a suitable auxiliary agent for each task step.

[0087] In step S330, based on the auxiliary intelligent agent and its functions, as well as the first task, multiple first task steps are obtained, thereby obtaining the first task execution scheme.

[0088] Based on the selected auxiliary agent, the large model is invoked for reasoning to obtain multiple first task steps, including the target agent and / or the auxiliary agent, which are then combined to form the first task execution scheme.

[0089] Furthermore, in some embodiments, steps S320 and S330 can be combined as follows: if the matching result indicates that the target agent cannot perform the first task alone, the large model is invoked to reason about the first task, and combined with the agent information dictionary, multiple first task steps including the target agent and / or the auxiliary agent are directly derived and combined to form the first task execution scheme.

[0090] In step S340, the first task execution plan is stored in a memory bank, wherein the memory bank is a local memory bank unique to the target intelligent agent and / or a public memory bank shared by the auxiliary intelligent agent and the target intelligent agent.

[0091] When intelligent agents handle complex tasks, they require the assistance of large models. Although large models have powerful logical reasoning capabilities, they consume a lot of computing power and have limited input instruction length. Therefore, although large models can engage in continuous dialogue, they are limited by the input length and cannot withstand the surge in input information when performing complex reasoning.

[0092] Based on this, this application sets up a memory to store historical information during the execution of the first task.

[0093] The memory bank includes at least the target agent's unique local memory bank and / or the shared memory bank between the auxiliary agent and the target agent.

[0094] It is important to emphasize that the essential purpose of setting up a memory bank is to store information that may need to be shared with other intelligent agents during the execution of the first task, so as to facilitate retrieval. Therefore, the local memory bank can be configured for all intelligent agents, or only for intelligent agents that need to perform complex reasoning. Similarly, the public memory bank can be configured for all intelligent agents to share, or it can be configured for the target intelligent agent and the auxiliary intelligent agents that need to retrieve information from the public memory bank.

[0095] According to the example embodiment, each agent maintains a local memory bank and collectively maintains a common memory bank. Building upon this, and based on the specific embodiment described above, to implement the agent's memory bank function, the agent further includes the `load_memory` method, the `save_memory` method, the `load_local_memory` method, and the `save_local_memory` method.

[0096] The `load_memory` method is used to write information to the public memory. The `save_memory` method is used to read specific information from the public memory. The `load_local_memory` method is used to write information to the local memory. The `save_local_memory` method is used to read specific information from the local memory.

[0097] In this embodiment, the first task execution plan is stored in the target agent's unique local memory and / or in the shared memory of the auxiliary agent and the target agent.

[0098] In step S350, the first task steps are executed sequentially based on the target intelligent agent and the auxiliary intelligent agent until the first task execution plan is completed.

[0099] According to some embodiments, refer to Figure 4 In the process of the target intelligent agent executing step S350, it can be specifically implemented through steps S1-S8.

[0100] S1: Take the first step in the first task execution plan as the current first task step.

[0101] Each first task step of the first task execution plan is executed sequentially, and each first task step is recorded as the current first task step during execution.

[0102] S2: Store the current task execution status in the memory bank, wherein the current task execution status includes at least one of the following: the execution progress of the current first task step, the execution progress of the first task execution plan, the execution details of the current first task step, and the current instruction reception status.

[0103] For example, during the first iteration of steps S1-S8, the target agent stores "the first step in the first task execution plan is being executed" as the current task execution state in the memory bank.

[0104] S3: If the executing entity of the current first task step is an auxiliary intelligent agent, based on the target intelligent agent, generate second intelligent agent scheduling information according to the current first task step and the auxiliary intelligent agent, send the second intelligent agent scheduling information to the corresponding auxiliary intelligent agent, and jump to step S5.

[0105] If the executing entity of the current first task step is an auxiliary intelligent agent, the target intelligent agent generates second intelligent agent scheduling information to communicate with the auxiliary intelligent agent and instruct it to execute the tasks included in the current first task step.

[0106] According to the example embodiment, the second agent scheduling information includes the sender (i.e., the target agent), the message content (i.e., the current first task step), and the receiver (i.e., the auxiliary agent).

[0107] Then, the generated second agent scheduling information is sent to the corresponding auxiliary agent so that the auxiliary agent can execute the tasks included in the current first task step, and then the process jumps to step S5.

[0108] It should be noted that when the auxiliary agent performs the tasks included in the current first task step, the specific execution process can refer to the steps of the target agent performing the first task, and this application will not elaborate on this.

[0109] S4: If the target intelligent agent is the subject of the current first task step, execute the current first task step, obtain the processing result, and jump to step S7.

[0110] If the target intelligent agent is the executing entity of the current first task step, the target intelligent agent directly executes the current first task step, obtains the processing result, and then jumps to step S7.

[0111] During the execution of the first task step by the target intelligent agent, the implementation can be based on the functions of the target intelligent agent or by calling tools in the tool library. This application does not impose any restrictions on this.

[0112] Furthermore, it is important to emphasize that the difference between an intelligent agent and a tool is that an intelligent agent is a stateful class capable of performing specific tasks, while a tool is a stateless method that implements or performs a certain function. The definition of a tool can refer to the definition method of an intelligent agent in the above embodiments, including the aforementioned attributes and methods.

[0113] S5: Obtain the third agent scheduling information sent by the auxiliary agent, wherein the third agent scheduling information is generated by the auxiliary agent after processing the second agent scheduling information.

[0114] After the auxiliary agent completes the first task step, it generates third agent scheduling information to communicate with the target agent and send the processing results to the target agent.

[0115] S6: Obtain the processing result based on the scheduling information of the third intelligent agent.

[0116] S7: Update the current task execution status and / or the next step of the current first task step using the processing results.

[0117] According to the example embodiment, the processing result includes the execution status and / or execution result of the current first task step. If the subsequent first task step is related to the current first task step, the received processing result is used to update the subsequent first task step.

[0118] At the same time, update the current task execution status based on the processing results.

[0119] S8: Take the next step of the current first task step as the current first task step, and repeat steps S2-S8 until the first task execution plan is completed.

[0120] According to some embodiments, refer to Figure 5 When the target agent executes step S7, it is specifically implemented by steps S510-S540.

[0121] In step S510, memory keywords are extracted based on the processing results.

[0122] In step S520, the target memory information is obtained by searching the memory bank based on the memory keywords.

[0123] In step S530, the target memory information is concatenated with the processing result, and the processing result is updated using the concatenated result.

[0124] In step S540, the current task execution status and / or the next first task step are updated using the updated processing results.

[0125] For example, after obtaining the processing result, the target memory information retrieved from the memory bank based on the processing result will be concatenated with the processing result. The concatenated result will update the processing result, and the updated processing result will be used to update the current task execution status and / or the next first task step.

[0126] During the update process, the target large model can be invoked to infer the updated processing results.

[0127] It is important to emphasize that when an agent receives a task instruction, it does not add all historical information to the task instruction by default. Instead, the agent retrieves relevant information from the local memory bank and the public memory bank according to the instruction, puts it into the instruction, and provides it together for the large model to perform inference. In this way, it not only saves the inference cost of the large model, but also realizes the management and utilization of memory.

[0128] According to some embodiments, refer to Figure 6 The auxiliary intelligent agent is used to execute steps S610-S640.

[0129] In step S610, the corresponding second task is obtained according to the second agent scheduling information.

[0130] According to the example embodiment, after the auxiliary agent receives the scheduling information of the second agent, since the scheduling information of the second agent includes the sender (i.e. the target agent), the message content (e.g. the current first task step), and the receiver (i.e. the auxiliary agent), it obtains the second task from the message content.

[0131] In step S620, the second task is performed to obtain the processing result.

[0132] When the auxiliary agent performs the second task, the specific execution process can refer to the steps of the target agent performing the first task, and this application will not elaborate on this.

[0133] In step S630, scheduling information for the third agent is generated based on the processing results of the auxiliary agent.

[0134] After the auxiliary agent completes the second task, it generates scheduling information for the third agent to communicate with the target agent and send the processing results to the target agent.

[0135] In step S640, the scheduling information of the third agent is sent to the target agent.

[0136] According to the example embodiment, the third agent scheduling information includes the sender (i.e., the auxiliary agent), the message content (i.e., the processing result), and the receiver (i.e., the target agent).

[0137] According to some embodiments, refer to Figure 7 In the process of the auxiliary intelligent agent performing step S620, it can be specifically implemented through steps S710-S750.

[0138] In step S710, the functions of the auxiliary agent and the second task are matched.

[0139] In step S720, if the matching result indicates that the auxiliary agent cannot perform the second task independently, at least one agent is selected from multiple agents as the second auxiliary agent based on the agent information dictionary.

[0140] In step S730, based on the second auxiliary intelligent agent and its functions, as well as the second task, multiple second task steps are obtained, thereby obtaining a second task execution scheme.

[0141] In step S740, the second task execution plan is stored in the memory bank.

[0142] In step S750, the second task steps are executed sequentially based on the auxiliary agent and the second auxiliary agent until the second task execution plan is completed.

[0143] The steps for the auxiliary agent to perform the second task are similar to steps S310-S350 for the target agent to perform the first task, and will not be described in detail here.

[0144] This application designs and implements a knowledge- and large-model-driven scalable distributed multi-agent framework. It achieves multi-agent collaboration through distributed multi-agent scheduling, realizes dynamic workflow through chained scheduling of large models, and solves the problems of unreliability and weak professional capabilities of large models through the application of knowledge. This multi-agent framework can be used to solve complex planning and decision-making problems.

[0145] Furthermore, in order to provide a more detailed description of the task processing system based on distributed multi-agent provided in this application, a specific embodiment is given below.

[0146] This embodiment constructs a multi-agent collaborative system for intelligent search and rescue in a virtual simulation environment, and implements a search agent, a path generation agent, and a transportation agent. The agent is integrated into the simulation environment, and its task is to locate the malfunctioning robot A and bring it back.

[0147] This embodiment also constructs a knowledge rule base to store search and rescue-related knowledge and rules. For example:

[0148] 1) Search and rescue steps: 1-1) Search for the target location; 1-2) Plan the route; 1-3) Bring the target back according to the route.

[0149] 2) Maneuvering method: In complex environments, it is necessary to avoid terrain such as ditches and leaves, and choose smooth and flat ground.

[0150] In this embodiment, the first task instruction input by the user is "bring the robot back to the current area".

[0151] The agent information dictionary stores information about four agents in dictionary format, as follows:

[0152] {

[0153] 'Search and rescue agent': 'Responsible for searching and rescuing designated targets within a specified area'.

[0154] 'Search agent': 'Equipped with sensors such as radar and vision, responsible for searching for designated targets within the execution area'.

[0155] 'Path planning agent': 'Receives visual information about the terrain and is responsible for planning the action path'.

[0156] 'Transportation agent': 'Possesses transportation capabilities and can move objects such as robots and robot dogs'.

[0157] }

[0158] The planning and reasoning module 120, based on a pre-stored agent information dictionary, utilizes the reasoning capabilities of a large model to select a search and rescue agent as the target agent according to the input first task instruction, so as to execute the first task contained in the first task instruction.

[0159] After the planning and reasoning module 120 completes the selection of the target intelligent agent, it generates the first intelligent agent scheduling information: (Planning and reasoning module, search in the current area to bring the robot back, search and rescue intelligent agent).

[0160] The planning and reasoning module 120 sends the first agent scheduling information (planning and reasoning module, bring the robot back in the current area, search and rescue agent) to the search and rescue agent through the message management module 130.

[0161] The search and rescue agent receives the scheduling information from the first agent and calls the `process_message` method to process the scheduling information. From this, the first task is extracted: "Bring the robot back to the current area."

[0162] The search and rescue agent analyzes the first task, "bring the robot back to the current area," and the analysis results indicate that the first task cannot be executed directly. Based on the first task, the agent generates the knowledge keyword "search and rescue."

[0163] The search and rescue agent retrieves relevant knowledge and rules from the knowledge rule base based on the knowledge keyword "search and rescue," resulting in the following K (i.e., the target knowledge dictionary):

[0164] {

[0165] Search and rescue steps: 1) Search for the target location; 2) Plan the route; 3) Bring the target back according to the route.

[0166] }

[0167] The search and rescue agent combines the target knowledge dictionary K with the first task "bring the robot back to the current area" to form a new first task.

[0168] Based on the functions of the search and rescue agent, the search and rescue agent cannot independently execute a new primary task.

[0169] Based on the new first task, the large model is invoked for reasoning, and combined with the agent information dictionary, the steps of the first task are as follows: 1) Dispatch the search agent to find the target location and record the terrain information; 2) Invoke the path planning agent to generate the action route; 3) Invoke the transportation agent to carry out the transportation task according to the action route. Steps 1), 2), and 3) constitute the execution plan of the first task.

[0170] The search and rescue agent stores the first mission execution plan in its local memory.

[0171] The search and rescue agent began executing the mission according to the first mission step of the first mission execution plan.

[0172] Step 1) dispatching the search agent to find the target location and record the terrain information is the current first task step.

[0173] The search and rescue agent stores the current mission execution status "Executing the first step of the search and rescue mission" in its local memory.

[0174] Since the current first task step is not executed by the search and rescue intelligent agent, the second intelligent agent scheduling information is generated based on the current first task step and the relevant auxiliary intelligent agent (i.e., the search intelligent agent): (search and rescue intelligent agent, explore the terrain and search for the robot in the area, search intelligent agent), and the second intelligent agent scheduling information: (search and rescue intelligent agent, explore the terrain and search for the robot in the area, search intelligent agent) is sent to the search intelligent agent.

[0175] After the message is sent, the search and rescue intelligent agent will cease operation.

[0176] The search agent receives scheduling information from the second agent: (Search and rescue agent, explore the terrain and search for robots in the area, search agent), extracts the second task "explore the terrain and search for robots in the area", and calls the large model to analyze and understand the message.

[0177] The search agent obtains a search and rescue plan: 1.1) Generate a search route; 1.2) Collect visual information and identify targets according to the search route. Steps 1.1) and 1.2) constitute the second task execution plan.

[0178] The search agent saves the second task execution plan (i.e., the search and rescue plan) to its local memory.

[0179] The search agent saves the information "Received instructions from the search and rescue agent: Explore the terrain and search for the robot in the area" to its local memory.

[0180] Step 1.1) generating the search route is taken as the current second task step of the search agent.

[0181] The search agent stores the current task execution status, "Current status: executing search task. Task steps: 1. Generate route; 2. Collect visual information and identify targets according to the route. Calling the path planning agent to plan the route," into its local memory.

[0182] The search agent's functions are matched with the tasks in the current second task step, and the path agent is selected as the second auxiliary agent.

[0183] The search agent generates scheduling information for a third agent (the search agent generates a search route within the region, and the path planning agent). This third agent scheduling information (the search agent generates a search route within the region, and the path planning agent) is then sent to the path planning agent.

[0184] After receiving the scheduling information from the third agent, the path planning agent executes the task, obtains the processing result (i.e., route information), generates the third agent scheduling information (path planning agent, route information, search agent), and sends it to the search agent.

[0185] After receiving scheduling information from the third agent (path planning agent, route information, and search agent), the search agent obtains the processing result (i.e., route information).

[0186] The local memory and processing results (i.e., route information) are concatenated. This concatenation is achieved through steps S510-S540, which will not be elaborated in this embodiment. The large model is then invoked for inference, and the result (i.e., the next step after the current second task step) is: search according to the route.

[0187] The search route in step 1.2) "Acquire visual information and identify targets according to the search route" is updated using the route information in the processing results, and the updated step 1.2) is used as the current second task step.

[0188] The search agent begins executing the search task in the current second task step, finds the target location, and obtains the processing result.

[0189] Extract memory keywords from the target location included in the processing results, retrieve the target memory information from the memory bank based on the memory keywords, concatenate the target memory information with the processing results, call the large model for reasoning, and obtain the result (i.e. the current task execution status): the target has been found, and send the target information and terrain information to the search and rescue agent.

[0190] The second task execution plan of the intelligence agent is completed, and the processing results are target location information and terrain information. The search intelligence agent generates the third intelligence agent scheduling information (search intelligence agent, target and terrain information, search and rescue intelligence agent) and sends a message to the search and rescue intelligence agent.

[0191] The search and rescue agent receives scheduling information from a third agent (search agent, target and terrain information, search and rescue agent), obtains the processing results (target location information and terrain information), and concatenates the processing results with local memory. This concatenation is implemented through steps S510-S540, which will not be elaborated in this embodiment. The concatenated result is sent to the large model for inference, and the result (i.e., the current task execution status) is: the first step has been completed, and the second step is now being executed.

[0192] Step 2) calls the path planning agent to generate the action route as the current first task step.

[0193] The search and rescue agent changes the current task execution status to "First step completed, second step in progress".

[0194] Since the current first task step is not executed by the search and rescue intelligent agent, the second intelligent agent scheduling information is generated based on the current first task step and the relevant auxiliary intelligent agent (i.e., the path planning intelligent agent): (search and rescue intelligent agent, generates action route based on target and terrain information, path planning intelligent agent) and sent to the path planning intelligent agent.

[0195] The path planning agent executes the tasks in the scheduling information of the second agent, performs path planning, and generates a route as the processing result. The specific process will not be described again.

[0196] The path planning agent sends the scheduling information of the third agent (path planning agent, action route, search and rescue agent) to the search and rescue agent.

[0197] The search agent receives scheduling information from the third agent and obtains the processing result (action route) from it.

[0198] The search and rescue agent reasoned based on local memory and processing results (action route) to obtain the result (i.e. the current task execution status): the second step has been completed and the third step has begun.

[0199] Step 3) is to call upon the transportation agent to carry out the transport task according to the action route as the current first task step.

[0200] The search and rescue agent changes the current mission execution status to "Executing Phase 3 mission - transporting the target back".

[0201] Since the current first task step is not executed by the search and rescue intelligent agent, the second intelligent agent scheduling information is generated based on the current first task step and the relevant auxiliary intelligent agent (i.e., the transportation intelligent agent): (Search and rescue intelligent agent, bring the robot back according to the route - route information is xxx - target is xxx, transportation intelligent agent).

[0202] The search and rescue agent adds the current task execution status: The route for moving the target is xxxx to the local memory bank.

[0203] The transportation agent receives scheduling information from the second agent: (Search and rescue agent, bring the robot back according to the route - route information is xxx - target is xxx, transportation agent), and begins to execute the transportation task according to the information.

[0204] After the transportation agent completes its task, it generates scheduling information for the third agent (transportation agent, task execution result, search and rescue agent) and sends it to the search and rescue agent.

[0205] The search and rescue agent receives scheduling information from a third agent: (transport agent, task execution result, search and rescue agent). It then concatenates the scheduling information (transport agent, task execution result, search and rescue agent) with its local memory, calls the large model for inference, and obtains the result (i.e., the current task execution status): task execution successful or task execution failed, and the task ends.

[0206] The search and rescue agent clears its local memory and terminates its operation.

[0207] The above describes the complete agent scheduling and coordination process. During this process, agents communicate and coordinate via messages, and this process is entirely asynchronous. After sending a message, an agent does not block and wait for the message result. Only when the agent receives another message will it deduce the next action based on its memory and the message's combined reasoning. The scheduling between agents is chain-like and dynamic, without a pre-defined scheduling order.

[0208] The distributed multi-agent task processing system provided in this application implements message-based multi-agent collaboration. It does not require prior construction of agent workflows; the scheduling and collaboration between agents are dynamically formed during operation, rather than being pre-defined. This maximizes the reasoning ability, problem-solving adaptability, and flexibility of the large model. Furthermore, message-based agent collaboration is similar to human communication, a natural and scalable form of collaboration. Any agent can collaborate with other agents through a unified message format without the need for additional interfaces, exhibiting great scalability.

[0209] The following describes method embodiments of this application, which can be used to control device embodiments of this application. For details not disclosed in the method embodiments of this application, please refer to the device embodiments of this application.

[0210] Figure 8 A flowchart is shown for a task processing method based on a distributed multi-agent system according to an exemplary embodiment.

[0211] See Figure 8 Referring to the preceding description, in step S810, the planning and reasoning module uses the first task instruction input by the user and, based on the pre-stored agent information dictionary, selects at least one agent as the target agent from multiple agents. Then, based on the first task instruction and the target agent, it generates first agent scheduling information. The agent information dictionary includes the names of multiple agents and their corresponding functions, and the multiple agents are used to process the agent scheduling information.

[0212] In step S820, the message management module is used to send the scheduling information of the first intelligent agent to the target intelligent agent;

[0213] In step S830, the target agent executes the task contained in the scheduling information of the first agent.

[0214] The method performs similar functions to the system provided above. Other steps can be found in the previous descriptions and will not be repeated here.

[0215] This application discloses an electronic device, including: a processor; and a memory storing a computer program, which, when executed by the processor, causes the processor to execute the above-described instruction generation method.

[0216] For example, refer to Figure 9 , Figure 9 The illustrated electronic device 900 includes a processor 901 and a memory 903. The processor 901 and the memory 903 are connected, for example, via a bus 902. Optionally, the electronic device 900 may also include a transceiver 904. It should be noted that in practical applications, the transceiver 904 is not limited to one type, and the structure of this electronic device 900 does not constitute a limitation on the embodiments of the present invention.

[0217] Processor 901 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in this disclosure. Processor 901 may also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0218] Bus 902 may include a pathway for transmitting information between the aforementioned components. Bus 902 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 902 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0219] The memory 903 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other storage medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0220] The memory 903 stores application code that executes the present invention, and its execution is controlled by the processor 901. The processor 901 executes the application code stored in the memory 903 to implement the content shown in the foregoing method embodiments.

[0221] Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0222] This application discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, causes the processor to execute an instruction generation method.

[0223] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0224] The above are only some embodiments of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A task processing system based on distributed multi-agent systems, characterized in that, include: Multiple intelligent agents are used to process the intelligent agent scheduling information; The planning and reasoning module is used to select at least one intelligent agent as the target intelligent agent from the plurality of intelligent agents based on the first task instruction input by the user and a pre-stored intelligent agent information dictionary, and to generate first intelligent agent scheduling information based on the first task instruction and the target intelligent agent, wherein the intelligent agent information dictionary includes the names of the plurality of intelligent agents and their corresponding functions; The message management module is used to send the first agent scheduling information to the target agent, so that the target agent executes the first task contained in the first agent scheduling information.

2. The system according to claim 1, characterized in that, The target intelligent agent is used for: The first task is analyzed by calling a preset large model, and the analysis results are obtained; If the analysis result indicates that the first task cannot be directly executed, knowledge keywords are generated based on the first task. Based on the knowledge keywords, a target knowledge dictionary is obtained by searching a pre-stored knowledge rule base. The target knowledge dictionary is concatenated with the first task, and the first task is updated using the concatenation result.

3. The system according to claim 2, characterized in that, The target intelligent agent is also used for: If the analysis result indicates that the first task can be executed directly, then the first task is executed directly.

4. The system according to claim 2 or 3, characterized in that, The target intelligent agent is also used for: Match the functions of the target intelligent agent with the first task; If the matching result indicates that the target agent cannot perform the first task independently, at least one agent is selected as an auxiliary agent from the plurality of agents based on the agent information dictionary. Based on the auxiliary intelligent agent and its functions, as well as the first task, multiple first task steps are obtained, thereby obtaining a first task execution scheme; The first task execution plan is stored in a memory bank, wherein the memory bank is a local memory bank unique to the target intelligent agent and / or a public memory bank shared by the auxiliary intelligent agent and the target intelligent agent; The first task steps are executed sequentially based on the target intelligent agent and the auxiliary intelligent agent until the first task execution plan is completed.

5. The system according to claim 4, characterized in that, The target intelligent agent is also used for: S1: Take the first step in the first task execution plan as the current first task step; S2: Store the current task execution status in the memory bank, wherein the current task execution status includes at least one of the following: the execution progress of the current first task step, the execution progress of the first task execution plan, the execution details of the current first task step, and the current instruction reception status; S3: If the executing entity of the current first task step is an auxiliary intelligent agent, based on the target intelligent agent, generate second intelligent agent scheduling information according to the current first task step and the auxiliary intelligent agent, send the second intelligent agent scheduling information to the corresponding auxiliary intelligent agent, and jump to step S5; S4: If the target intelligent agent is the executing entity of the current first task step, execute the current first task step, obtain the processing result, and jump to step S7; S5: Obtain the third agent scheduling information sent by the auxiliary agent, wherein the third agent scheduling information is generated by the auxiliary agent after processing the second agent scheduling information; S6: Obtain the processing result based on the scheduling information of the third intelligent agent; S7: Update the current task execution status and / or the next step of the current first task step using the processing result; S8: Take the next step of the current first task step as the current first task step, and repeat steps S2-S8 until the first task execution plan is completed.

6. The system according to claim 5, characterized in that, The target intelligent agent is also used for: Extract memory keywords based on the processing results; Based on the memory keywords, the target memory information is obtained by searching the memory bank. The target memory information is concatenated with the processing result, and the processing result is updated using the concatenated result; Update the current task execution status and / or the next first task step using the updated processing results.

7. The system according to claim 5, characterized in that, The auxiliary intelligent agent is used for: The corresponding second task is obtained based on the scheduling information of the second intelligent agent; Execute the second task to obtain the processing result; Based on the processing results of the auxiliary intelligent agent, third intelligent agent scheduling information is generated; The scheduling information of the third agent is sent to the target agent.

8. The system according to claim 7, characterized in that, The auxiliary agent is also used for: The functions of the auxiliary intelligent agent are matched with the second task; If the matching result indicates that the auxiliary agent cannot perform the second task independently, at least one agent is selected from the plurality of agents as the second auxiliary agent based on the agent information dictionary. Based on the second auxiliary intelligent agent and its functions, as well as the second task, multiple second task steps are obtained, thereby obtaining a second task execution scheme; Store the second task execution plan into the memory bank; The second task steps are executed sequentially based on the first auxiliary agent and the second auxiliary agent until the second task execution plan is completed.

9. A task processing method based on distributed multi-agent systems, characterized in that, include: Using the planning and reasoning module, based on the first task instruction input by the user and a pre-stored agent information dictionary, at least one agent is selected as the target agent from multiple agents. Based on the first task instruction and the target agent, first agent scheduling information is generated. The agent information dictionary includes the names of the multiple agents and their corresponding functions. The multiple agents are used to process the agent scheduling information. Using the message management module, the scheduling information of the first intelligent agent is sent to the target intelligent agent; The target agent is used to execute the tasks contained in the scheduling information of the first agent.

10. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in claim 9.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they implement the method as described in claim 9.

Images