Task management method and device based on multi-agent collaboration and medium

Abstract

The invention discloses a task management method and device based on multi-agent cooperation and a medium. The method comprises the steps that an original task request initiated by a user is received; performing task decomposition on the original task request through a preset cue word based on a core reasoning model so as to decompose the original task request into a sub-task sequence formed by a plurality of sub-tasks; wherein the subtask sequence comprises a dependency relationship among the subtasks; target sub-agents matched with the sub-tasks are screened out from a sub-agent library, and the sub-tasks are distributed to the target sub-agents, so that the target sub-agents execute the sub-tasks according to the dependency relationship; and integrating the sub-task execution results returned by the target sub-agents to obtain a task result corresponding to the original task request, and feeding back the task result to the user.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a task management method, device, and medium based on multi-agent collaboration. Background Technology

[0002] With the rapid development of large language model technology, intelligent task processing systems based on large language models are evolving from single-agent architectures to multi-agent collaborative architectures. However, existing systems still face numerous challenges when dealing with complex tasks. Traditional single-agent systems are limited by fixed context windows, making it difficult to handle long-flowing, complex tasks. Existing multi-agent systems have significant shortcomings in task decomposition, mainly relying on preset rules or simple keyword matching for task breakdown. They cannot dynamically generate sub-task sequences with precise dependencies based on task semantics. This unstructured decomposition method leads to rigid task execution processes and makes it difficult to accurately express the logical relationships and data flow between sub-tasks. Furthermore, at the agent collaboration level, communication and scheduling mechanisms have significant efficiency bottlenecks. On the one hand, there is a lack of unified communication standards among agents, and fragmented message formats lead to high parsing costs and compatibility issues. On the other hand, centralized scheduling architectures easily become system performance bottlenecks when facing large-scale agent clusters, causing task distribution delays. In addition, the lack of effective routing strategies often leads to message loss, duplication, or out-of-order delivery, affecting the efficiency of agent scheduling and collaboration. Summary of the Invention

[0003] To address the aforementioned problems, this application proposes a task management method based on multi-agent cooperation, comprising: Receive the original task request initiated by the user; Based on the core reasoning model, the original task request is decomposed into a sequence of subtasks consisting of multiple subtasks by using preset prompts; wherein, the sequence of subtasks includes the dependencies between the subtasks. Select target sub-agents that match the sub-task from the sub-agent library, and assign the sub-tasks to the target sub-agents so that the target sub-agents execute the sub-tasks according to the dependencies; The subtask execution results returned by each target sub-agent are integrated to obtain the task result corresponding to the original task request, and the task result is fed back to the user.

[0004] In one implementation of this application, before decomposing the original task request, the method further includes: The original task request is subjected to intent recognition in order to determine the business domain to which the original task request belongs; Based on the business corpus corresponding to the business domain, the preset large language model is fine-tuned to obtain the core reasoning model applied to the original task request; The original task request is decomposed into a sequence of subtasks consisting of multiple subtasks by using preset prompts, specifically including: Based on the general task decomposition rules and the task decomposition rules of the business domain, construct multi-layered prompt words for decomposing the original task request; The multi-layered prompts and the original task request are input into the core reasoning model. The core reasoning model performs reasoning and outputs a sequence of subtasks obtained by decomposing the original task request. The subtask sequence consists of multiple subtasks, and each subtask includes a task description, task dependencies, and input parameters.

[0005] In one implementation of this application, a multi-layered prompt word for decomposing the original task request is constructed based on general task decomposition rules and the task decomposition rules of the business domain, specifically including: The system-level prompt words corresponding to the general task decomposition rules and the domain-level prompt words corresponding to the task decomposition rules of the business domain are concatenated to obtain the prompt words; Based on the prompt words, the original task request, and the output format constraints, a multi-layered prompt word is constructed to decompose the original task request.

[0006] In one implementation of this application, selecting target sub-agents that match the sub-task from a sub-agent library specifically includes: For the subtask to be assigned, calculate the matching degree between the subtask and the sub-agents in the sub-agent library, and select the sub-agents with matching degrees greater than a preset matching degree as candidate sub-agents; Determine whether the skills of the candidate sub-agents can support the execution of the sub-task. If so, select the candidate sub-agent with the highest matching degree as the target sub-agent for executing the sub-task.

[0007] In one implementation of this application, calculating the matching degree between the subtask and the sub-agents in the sub-agent library specifically includes: Obtain real-time status information corresponding to sub-agents in the sub-agent library; wherein, the real-time status information includes skill proficiency, current load, and average response time; Calculate the semantic similarity between the skill requirements of the subtask and the registered skills of the sub-agent. Then, using a weighted fusion algorithm, calculate the matching degree between the subtask and the sub-agent based on the real-time state information and the semantic similarity.

[0008] In one implementation of this application, determining whether the skills of the candidate sub-agent can support the execution of the sub-task specifically includes: Determine whether the input parameters of the subtask completely match the skill parameters of the candidate sub-agent, and whether the candidate sub-agent has the data access permissions required to execute the subtask; If so, it is determined that the skills of the candidate sub-agent are capable of supporting the execution of the sub-task.

[0009] In one implementation of this application, after assigning the subtask to the target sub-agent so that the target sub-agent executes the subtask according to the dependency relationship, the method further includes: Obtain the execution metrics corresponding to the target sub-agent when executing the sub-task; Based on the execution metrics, the skill proficiency of the target sub-agent corresponding to the sub-task is updated, so as to adjust the matching degree between the target sub-agent and the sub-task through the updated skill proficiency.

[0010] In one implementation of this application, the subtask execution results returned by each target sub-agent are integrated to obtain the task result corresponding to the original task request, specifically including: Based on the routing component, the subtask execution results returned by each target sub-agent are transmitted to the scheduling and coordination module according to the preset routing rules, so that the scheduling and coordination module can integrate the subtask execution results to obtain the task result corresponding to the original task request.

[0011] This application provides a task management device based on multi-agent collaboration, the device comprising: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a task management method based on multi-agent cooperation as described above.

[0012] This application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: As described in any of the preceding items, this is a task management method based on multi-agent collaboration.

[0013] The task management method based on multi-agent cooperation proposed in this application can bring the following beneficial effects: By leveraging a core reasoning model and prompts, intelligent task decomposition is achieved. This enables the dynamic generation of sub-task sequences with dependencies based on task semantics, eliminating reliance on pre-defined rules and ensuring both flexibility and logical rigor in task planning. Secondly, by integrating the sub-task execution results from different target agents, the fragmentation of message formats between agents is eliminated, achieving efficient and reliable task distribution and result transmission. This significantly improves the accuracy of task decomposition and the optimization of execution paths when handling complex, long-process tasks in a multi-agent system, while also substantially reducing communication and scheduling overhead, effectively enhancing task processing efficiency. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a task management method based on multi-agent collaboration provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a task management device based on multi-agent collaboration, provided as an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] Existing single-agent systems rely on a single Large Language Model (LLM) to complete all task processing. Limited by the LLM's context window (e.g., GPT-4's context window is 128K tokens), they are highly susceptible to inference errors due to context overload when handling complex, multi-step tasks, such as "generating an annual financial analysis report and synchronizing it to a cloud collaboration platform, while simultaneously generating an executable business optimization plan based on the report's conclusions." Furthermore, single agents lack task decentralization capabilities, requiring all logic to be coupled into a single prompt word, resulting in extremely complex prompt word design and low task processing efficiency. Taking robo-advisory scenarios as an example, a single agent must simultaneously handle multiple sub-tasks such as market analysis, asset allocation, risk assessment, and portfolio generation. This not only results in inference times of several minutes but also an accuracy rate of less than 60%, far from meeting the accuracy requirements of the financial sector.

[0017] Based on this, multi-agent systems, as a key technology for achieving complex task collaboration, have shown great application potential in fields such as intelligent customer service, enterprise digital transformation, and intelligent decision-making. However, in existing technologies, multi-agent systems mostly rely on manual rules or simple keyword matching when decomposing tasks, lacking standardized task breakdown and dependency definition. For example, in e-commerce order processing scenarios, existing systems may solidify the execution order of sub-tasks such as "user order placement, payment verification, logistics allocation, and after-sales processing" into a fixed process, unable to dynamically adjust according to actual business needs (such as the process differences when users choose cash on delivery or online payment), resulting in a serious lack of flexibility and accuracy in task decomposition. At the same time, there is a lack of explicit definition of task priority and resource dependencies, which can easily lead to sub-task execution conflicts or resource competition. For example, in medical diagnosis scenarios, if image interpretation and medical history retrieval sub-tasks simultaneously call the same medical database, it may lead to database access congestion and significantly reduce system response speed.

[0018] Furthermore, the core value of multi-agent systems lies in collaboration, but existing technologies have significant shortcomings in communication standards and scheduling efficiency between agents. Firstly, there is the problem of fragmented communication formats. Different agents may use various formats such as JSON, XML, and custom text for communication, leading to high message parsing costs and error rates. For example, in a medical system, the image interpretation agent outputs image features in XML format, while the diagnostic reasoning agent only supports JSON input, requiring the development of an additional format conversion module, increasing system complexity and potential failure points. Secondly, most multi-agent systems rely on a single scheduler for task allocation. In large-scale agent clusters, the scheduler easily becomes a performance bottleneck, causing task distribution delays of up to tens of seconds, severely impacting user experience. Moreover, the lack of clear routing rules for message transmission between agents makes message loss, duplication, or out-of-order delivery prone to occur. For instance, in financial trading scenarios, if the results of the risk assessment agent are not promptly transmitted to the transaction execution agent, it may lead to delayed trading decisions and missed optimal trading opportunities.

[0019] To address the aforementioned technical problems, embodiments of this application provide a task management method and system based on multi-agent collaboration, which achieves efficient processing of complex tasks through a multi-agent collaborative task management and scheduling architecture.

[0020] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] like Figure 1 As shown in the figure, an embodiment of this application provides a task management method based on multi-agent cooperation, including: S101: Receive the original task request initiated by the user.

[0022] This application provides a task management system based on multi-agent collaboration. The system includes a task management module and a scheduling and coordination module. The task management module is responsible for parsing user intent, decomposing complex tasks, and planning the execution order of subtasks. The scheduling and coordination module is responsible for the final distribution decision of subtasks. The system can receive original task requests initiated by users and analyze them through the task management module to identify the user's current intent and the specific business domain involved in the original task request.

[0023] S102: Based on the core reasoning model, the original task request is decomposed into a sequence of subtasks consisting of multiple subtasks by using preset prompts; wherein, the sequence of subtasks includes the dependencies between the subtasks.

[0024] The core inference model is finely tuned from a mainstream large language model. Through prompt engineering, it guides the model to accurately interpret user intent and automatically decompose tasks. In other words, the core inference model uses pre-set prompts to decompose the original task request, breaking down complex tasks into a sequence of subtasks. This subtask sequence includes dependencies between subtasks, clarifying the execution order and mutual influence of each subtask. For example, some subtasks may require the completion of other subtasks before execution can begin, or the output of some subtasks may serve as input for others. By clearly defining these dependencies, the execution order of subtasks can be arranged more rationally, improving the efficiency and accuracy of task execution.

[0025] In one embodiment, before decomposing the original task request, a suitable LLM model needs to be selected, and then the core inference model is obtained through supervised fine-tuning. When selecting an LLM model, it is necessary to choose an appropriate model based on the characteristics of the business scenario. For example, the financial field has high compliance requirements, so an LLM model fine-tuned based on financial corpora can be selected; the medical field has high professional requirements, so an LLM model fine-tuned based on medical literature, such as Med-PaLM, can be selected. For general scenarios, an open-source LLM model can be used and fine-tuned in conjunction with business domain data to balance the model's cost and performance. First, intent recognition needs to be performed on the original task request to determine the business domain to which the original task request belongs. Then, based on the business corpus of that business domain, a paired dataset containing user requests and task decomposition results is constructed. Based on this dataset, a supervised fine-tuning method is used to fine-tune the pre-set large language model, i.e., the LLM model, so that the LLM learns the logic and format of task decomposition, thereby obtaining a core inference model that can be applied to the original task request.

[0026] Using the core reasoning model described above, complex tasks can be decomposed using prompts. The prompts employ a hierarchical design, including system-level and domain-level prompts. System-level prompts define general task decomposition rules, such as "subtasks should cover all user requests without omissions or redundancy" and "dependencies should accurately reflect the execution order of subtasks, supporting both parallel and sequential execution." Domain-level prompts refer to industry-specific task decomposition rules. For example, in the financial field, the prompt is "prioritize market data retrieval, then risk assessment, and finally portfolio generation"; in the medical field, it's "symptom analysis and image interpretation can be performed in parallel, and diagnostic reasoning depends on the results of both." Based on these general and domain-specific task decomposition rules, multi-layered prompts can be constructed accordingly.

[0027] It should be noted that for system-level prompts corresponding to general task decomposition rules and domain-level prompts corresponding to business domain task decomposition rules, the two need to be concatenated to form the final prompts. The original task request is the content input by the user, which can include multimodal data such as text, images, and audio. Format constraints refer to the output content format of the model, which needs to be enforced through JSON Schema. The prompt structure used in this embodiment is in the form of "prompt + user input + format constraints". Therefore, based on the above prompts, original task requests, and output format constraints, a multi-layered prompt structure for decomposing the original task request can be constructed.

[0028] The core inference model inputs multi-layered prompts and the original task request. Guided by these prompts, it performs deep semantic understanding and logical reasoning, ultimately outputting a sequence of subtasks derived from the original task request. Each subtask sequence comprises multiple subtasks, each explicitly including a task description (defining what to do), task dependencies, and input parameters (indicating the specific data or resources required for execution). By understanding the user's task requirements and decomposing them through the core inference model, it is no longer limited to rigid process templates but can adapt to complex user requests, flexibly adjusting the content and execution order of subtasks. Furthermore, the dependencies within the subtask sequence clearly reflect the logical connections between subtasks, avoiding execution conflicts or resource contention and improving task execution efficiency and accuracy.

[0029] It should be noted that this application's embodiments introduce a hybrid contextual retrieval mechanism based on vector databases and time-series databases, endowing the system with continuous memory and cognitive capabilities during the reasoning process. By utilizing models such as Sentence-BERT, historical dialogue text is transformed into high-dimensional vectors and stored in vector databases such as Milvus, thereby achieving semantic-level contextual retrieval and accurately recalling historical information related to the current task. Simultaneously, time-series databases such as InfluxDB are used to completely record the temporal context of the dialogue and task execution, ensuring strict temporal consistency of the contextual information. When a user initiates a new request, the system dynamically integrates the current instruction with the recalled historical context. For example, when processing a "adjust portfolio" request, it automatically associates the user's past risk assessment results and holding records, providing rich background knowledge for the core reasoning model. This allows its task decomposition and decision-making to be based not only on the current input but also on a deep understanding of the user's historical state, thus significantly improving the accuracy of task decomposition.

[0030] S103: Select target sub-agents that match the sub-tasks from the sub-agent library, and assign the sub-tasks to the target sub-agents so that the target sub-agents can execute the sub-tasks according to the dependencies.

[0031] The sub-agent library stores various sub-agents with different capabilities, including their attributes, skills, and workloads. The library supports dynamic registration and retrieval of agents. When registering an agent, its attributes, skills, and workloads must be specified and registered in the library for subsequent scheduling. When selecting target sub-agents, the library matches the requirements of each subtask with the attributes, skills, and workloads of the agents in the library. Each subtask is assigned to its corresponding target sub-agent for processing. The target sub-agent executes its tasks based on prompts, including input parameters, dependencies, and execution rules, following the dependencies in the subtask sequence. If a subtask has prerequisite dependencies, the target sub-agent must wait for the prerequisite subtask to complete and obtain its output before it can begin executing its own subtask. For example, in a software development project, the code testing subtask can only begin after the code writing subtask is completed; the target sub-agent must work systematically according to these dependencies.

[0032] In one embodiment, retrieving corresponding sub-agent resources for each sub-task requires the collaborative work of the sub-agent library retrieval module and the skill library retrieval module to match suitable sub-agents and skills for the sub-task. The sub-agent library retrieval module retrieves sub-agents based on the matching degree between the sub-task and the sub-agents in the sub-agent library. For each currently assigned sub-task, its matching degree with each sub-agent in the sub-agent question bank needs to be calculated. Only sub-agents with a matching degree greater than a preset matching degree are added to the candidate sub-agent set as candidate sub-agents.

[0033] When calculating the matching degree between subtasks and sub-agents, a weighted fusion retrieval method is required, comprehensively considering multiple factors such as the sub-agent's attributes, skills, and workload. First, the real-time state information of the sub-agents in the sub-agent database is obtained, including skill proficiency, current workload, and average response time. Then, the semantic similarity between the skill requirements of the subtask and the registered skills of the sub-agent is calculated. Semantic similarity reflects the degree of matching between the sub-agent's skills and the sub-task requirements, thus measuring the suitability of the sub-agent and sub-task from a skill perspective. Next, a weighted fusion algorithm is used to sum the real-time state information and semantic similarity of the sub-agent to obtain the matching degree between the subtask and the sub-agent. Calculating the matching degree using semantic similarity and real-time state information allows for a comprehensive evaluation of the sub-agent's skill suitability and real-time state, thereby selecting the most suitable target sub-agent and improving the accuracy and rationality of subtask allocation.

[0034] After obtaining the set of candidate sub-agents, further screening of these sub-agents is required. The skill base retrieval module stores information on the various skills possessed by the sub-agents. For each candidate sub-agent, the skill base retrieval module determines whether the candidate sub-agent's skills can support the execution of the sub-task.

[0035] First, it's necessary to determine whether the input parameters of the subtask completely match the skill parameters of the candidate sub-agent. Input parameters refer to the structured data defined during the task decomposition phase, which is essential for executing the task. The skill parameters of the sub-agent are the standardized input interface specifications declared when registering its skills with the system. These specifications define the types of information required to execute the skill, along with the corresponding information names, data types, and formats. For example, the skill parameter definition for a weather query agent might require an input object containing "city" (a string) and "date" (a date format string). Only when the input parameters and skill parameters completely match can the sub-agent correctly understand and receive the subtask parameters. After confirming the match, it's further necessary to verify whether the candidate sub-agent has the necessary data access permissions to execute the subtask. Only when the agent is authorized to access specific data resources or perform specific operations can it successfully execute the subtask; otherwise, even if the skills match perfectly, the sub-agent will not be selected due to insufficient permissions. By enforcing mandatory permission verification, it is ensured that sensitive data and critical operations can only be accessed by authorized sub-agents, effectively preventing unauthorized behavior and meeting data security requirements.

[0036] Only if both of the above conditions are met will the system confirm that the sub-agent's skills can support the execution of the sub-task. If the sub-agent meets both the matching degree requirement and the skill requirements for completing the sub-task, then the candidate sub-agent with the highest matching degree will be identified as the target sub-agent for executing the sub-task.

[0037] To ensure efficient and stable task execution, the system monitors the execution process of the target sub-agents in real time. When an execution anomaly is detected in a target sub-agent—such as failing to complete a task for an extended period or producing unexpected results—the system takes timely measures. These measures may include reassigning the sub-task to another suitable target sub-agent or diagnosing and repairing the malfunctioning sub-agent to ensure successful task completion. Furthermore, the system records execution metrics for the target sub-agents, including execution results, execution time, and resource consumption, for subsequent analysis and optimization of the task execution process.

[0038] In one embodiment, the sub-agent updates its skill proficiency in real time based on its actual performance during the execution of sub-tasks. The skill proficiency of the sub-agent is evaluated by acquiring its execution metrics in real time. This evaluation can be calculated using a pre-defined evaluation function. The skill proficiency calculated from the execution metrics can be used to update the skill proficiency of the target sub-agent corresponding to the sub-task. This updated skill proficiency allows for adjustments to the matching degree between the target sub-agent and the sub-task in subsequent sub-task allocation and scheduling. If the target sub-agent executes the sub-task accurately, with short execution time and low resource consumption, its skill proficiency will be significantly improved; conversely, if the execution result is flawed, the execution time is too long, or the resource consumption is too high, the skill proficiency will decrease accordingly.

[0039] Real-time updates to skill proficiency can reflect changes in the capabilities of sub-agents during actual task execution. Furthermore, the updated skill proficiency will affect the allocation of subsequent sub-tasks. When selecting target sub-agents in the future, sub-agents with higher skill proficiency will be more likely to be selected to execute sub-tasks, which can effectively improve the efficiency and quality of sub-task execution.

[0040] S104: Integrate the subtask execution results returned by each target sub-agent to obtain the task result corresponding to the original task request, and feed the task result back to the user.

[0041] After each target sub-agent completes its corresponding sub-task, it transmits the sub-task execution results returned by each target sub-agent to the scheduling and coordination module based on the routing component and according to preset routing rules. Routing rules define the data transmission path and target, ensuring that the sub-task execution results are accurately transmitted from the target sub-agent to the scheduling and coordination module. Routing rules can be flexibly configured based on factors such as the nature of the task, the distribution of sub-agents, and the system architecture. Upon receiving the sub-task execution results, the scheduling and coordination module integrates them to obtain the final task result corresponding to the original task request. The task result is then fed back to the user through the scheduling and coordination module, enabling the user to obtain the required task result in a timely manner. This helps the user understand the task's progress and make appropriate decisions. During the integration process, the routing rules standardize the transmission of sub-task execution results, ensuring the accuracy and efficiency of result transmission, avoiding data loss or transmission errors, and guaranteeing the smooth execution of the entire task process.

[0042] The above are embodiments of the methods proposed in this application. Based on the same idea, some embodiments of this application also provide devices and non-volatile computer storage media corresponding to the above methods.

[0043] Figure 2 This is a schematic diagram of the structure of a task management device based on multi-agent collaboration, provided as an embodiment of this application. Figure 2 As shown, it includes: At least one processor; and, At least one processor-communication-connected memory; wherein, The memory stores instructions that can be executed by at least one processor, which enables the at least one processor to perform a multi-agent cooperative task management method as described in any of the preceding claims.

[0044] This application provides a non-volatile computer storage medium storing computer-executable instructions, which are configured as follows: As described in any of the preceding items, this is a task management method based on multi-agent collaboration.

[0045] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0046] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0047] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0052] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0053] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0054] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0055] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A task management method based on multi-agent collaboration, characterized in that, The method includes: Receive the original task request initiated by the user; Based on the core reasoning model, the original task request is decomposed into a sequence of subtasks consisting of multiple subtasks by using preset prompts; wherein, the sequence of subtasks includes the dependencies between the subtasks. Select target sub-agents that match the sub-task from the sub-agent library, and assign the sub-tasks to the target sub-agents so that the target sub-agents execute the sub-tasks according to the dependencies; The subtask execution results returned by each target sub-agent are integrated to obtain the task result corresponding to the original task request, and the task result is fed back to the user.

2. The task management method based on multi-agent collaboration according to claim 1, characterized in that, Before decomposing the original task request into tasks, the method further includes: The original task request is subjected to intent recognition in order to determine the business domain to which the original task request belongs; Based on the business corpus corresponding to the business domain, the preset large language model is fine-tuned to obtain the core reasoning model applied to the original task request; The original task request is decomposed into a sequence of subtasks consisting of multiple subtasks by using preset prompts, specifically including: Based on the general task decomposition rules and the task decomposition rules of the business domain, construct multi-layered prompt words for decomposing the original task request; The multi-layered prompts and the original task request are input into the core reasoning model. The core reasoning model performs reasoning and outputs a sequence of subtasks obtained by decomposing the original task request. The subtask sequence consists of multiple subtasks, and each subtask includes a task description, task dependencies, and input parameters.

3. The task management method based on multi-agent collaboration according to claim 2, characterized in that, Based on the general task decomposition rules and the task decomposition rules of the business domain, a multi-layered prompt word is constructed to decompose the original task request, specifically including: The system-level prompt words corresponding to the general task decomposition rules and the domain-level prompt words corresponding to the task decomposition rules of the business domain are concatenated to obtain the prompt words; Based on the prompt words, the original task request, and the output format constraints, a multi-layered prompt word is constructed to decompose the original task request.

4. The task management method based on multi-agent collaboration according to claim 1, characterized in that, Selecting target sub-agents that match the sub-task from the sub-agent library specifically includes: For the subtask to be assigned, calculate the matching degree between the subtask and the sub-agents in the sub-agent library, and select the sub-agents with matching degrees greater than a preset matching degree as candidate sub-agents; Determine whether the skills of the candidate sub-agents can support the execution of the sub-task. If so, select the candidate sub-agent with the highest matching degree as the target sub-agent for executing the sub-task.

5. The task management method based on multi-agent collaboration according to claim 4, characterized in that, Calculating the matching degree between the subtask and the sub-agents in the sub-agent library specifically includes: Obtain real-time status information corresponding to sub-agents in the sub-agent library; wherein, the real-time status information includes skill proficiency, current load, and average response time; Calculate the semantic similarity between the skill requirements of the subtask and the registered skills of the sub-agent. Then, using a weighted fusion algorithm, calculate the matching degree between the subtask and the sub-agent based on the real-time state information and the semantic similarity.

6. The task management method based on multi-agent collaboration according to claim 5, characterized in that, Determining whether the skills of the candidate sub-agent can support the execution of the sub-task specifically includes: Determine whether the input parameters of the subtask completely match the skill parameters of the candidate sub-agent, and whether the candidate sub-agent has the data access permissions required to execute the subtask; If so, it is determined that the skills of the candidate sub-agent are capable of supporting the execution of the sub-task.

7. The task management method based on multi-agent collaboration according to claim 1, characterized in that, After assigning the subtask to the target sub-agent so that the target sub-agent executes the subtask according to the dependency relationship, the method further includes: Obtain the execution metrics corresponding to the target sub-agent when executing the sub-task; Based on the execution metrics, the skill proficiency of the target sub-agent corresponding to the sub-task is updated, so as to adjust the matching degree between the target sub-agent and the sub-task through the updated skill proficiency.

8. The task management method based on multi-agent collaboration according to claim 1, characterized in that, The subtask execution results returned by each target sub-agent are integrated to obtain the task result corresponding to the original task request, specifically including: Based on the routing component, the subtask execution results returned by each target sub-agent are transmitted to the scheduling and coordination module according to the preset routing rules, so that the scheduling and coordination module can integrate the subtask execution results to obtain the task result corresponding to the original task request.

9. A task management device based on multi-agent collaboration, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a task management method based on multi-agent collaboration as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: A task management method based on multi-agent collaboration as described in any one of claims 1-8.

Images