Multi-agent collaborative task decomposition and execution method, device, equipment and medium

Abstract

The application relates to a multi-agent collaborative task decomposition and execution method, device, equipment and medium, relates to the field of artificial intelligence, can be applied to the fields of finance and medicine, and comprises the following steps: analyzing logical dependency relationships among each subtask in a candidate subtask list to construct a task directed acyclic graph; each subtask in the task directed acyclic graph is assigned to a target agent model for execution; if the subtask execution state is a completion state, whether the dependency condition of a post-subtask is satisfied is checked based on the dependency edge in the task directed acyclic graph; if the condition is satisfied, scheduling and execution of the post-subtask are triggered; and if the execution state is an execution abnormal state, the subtask with the execution abnormality is dynamically rescheduled and executed. According to the application, complex medical tasks are upgraded from artificial static management and multi-agent disordered independent execution to machine dynamic arrangement and multi-agent efficient collaborative execution, and problems such as task logic confusion, idle computing resources and medical risk hidden dangers are completely eliminated.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and can be applied to the financial and medical fields. Specifically, it relates to a method, apparatus, device, and medium for multi-agent collaborative task decomposition and execution. Background Technology

[0002] With the deepening application of artificial intelligence technology in the healthcare field, the level of intelligent processing of medical tasks has been significantly improved. However, in real-world medical scenarios, the processing of complex tasks often requires the collaborative participation of multiple professional roles. Taking "multidisciplinary consultation for cancer patients" as an example, this task not only requires radiologists to interpret medical images and pathologists to analyze biopsy tissues, but also requires oncologists to develop chemotherapy plans, surgeons to assess surgical feasibility, and radiation oncologists to design radiotherapy plans. This division of labor and collaboration among multiple professional roles places extremely high demands on existing intelligent processing systems.

[0003] Currently, most solutions for handling complex medical tasks involve a combination of manual decomposition and independent execution. Specifically, this involves manually breaking down complex medical tasks into multiple executable subtasks, which are then executed by a single or multiple independent intelligent agents. However, because these subtasks often have complex logical dependencies, existing systems lack the ability to dynamically manage and automatically orchestrate these complex dependencies. Manual management is prone to errors, leading to disordered task execution order, which in turn consumes significant computing and medical human resources and may even trigger medical risks. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, device and medium for multi-agent collaborative task decomposition and execution. The main purpose is to solve the technical problem that the existing technology mainly relies on manual decomposition of complex medical tasks into multiple executable sub-tasks, which are then executed by a single agent or multiple independent agents respectively. Since there are often complex logical dependencies between sub-tasks, the existing system lacks the ability to dynamically manage and automatically arrange such complex dependencies. Manual management is prone to errors, resulting in chaotic task execution order, which in turn greatly consumes computing resources and medical human resources, and may even cause medical risks.

[0005] According to a first aspect of this application, a method for decomposing and executing multi-agent cooperative tasks is provided, the method comprising: Receive description information of the target medical task, and perform deep semantic parsing on the description information to extract key task elements; Based on the medical knowledge base and the key task elements, a list of candidate subtasks required to complete the target medical task is generated, and the logical dependencies between the subtasks in the candidate subtask list are resolved. Based on the logical dependencies, construct a directed acyclic graph of tasks containing the list of candidate subtasks; Based on multiple agent models and their corresponding real-time load conditions, each subtask in the directed acyclic graph of the task is assigned to a target agent model with matching capabilities and reasonable load for execution. The execution status of each subtask is monitored in real time. If the execution status is that the subtask is completed, the dependency conditions of its subsequent subtasks are checked based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is that the execution is abnormal, the subtask with execution abnormality is dynamically rescheduled for execution. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. After all the subtasks have been completed, the scattered execution results of the multiple agent models are collected, and the scattered execution results are integrated into a unified output result according to the fusion strategy.

[0006] According to a second aspect of this application, a multi-agent cooperative task decomposition and execution apparatus is provided, the apparatus comprising: The receiving module is used to receive the description information of the target medical task and perform deep semantic parsing processing on the description information to extract key task elements. The generation module is used to generate a list of candidate subtasks required to complete the target medical task based on a medical knowledge base and the key task elements, and to parse the logical dependencies between the subtasks in the candidate subtask list. A construction module is used to construct a directed acyclic graph of tasks containing the list of candidate subtasks based on the logical dependencies. The allocation module is used to allocate each subtask in the directed acyclic graph of the task to the target intelligent agent model with matching capabilities and reasonable load for execution, based on multiple intelligent agent models and the corresponding real-time load of the multiple intelligent agent models. The monitoring module is used to monitor the execution status of each subtask in real time. If the execution status is a subtask completion status, it checks whether the dependency conditions of its subsequent subtasks are met based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is an execution exception status, the execution of the subtask with execution exception is dynamically rescheduled. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. The integration module is used to collect the scattered execution results of the multiple intelligent agent models after all the sub-tasks have been executed, and to integrate the scattered execution results into a unified output result according to the fusion strategy.

[0007] According to a third aspect of this application, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method of the first aspect described above.

[0008] According to a fourth aspect of this application, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of the first aspect described above.

[0009] Compared with existing technologies, the multi-agent collaborative task decomposition and execution method, apparatus, device, and medium provided in this application can receive description information of a target medical task, perform deep semantic parsing processing on the description information to extract key task elements; generate a candidate sub-task list required to complete the target medical task based on a medical domain knowledge base and key task elements, and parse the logical dependencies between each sub-task in the candidate sub-task list; construct a directed acyclic graph of tasks containing the candidate sub-task list based on the logical dependencies; and, according to multiple agent models and their corresponding real-time load conditions, decompose and execute each sub-task in the directed acyclic graph of tasks. Tasks are assigned to target agent models with matching capabilities and reasonable loads for execution. The execution status of each subtask is monitored in real time. If the execution status is that the subtask is completed, the dependency conditions of its subsequent subtasks are checked based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtasks are triggered. If the execution status is that the execution is abnormal, the abnormal subtasks are dynamically rescheduled for execution. The dependency condition is that all the preceding subtasks required by the subsequent subtasks have been completed. After all subtasks have been executed, the scattered execution results of multiple agent models are collected, and the scattered execution results are integrated into a unified output result according to the fusion strategy.

[0010] The solution proposed in this application enables AI to automatically understand task intent, automatically generate sub-tasks by combining with a medical knowledge base, and automatically identify prerequisite dependencies, parallel possibilities, and information flow dependencies, transforming them into a structured DAG graph. This not only frees medical staff from tedious process arrangement but also fundamentally eliminates task decomposition errors and sequence confusion caused by limitations or negligence in human cognition.

[0011] This application does not simply dump tasks onto the intelligent agent, but strictly follows the DAG topology. The subsequent task is only automatically triggered when all the preceding subtasks required by the subsequent subtask are completed (information flow is complete). This completely avoids the logical collapse and medical risks caused by blindly executing the subsequent task before the preceding results are available, and ensures the rigor and safety of executing complex clinical pathways.

[0012] In real-world medical computing environments, agents may malfunction or time out. This application can detect anomalies in real time and automatically trigger a rescheduling mechanism (such as assigning to a backup agent), achieving process self-healing without human intervention. This effectively prevents the cascading risk of local anomalies spreading to the global system and significantly reduces the probability of medical safety incidents caused by system failures.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of the application environment for a multi-agent cooperative task decomposition and execution method provided in an embodiment of this application; Figure 2 A flowchart illustrating a multi-agent collaborative task decomposition and execution method provided in an embodiment of this application; Figure 3 A task directed acyclic graph provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a multi-agent cooperative task decomposition and execution device provided in an embodiment of this application; Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of another computer device provided in an embodiment of this application. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0018] The multi-agent cooperative task decomposition and execution method provided in this application embodiment can be applied to, for example, Figure 1 In this application environment, the client communicates with the server via a network. The server can receive the description information of the target medical task from the client, perform deep semantic parsing on the description information to extract key task elements; based on the medical domain knowledge base and key task elements, generate a list of candidate subtasks required to complete the target medical task, and parse the logical dependencies between the subtasks in the candidate subtask list; based on the logical dependencies, construct a directed acyclic graph (DAG) containing the candidate subtask list; according to multiple agent models and their corresponding real-time load, assign each subtask in the DAG to the target agent model with matching capabilities and reasonable load for execution; monitor the execution status of each subtask in real time; if the execution status is that the subtask is completed, check whether the dependency conditions of its subsequent subtasks are met based on the dependency edges in the DAG; if they are met, the subsequent subtasks are scheduled for execution; if the execution status is that the execution is abnormal, the abnormal subtasks are dynamically rescheduled for execution, wherein the dependency condition is that all the preceding subtasks required by the subsequent subtasks have been completed; after all subtasks have been executed, collect the scattered execution results of multiple agent models, and integrate the scattered execution results into a unified output result according to the fusion strategy.

[0019] The solution proposed in this application enables AI to automatically understand task intent, automatically generate sub-tasks by combining with a medical knowledge base, and automatically identify prerequisite dependencies, parallel possibilities, and information flow dependencies, transforming them into a structured DAG graph. This not only frees medical staff from tedious process arrangement but also fundamentally eliminates task decomposition errors and sequence confusion caused by limitations or negligence in human cognition.

[0020] This application does not simply dump tasks onto the intelligent agent, but strictly follows the DAG topology. The subsequent task is only automatically triggered when all the preceding subtasks required by the subsequent subtask are completed (information flow is complete). This completely avoids the logical collapse and medical risks caused by blindly executing the subsequent task before the preceding results are available, and ensures the rigor and safety of executing complex clinical pathways.

[0021] In real-world medical computing environments, agents may malfunction or time out. This application can detect anomalies in real time and automatically trigger a rescheduling mechanism (such as assigning to a backup agent), achieving process self-healing without manual intervention. This effectively prevents the cascading risk of local anomalies spreading globally and significantly reduces the probability of medical safety incidents caused by system failures. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this application.

[0022] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the multi-agent cooperative task decomposition and execution method provided in this application embodiment includes the following steps: Step 101: Receive the description information of the target medical task, and perform deep semantic parsing on the description information to extract key task elements.

[0023] Specifically, the descriptive information of the target medical task input by the user is usually unstructured natural language information, such as: "The patient is a 68-year-old male, recently diagnosed with lung adenocarcinoma, clinical stage IIIA. He has mild chronic obstructive pulmonary disease (COPD), a 40-year smoking history, and has quit smoking for 5 years. The patient and his family hope to preserve lung function as much as possible and have high requirements for quality of life. Please develop a personalized comprehensive treatment plan for this patient, which needs to consider a reasonable combination of multiple methods such as surgery, chemotherapy, and radiotherapy."

[0024] After receiving the description information, the system can use a large language model (LLM) to perform deep semantic parsing on the description information and identify the parsing results of the target medical task. The parsing results may include core objectives (such as developing a personalized comprehensive treatment plan), key constraints (such as the need to comprehensively consider multiple methods and patient preferences), and expected outputs (such as a complete treatment plan). Subsequently, key task elements can be extracted from the parsing results.

[0025] In this embodiment, key task elements may include patient characteristics, disease information, task objectives, and constraints. Patient characteristics may include age, gender, and underlying diseases; disease information includes diagnosis, stage, and gene mutation status; constraints include medical insurance restrictions, patient preferences, and medical resource limitations.

[0026] For example: Patient characteristics: 68-year-old male, underlying disease is mild chronic obstructive pulmonary disease (COPD), personal history is 40 years of smoking but has quit smoking; disease information: diagnosed with lung adenocarcinoma, stage IIIA; task goal: to develop a personalized comprehensive treatment plan; constraints: patient preference is to preserve lung function and prioritize quality of life; implicit medical resource limitations and medical insurance limitations, etc.

[0027] Step 102: Based on the medical domain knowledge base and key task elements, generate a list of candidate subtasks required to complete the target medical task, and parse the logical dependencies between the subtasks in the candidate subtask list.

[0028] Among them, the medical knowledge base can be used to store medical knowledge, clinical guidelines, treatment pathways, drug information, disease standards, etc., and provide knowledge support for task parsing and decomposition.

[0029] Specifically, the system can call upon a pre-stored medical knowledge base, and based on the aforementioned extracted key task elements, deduce and generate a list of candidate sub-tasks required to complete the target medical task. For example, based on the aforementioned lung cancer case, the system generates the following list of candidate sub-tasks: T1: Reinterpretation of imaging data (clarifying the precise location of the tumor, its relationship with surrounding tissues, and the status of lymph nodes). T2: Pathological section review (confirming subtype and gene mutation status) T3: Precise assessment of lung function (to assess surgical tolerance) T4: Surgical feasibility assessment (based on imaging and lung function) T5: Recommended chemotherapy regimen (based on pathology and genetic results) T6: Radiation therapy plan design (for cases where complete resection may not be possible) T7: Comprehensive Solution Integration (Main Task) Subsequently, the logical dependencies between the various subtasks can be resolved. These logical dependencies can include prerequisite dependencies, parallel execution dependencies, and information flow dependencies. Specifically, the resolution process may include: Based on the execution timing constraints and logical order of subtasks in the medical clinical pathway, the logical dependencies between each subtask in the candidate subtask list are determined as prerequisite dependencies. For example, surgical evaluation must be performed after image interpretation and pulmonary function assessment, i.e., T4 prerequisite dependencies T1 and T3; chemotherapy recommendation requires pathology and genetic results, i.e., T5 prerequisite dependencies T2; radiotherapy design requires precise image localization, i.e., T6 prerequisite dependencies T1; comprehensive integration requires all professional opinions, i.e., T7 prerequisite dependencies T4, T5, and T6. Based on the data interaction requirements and mutual independence between subtasks, the logical dependencies between the subtasks in the candidate subtask list are determined to be parallel execution relationships; for example, T1 (imaging), T2 (pathology), and T3 (lung function) are independent of each other, do not require data interaction, and can be executed simultaneously. The logical dependencies between the subtasks in the candidate subtask list are determined by the correspondence between the input data of the subsequent subtask and the output data of the preceding subtask. For example, the input of T4 requires the output of T1 (image interpretation report) and the output of T3 (pulmonary function assessment report) as data flow input, thus determining the information flow dependency from T1, T3 to T4.

[0030] Step 103: Based on logical dependencies, construct a directed acyclic graph of tasks containing a list of candidate subtasks.

[0031] For embodiments of this disclosure, the task directed acyclic graph (DAG) automatic builder can automatically construct a task directed acyclic graph containing all subtask nodes in the candidate subtask list, dependency edges between nodes, critical paths between nodes, and a set of nodes that can be executed in parallel, based on the logical dependencies resolved in step 102.

[0032] The task-oriented directed acyclic graph structure constructed in this embodiment is as follows: Figure 3 As shown, T1, T2, and T3 serve as starting nodes in parallel (i.e., a set of nodes that can be executed in parallel), converge to nodes T4, T5, and T6, and finally converge to node T7. The system can dynamically determine the critical path based on the actual execution time, and can display the generated DAG visualization on the user's terminal interface through the DAG visualization and verification interface for manual review and adjustment. After confirming that the dependencies are correct, the system enters the scheduling phase.

[0033] Step 104: Based on the multiple agent models and their corresponding real-time load conditions, assign each subtask in the directed acyclic graph of tasks to the target agent model with matching capabilities and reasonable load for execution.

[0034] In this embodiment of the disclosure, the agent scheduling module can perform allocation according to preset rules, and the specific process is as follows: First, each subtask in the directed acyclic graph of tasks can be matched with the agent capability catalog and task-capability mapping matrix to filter out a list of candidate agents with the corresponding capabilities and level requirements. The agent capability catalog maintains the professional domain, capability level, supported task types, input and output format requirements, and historical execution quality indicators of each agent model. Secondly, the current load status (such as the number of tasks being executed) and estimated completion time of each intelligent agent model can be monitored in real time, and the intelligent agent model with matching capabilities and reasonable load can be selected from the candidate intelligent agent list as the target intelligent agent model. Finally, based on the urgency of the task or a preset scheduling strategy, the shortest waiting time priority strategy, the highest capability level priority strategy, or the load balancing priority strategy can be dynamically selected to assign subtasks to the target intelligent agent model.

[0035] For example, for the initial parallel execution of T1, T2, and T3, for T1 (image interpretation), the capability matching engine can filter out "Image Diagnosis Agent - Standard Version" and "Image Diagnosis Agent - Expert Version". The load-aware scheduler monitors that the "Expert Version" queue is currently idle (low load), while the "Standard Version" queue is busy. If the current strategy is "highest capability level priority" or "shortest waiting time priority", the system can assign T1 to the "Image Diagnosis Agent - Expert Version" for execution. Similarly, T2 can be assigned to the "Pathology Analysis Agent - Gene Version", and T3 can be assigned to the "Lung Function Assessment Agent". All three can be started and executed in parallel through the task distribution interface.

[0036] Step 105: Monitor the execution status of each subtask in real time. If the execution status is that the subtask is completed, check whether the dependency conditions of its subsequent subtasks are met based on the dependency edges in the directed acyclic graph of the task. If they are met, the subsequent subtasks are scheduled and executed. If the execution status is that the execution is abnormal, the subtasks with execution abnormalities are dynamically rescheduled and executed. The dependency condition is that all the preceding subtasks required by the subsequent subtasks have been completed.

[0037] In this embodiment of the disclosure, the real-time monitor of the task collaboration module can track the execution status of distributed sub-tasks (such as in progress, completed, failed, delayed, etc.): When T1, T2, and T3 are completed sequentially (status: completed), the intermediate result transmitter stores the output results in intermediate storage. The dependency triggering engine automatically checks whether the dependency conditions of its subsequent tasks (i.e., all preceding subtasks required by the subsequent subtask have been completed) are met: When T1 and T3 are detected to be completed normally, and the task collaboration module detects that the prerequisite conditions for T4 (surgical assessment) have been met, T4 scheduling is triggered and assigned to the "surgical assessment agent". When T2 is detected to be completed and the prerequisite conditions for T5 (chemotherapy recommendation) are met, T5 scheduling is triggered and assigned to the "chemotherapy recommendation agent". When T1 is detected to be completed and the prerequisite conditions for T6 (radiotherapy design) are met, T6 scheduling is triggered and assigned to the "radiotherapy planning agent".

[0038] If the execution status is an abnormal execution status (which may include at least subtask execution failure, delay exceeding a preset threshold, or failure of the agent model executing the subtask), then the abnormal subtask will be dynamically rescheduled for execution, specifically including: If the execution status is an execution exception, then re-match the candidate agent model from the agent capability catalog; The outputs of the subtasks that fail to execute and their required preceding subtasks are redistributed to alternative agent models to continue execution, and the state of the directed acyclic graph of the task is updated synchronously.

[0039] For example, if the "surgical evaluation agent" suddenly fails during the execution of T4, the system will execute the following rescheduling mechanism: First, if the task collaboration module detects a task failure (abnormal execution state), it will automatically trigger a rescheduling mechanism to rematch alternative agent models (such as "backup surgical assessment agent") from the agent capability catalog. The output results of the abnormal subtask T4 and its required preceding subtasks (T1 image report and T3 lung function report) will be redistributed to the alternative agent models to continue execution, and the state of the task's directed acyclic graph will be updated synchronously. The system records faults and rescheduling processes, but does not affect the overall process; T5 and T6 still execute in parallel.

[0040] Step 106: After all subtasks have been completed, collect the scattered execution results of multiple agent models and integrate the scattered execution results into a unified output result according to the fusion strategy.

[0041] In this embodiment of the disclosure, after all subtasks are completed, the result collector can collect a structured intermediate result set (i.e., distributed execution results) output by each agent model, for example: T1 output: Image interpretation report (tumor location, lymph node status) T2 output: Pathology report (adenocarcinoma, EGFR mutation positive) T3 Output: Pulmonary Function Assessment Report (Mild to Moderate Obstructive Ventilation Dysfunction) T4 output: Surgical assessment (operation is possible, but lung function needs to be considered; sublobar resection is recommended). T5 output: Chemotherapy recommended (EGFR-TKI targeted therapy preferred) T6 Output: Radiotherapy Design (SBRT Alternative) The fusion execution engine can select the corresponding fusion strategy from the fusion strategy library according to the task type, and integrate the scattered execution results into a unified output result according to the fusion strategy. The fusion strategy includes at least the sequential concatenation strategy, the weighted synthesis strategy, the conflict resolution strategy, and the complementary enhancement strategy.

[0042] Among them, the sequential splicing strategy can be to splice the results in a specific order; the weighted synthesis strategy can be to perform weighted synthesis of results from different sources; the conflict resolution strategy can be to resolve conflicts based on rules or confidence levels when there are contradictions in the results; and the complementary enhancement strategy can be to integrate complementary information to form a more comprehensive conclusion.

[0043] Finally, the system generates interpretable output, which includes a description of the fusion process (contribution of each subtask, how the fusion is performed, and uncertainties). The output results are as follows: Recommendations for comprehensive treatment of lung cancer patients Overall assessment conclusion: The patient has stage IIIA lung adenocarcinoma, is EGFR mutation positive, has mild to moderate COPD, and requires careful protection of lung function.

[0044] Recommended treatment plan: (1) Surgical treatment: Thoracoscopic sublobar resection (preserving more lung function) + lymph node dissection is recommended. (2) Postoperative adjuvant therapy: Based on the positive EGFR mutation, it is recommended to perform EGFR-TKI targeted therapy for 2 years after surgery. (3) Alternative options: If the patient refuses surgery or the surgical risk is too high, stereotactic body radiotherapy (SBRT) may be considered. Reasoning path (explainability): (1) Surgical feasibility is based on imaging localization and lung function assessment (T1+T3→T4) (2) Targeted therapy recommendations are based on EGFR mutation detection (T2→T5) (3) Radiotherapy alternatives based on precise image localization (T1→T6) (4) The final plan integrates the professional opinions of surgery, targeted therapy and radiotherapy (T4+T5+T6→T7) Uncertainty warning: (1) There are individual differences in the actual recovery of lung function after surgery. (2) The long-term efficacy of targeted therapy requires regular follow-up assessments. The consultation team received the comprehensive plan and interpretability explanation generated by the system as a basis for discussion. During the consultation, the expert team confirmed the plan's rationality and supplemented it with suggestions for nutritional support and psychological counseling. These supplementary suggestions served as feedback input to the system, used to optimize future task decomposition and integration strategies. Through this system, data collection and preliminary plan formulation, which previously required several days, can now be completed within hours; the collaborative work of various professional intelligent agents ensured the comprehensiveness and professionalism of the plan; and the final interpretable output enabled the expert team to quickly understand and verify the AI's suggestions, truly achieving precision medicine through human-machine collaboration.

[0045] In summary, the multi-agent collaborative task decomposition and execution method provided in this application, compared with the prior art, can receive the description information of the target medical task, perform deep semantic parsing on the description information to extract key task elements; generate a candidate sub-task list required to complete the target medical task based on the medical domain knowledge base and key task elements, and parse the logical dependencies between each sub-task in the candidate sub-task list; construct a directed acyclic graph of tasks containing the candidate sub-task list based on the logical dependencies; and divide each sub-task in the directed acyclic graph of tasks according to the real-time load of multiple agent models. The target intelligent agent model with matching capabilities and reasonable load is allocated for execution; the execution status of each subtask is monitored in real time. If the execution status is that the subtask is completed, the dependency conditions of its subsequent subtasks are checked based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtasks are triggered. If the execution status is that the execution is abnormal, the abnormal subtasks are dynamically rescheduled for execution. The dependency condition is that all the preceding subtasks required by the subsequent subtasks have been completed. After all subtasks have been executed, the scattered execution results of multiple intelligent agent models are collected, and the scattered execution results are integrated into a unified output result according to the fusion strategy.

[0046] The solution proposed in this application enables AI to automatically understand task intent, automatically generate sub-tasks by combining with a medical knowledge base, and automatically identify prerequisite dependencies, parallel possibilities, and information flow dependencies, transforming them into a structured DAG graph. This not only frees medical staff from tedious process arrangement but also fundamentally eliminates task decomposition errors and sequence confusion caused by limitations or negligence in human cognition.

[0047] This application does not simply dump tasks onto the intelligent agent, but strictly follows the DAG topology. The subsequent task is only automatically triggered when all the preceding subtasks required by the subsequent subtask are completed (information flow is complete). This completely avoids the logical collapse and medical risks caused by blindly executing the subsequent task before the preceding results are available, and ensures the rigor and safety of executing complex clinical pathways.

[0048] In real-world medical computing environments, agents may malfunction or time out. This application can detect anomalies in real time and automatically trigger a rescheduling mechanism (such as assigning to a backup agent), achieving process self-healing without human intervention. This effectively prevents the cascading risk of local anomalies spreading to the global system and significantly reduces the probability of medical safety incidents caused by system failures.

[0049] Based on the above Figure 2 The specific implementation of the method shown in this embodiment provides a multi-agent cooperative task decomposition and execution device, such as... Figure 4As shown, the device includes: a receiving module 31, a generating module 32, a building module 33, an allocation module 34, a monitoring module 35, and an integration module 36; The receiving module 31 is used to receive the description information of the target medical task and perform deep semantic parsing processing on the description information to extract key task elements. The generation module 32 is used to generate a list of candidate subtasks required to complete the target medical task based on the medical field knowledge base and the key task elements, and to parse the logical dependencies between the subtasks in the candidate subtask list. Module 33 is used to construct a directed acyclic graph of tasks containing the list of candidate subtasks based on the logical dependencies. The allocation module 34 is used to allocate each subtask in the directed acyclic graph of the task to the target intelligent agent model with matching capabilities and reasonable load for execution, based on the multiple intelligent agent models and the corresponding real-time load of the multiple intelligent agent models. The monitoring module 35 is used to monitor the execution status of each subtask in real time. If the execution status is a subtask completion status, it checks whether the dependency conditions of its subsequent subtasks are met based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is an execution exception status, the execution of the subtask with execution exception is dynamically rescheduled. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. The integration module 36 is used to collect the scattered execution results of the multiple intelligent agent models after all the sub-tasks have been executed, and to integrate the scattered execution results into a unified output result according to the fusion strategy.

[0050] In a specific application scenario, the receiving module 31 can be used to perform deep semantic parsing processing on the description information using a large language model, and identify the parsing results of the target medical task, wherein the parsing results include core objectives, key constraints and expected outputs; and extract the key task elements from the parsing results, wherein the key task elements include patient characteristics, disease information, task objectives and constraints.

[0051] In specific application scenarios, the logical dependencies include prerequisite dependencies, parallel execution relationships, and information flow dependencies; the generation module 32 can be used to determine the logical dependencies between each subtask in the candidate subtask list as prerequisite dependencies based on the execution timing constraints and logical order of subtasks in the medical clinical pathway. Based on the data interaction requirements and mutual independence between subtasks, the logical dependencies between the subtasks in the candidate subtask list are determined to be parallel execution relationships. Based on the correspondence between the input data of the subsequent subtask and the output data of the preceding subtask, the logical dependency relationship between each subtask in the candidate subtask list is determined to be an information flow dependency relationship.

[0052] In specific application scenarios, the construction module 33 can be used to construct a directed acyclic graph of tasks based on the logical dependencies, which includes all subtask nodes in the candidate subtask list, dependency edges between nodes, critical paths between nodes, and a set of nodes that can be executed in parallel, and to visualize the directed acyclic graph of tasks.

[0053] In specific application scenarios, the allocation module 34 can be used to match each subtask in the directed acyclic graph of the task with the agent capability catalog to filter out a list of candidate agents with corresponding capabilities and level requirements. The current load and estimated completion time of each intelligent agent model are monitored in real time, and an intelligent agent model with matching capabilities and reasonable load is selected from the candidate intelligent agent list as the target intelligent agent model. Based on the urgency of the task or a preset scheduling strategy, the shortest waiting time priority strategy, the highest capability level priority strategy, or the load balancing priority strategy are dynamically selected to assign subtasks to the target intelligent agent model.

[0054] In specific application scenarios, the monitoring module 35 can be used to re-match candidate intelligent agent models from the intelligent agent capability catalog if the execution status is an execution abnormal status. The output of the subtask that failed to execute and its required preceding subtasks are redistributed to the candidate agent model for continued execution, and the state of the directed acyclic graph of the task is updated synchronously. Among them, abnormal execution states include at least the failure of subtask execution, delay exceeding a preset threshold, or failure of the agent model executing the subtask.

[0055] In specific application scenarios, the integration module 36 can be used to select the corresponding integration strategy from the integration strategy library according to the task type, and integrate the scattered execution results into a unified output result according to the integration strategy. The integration strategy includes at least the sequential splicing strategy, the weighted synthesis strategy, the conflict resolution strategy, and the complementary enhancement strategy.

[0056] Specific limitations regarding the multi-agent cooperative task decomposition and execution device can be found in the limitations of the multi-agent cooperative task decomposition and execution method described above, and will not be repeated here. Each module in the aforementioned multi-agent cooperative task decomposition and execution device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0057] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a multi-agent cooperative task decomposition and execution method on the server side.

[0058] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the client-side functions or steps of a multi-agent cooperative task decomposition and execution method.

[0059] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Receive description information of the target medical task, and perform deep semantic parsing on the description information to extract key task elements; Based on the medical knowledge base and the key task elements, a list of candidate subtasks required to complete the target medical task is generated, and the logical dependencies between the subtasks in the candidate subtask list are resolved. Based on the logical dependencies, construct a directed acyclic graph of tasks containing the list of candidate subtasks; Based on multiple agent models and their corresponding real-time load conditions, each subtask in the directed acyclic graph of the task is assigned to a target agent model with matching capabilities and reasonable load for execution. The execution status of each subtask is monitored in real time. If the execution status is that the subtask is completed, the dependency conditions of its subsequent subtasks are checked based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is that the execution is abnormal, the subtask with execution abnormality is dynamically rescheduled for execution. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. After all the subtasks have been completed, the scattered execution results of the multiple agent models are collected, and the scattered execution results are integrated into a unified output result according to the fusion strategy.

[0060] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Receive description information of the target medical task, and perform deep semantic parsing on the description information to extract key task elements; Based on the medical knowledge base and the key task elements, a list of candidate subtasks required to complete the target medical task is generated, and the logical dependencies between the subtasks in the candidate subtask list are resolved. Based on the logical dependencies, construct a directed acyclic graph of tasks containing the list of candidate subtasks; Based on multiple agent models and their corresponding real-time load conditions, each subtask in the directed acyclic graph of the task is assigned to a target agent model with matching capabilities and reasonable load for execution. The execution status of each subtask is monitored in real time. If the execution status is that the subtask is completed, the dependency conditions of its subsequent subtasks are checked based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is that the execution is abnormal, the subtask with execution abnormality is dynamically rescheduled for execution. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. After all the subtasks have been completed, the scattered execution results of the multiple agent models are collected, and the scattered execution results are integrated into a unified output result according to the fusion strategy.

[0061] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0062] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0064] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0065] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for multi-agent cooperative task decomposition and execution, characterized in that, The method includes: Receive description information of the target medical task, and perform deep semantic parsing on the description information to extract key task elements; Based on the medical knowledge base and the key task elements, a list of candidate subtasks required to complete the target medical task is generated, and the logical dependencies between the subtasks in the candidate subtask list are resolved. Based on the logical dependencies, construct a directed acyclic graph of tasks containing the candidate subtask list; Based on multiple agent models and their corresponding real-time load conditions, each subtask in the directed acyclic graph of the task is assigned to a target agent model with matching capabilities and reasonable load for execution. The execution status of each subtask is monitored in real time. If the execution status is that the subtask is completed, the dependency conditions of its subsequent subtasks are checked based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is that the execution is abnormal, the subtask with execution abnormality is dynamically rescheduled for execution. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. After all the subtasks have been completed, the scattered execution results of the multiple agent models are collected, and the scattered execution results are integrated into a unified output result according to the fusion strategy.

2. The multi-agent cooperative task decomposition and execution method according to claim 1, characterized in that, The deep semantic parsing process performed on the description information to extract key task elements specifically includes: The descriptive information is processed by deep semantic parsing using a large language model to identify the parsing results of the target medical task, wherein the parsing results include the core objective, key constraints and expected output; The key task elements are extracted from the analysis results, wherein the key task elements include patient characteristics, disease information, task objectives, and constraints.

3. The multi-agent cooperative task decomposition and execution method according to claim 1, characterized in that, The logical dependencies include prerequisite dependencies, parallel execution dependencies, and information flow dependencies; The process of parsing the logical dependencies between the subtasks in the candidate subtask list specifically includes: Based on the execution timing constraints and logical order of subtasks in the medical clinical pathway, the logical dependencies between the subtasks in the candidate subtask list are determined as prerequisite dependencies. Based on the data interaction requirements and mutual independence between subtasks, the logical dependencies between the subtasks in the candidate subtask list are determined to be parallel execution relationships. Based on the correspondence between the input data of the subsequent subtask and the output data of the preceding subtask, the logical dependency relationship between each subtask in the candidate subtask list is determined to be an information flow dependency relationship.

4. The multi-agent cooperative task decomposition and execution method according to claim 1, characterized in that, The construction of a directed acyclic graph of tasks containing the candidate subtask list based on the logical dependencies specifically includes: Based on the logical dependencies, a directed acyclic graph of tasks is constructed, which includes all subtask nodes in the candidate subtask list, dependency edges between nodes, critical paths between nodes, and a set of nodes that can be executed in parallel. The directed acyclic graph of tasks is then visualized.

5. The multi-agent cooperative task decomposition and execution method according to claim 1, characterized in that, The step of assigning each subtask in the directed acyclic graph of the task to a target agent model with matching capabilities and reasonable load, based on multiple agent models and their corresponding real-time load conditions, specifically includes: Each subtask in the directed acyclic graph of the task is matched with the agent capability catalog to filter out a list of candidate agents with the corresponding capabilities and level requirements. The current load and estimated completion time of each intelligent agent model are monitored in real time, and an intelligent agent model with matching capabilities and reasonable load is selected from the candidate intelligent agent list as the target intelligent agent model. Based on the urgency of the task or a preset scheduling strategy, the shortest waiting time priority strategy, the highest capability level priority strategy, or the load balancing priority strategy are dynamically selected to assign subtasks to the target intelligent agent model.

6. The multi-agent cooperative task decomposition and execution method according to claim 5, characterized in that, If the execution state is an execution exception state, then the subtasks with execution exceptions are dynamically rescheduled for execution, specifically including: If the execution status is an execution abnormality, then a candidate agent model is re-matched from the agent capability catalog; The output of the subtask that failed to execute and its required preceding subtasks are redistributed to the candidate agent model for continued execution, and the state of the directed acyclic graph of the task is updated synchronously. Among them, abnormal execution states include at least the failure of subtask execution, delay exceeding a preset threshold, or failure of the agent model executing the subtask.

7. The multi-agent cooperative task decomposition and execution method according to claim 1, characterized in that, The process of integrating the distributed execution results into a unified output result according to the fusion strategy specifically includes: According to the task type, a corresponding fusion strategy is selected from the fusion strategy library, and the scattered execution results are integrated into a unified output result according to the fusion strategy. The fusion strategy includes at least a sequential splicing strategy, a weighted synthesis strategy, a conflict resolution strategy, and a complementary enhancement strategy.

8. A multi-agent collaborative task decomposition and execution device, characterized in that, include: The receiving module is used to receive the description information of the target medical task and perform deep semantic parsing processing on the description information to extract key task elements. The generation module is used to generate a list of candidate subtasks required to complete the target medical task based on a medical knowledge base and the key task elements, and to parse the logical dependencies between the subtasks in the candidate subtask list. A construction module is used to construct a directed acyclic graph of tasks containing the list of candidate subtasks based on the logical dependencies. The allocation module is used to allocate each subtask in the directed acyclic graph of the task to the target intelligent agent model with matching capabilities and reasonable load for execution, based on multiple intelligent agent models and the corresponding real-time load of the multiple intelligent agent models. The monitoring module is used to monitor the execution status of each subtask in real time. If the execution status is a subtask completion status, it checks whether the dependency conditions of its subsequent subtasks are met based on the dependency edges in the directed acyclic graph of the task. If they are met, the scheduling and execution of the subsequent subtask is triggered. If the execution status is an execution exception status, the execution of the subtask with execution exception is dynamically rescheduled. The dependency condition is that all the preceding subtasks required by the subsequent subtask have been completed. The integration module is used to collect the scattered execution results of the multiple intelligent agent models after all the sub-tasks have been executed, and to integrate the scattered execution results into a unified output result according to the fusion strategy.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent cooperative task decomposition and execution method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-agent cooperative task decomposition and execution method as described in any one of claims 1 to 7.

Images