Multi-agent scheduling method and system based on task intention matching

Abstract

The invention relates to a multi-agent scheduling method and system based on task intention matching. The method comprises the following steps that S1, tasks are received and analyzed; s2, intelligent agent matching candidates are obtained; s3, selecting an optimal agent by a scoring mechanism; s4, task assignment, execution monitoring and result acquisition; s5, performing reflection evaluation and strategy updating: dynamically updating an intelligent agent score, and adjusting a task label matching rule or other scheduling parameters so as to optimize a distribution decision of a subsequent task; and S6, result combination and output: when the task comprises a plurality of sub-tasks executed by a plurality of intelligent agents, the results are verified, sorted and combined. Compared with the prior art, the method has the advantages that the task target can be dynamically analyzed, the optimal agent can be intelligently matched to execute the task, and the scheduling strategy is continuously optimized through feedback after the task is completed, so that the task execution efficiency and effect of the multi-agent system are remarkably improved, and the accuracy, the adaptability and the intelligent level of task allocation are improved.

Description

Technical Field

[0001] This invention relates to the field of task scheduling technology in the field of artificial intelligence, and in particular to a multi-agent scheduling method and system based on task intent matching. Background Technology

[0002] With the development of artificial intelligence and robotics, multi-agent systems are widely used in industrial production, service robots, distributed computing, and other fields. In these systems, assigning appropriate tasks to the most suitable agents is key to improving overall system efficiency. However, existing task scheduling and allocation methods still have many shortcomings.

[0003] Traditional task scheduling schemes often rely on pre-defined static rules or simple task type matching. For example, in a scheduling system with fixed rules, each task type is pre-assigned to a specific agent or a fixed sequence, lacking the ability to dynamically adjust according to actual circumstances. When the task environment or task requirements change, static rules struggle to respond promptly, resulting in inflexible task allocation. While some scheduling systems can select agents based on task type to some extent, they often only match based on the superficial category of the task, failing to deeply understand the task intent and thus making it difficult to guarantee that the selected agent is truly adept at performing that specific task.

[0004] Furthermore, in multi-agent collaborative work, different agents may possess their own strengths and weaknesses (i.e., heterogeneity). Some existing task allocation methods do not adequately consider the differences in abilities between agents, potentially simply distributing tasks evenly or based solely on idle status. This lack of precise matching may result in some tasks being assigned to agents that are not optimally suited, impacting task completion quality and efficiency. For example, a highly specialized task, if assigned to an agent with average abilities, may execute slowly or err, while a more suitable agent with specialized skills may not be selected—a common problem in existing scheduling mechanisms.

[0005] On the other hand, most current multi-agent task scheduling systems lack a feedback learning mechanism after task execution. Typically, after a task is completed, the system does not conduct in-depth analysis and utilization of the agent's performance, nor does it adjust the scheduling strategy based on past successes or failures. This means that regardless of whether an agent has previously performed poorly on similar tasks, the system may continue to assign it the same type of task, failing to achieve intelligent scheduling that enables continuous improvement.

[0006] In summary, existing technologies in the field of multi-agent task scheduling have the following problems: lack of in-depth understanding of task intent, and inaccurate matching between tasks and agents; static and rigid scheduling strategies that cannot adapt to dynamically changing task requirements; insufficient consideration of heterogeneous agent performance leading to suboptimal task allocation; and lack of a reflection and evaluation mechanism after task completion, making it impossible to optimize future scheduling strategies based on historical performance. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a multi-agent scheduling method and system based on task intent matching. This method can dynamically parse task objectives, intelligently match the optimal agent to execute the task, and continuously optimize the scheduling strategy through feedback after the task is completed. This significantly improves the task execution efficiency and effectiveness of the multi-agent system, and enhances the accuracy, adaptability, and intelligence of task allocation.

[0008] This invention provides a multi-agent scheduling method based on task intent matching, comprising the following steps:

[0009] S1: Task Reception and Parsing: Obtain the description information of the task to be executed, the task's objective and implicit intent, perform natural language processing on the task description information, and generate structured task tags containing the task intent; structured task tags include environmental monitoring, temperature monitoring, alarm capabilities, etc. Structured task tags are generated through a predefined domain knowledge base.

[0010] S2: Agent matching candidate acquisition: Based on the structured task labels, select candidate agents with corresponding capabilities or attributes from the pre-registered set of multiple agents;

[0011] S3: Scoring mechanism to select the optimal agent: For the selected candidate agents, their ability to perform the task is scored based on a predetermined scoring mechanism. The scoring mechanism considers historical task performance, current state, response ability, and skill fit, and selects the agent with the highest score as the optimal agent.

[0012] S4: Task Assignment, Execution Monitoring and Result Acquisition: The selected optimal agent executes the task, and monitors the task progress and important status data during the execution process; after the task is completed, the execution results and process data are collected.

[0013] S5: Reflection Evaluation and Strategy Update: Based on the collected execution results and process data, reflect on and evaluate the performance of the agent, and feed the evaluation results back to the scoring mechanism and scheduling strategy to dynamically update the agent score, adjust the task label matching rules or other scheduling parameters, thereby optimizing the allocation decision of subsequent tasks.

[0014] S6: Result merging and output: When a task contains multiple subtasks executed by multiple agents, the execution results of each subtask are collected, and the results are verified, organized, and merged to obtain the complete task output result.

[0015] Furthermore, in S1, semantic analysis technology is used to extract task target keywords. Through knowledge base mapping, the task description is converted into structured task information, and structured task tags that can represent task requirements are extracted. The structured task tags include: task type (e.g., navigation, grasping, data processing, etc.), required skills or resources (e.g., high-speed computing, robotic arm operation), urgency, and accuracy requirements.

[0016] Furthermore, in S2, in the pre-registered set of multiple agents, each agent is pre-associated with capability tags, proficient task types, and current state information. The filtering process includes comparing task tags with agent capability tags to select candidate agents that meet the task requirements.

[0017] Furthermore, in S3, the scoring mechanism comprehensively considers factors including: the success rate and quality of each agent in completing similar tasks in the past, the current workload and availability, response speed, and context-related capability indicators; by scoring each candidate agent, the system can quantitatively evaluate the expected performance of each agent in performing the task, and select the agent with the highest score as the optimal agent.

[0018] Furthermore, in S3, the scoring mechanism includes the following sub-modules: a performance evaluation module for recording the historical task success rate and quality evaluation of each agent; a load monitoring module for monitoring the real-time status of agents; and a scoring calculation module for calculating the agent's task adaptation score based on the above information. The scoring mechanism sorts candidate agents according to their scores, prioritizing the agent with the highest score for task execution. Compared to traditional methods that select agents based on only a single indicator (such as idle state or distance), the scoring selection method of this invention is more comprehensive and objective, ensuring that the selected agent is the most capable of efficiently completing the task.

[0019] Furthermore, in S5, the reflective evaluation includes: result achievement, efficiency and quality, and agent performance. Specifically, it involves comparing the actual results with the expected goals after task completion, recording key performance indicators and anomalies encountered by the agent during execution, and analyzing and providing feedback on the execution process, including: whether the task was completed on time and with high quality, whether errors or delays occurred during the process, and whether the agent's actions effectively fulfilled the task intent. Through this reflective evaluation, the system obtains feedback on the effectiveness of the scheduling decision. This feedback is used to update the agent's performance records and scoring parameters, and to adjust and optimize future scheduling strategies.

[0020] Further adjustments and optimizations to future scheduling strategies include: updating agent rating parameters (if an agent efficiently completes a certain type of task multiple times, its corresponding task label matching priority can be increased; otherwise, it can be decreased), refining task label definitions or weights, and improving the parameter configuration of task matching and rating algorithms; when consecutive task failures or inefficient executions are detected, adjusting the rules for matching task labels with agents or increasing the rating weights of other candidate agents in similar tasks, in order to achieve adaptive optimization of the scheduling strategy.

[0021] In this way, the system can achieve adaptive scheduling through continuous learning, making task allocation decisions more and more intelligent over time, and continuously improving overall performance.

[0022] Furthermore, in S6, the result merging output includes: for multiple subtasks executed in parallel or step-by-step, collecting all subtask results according to task dependencies; performing format standardization and integrity verification on the result data, and integrating multiple subtask results into the final result using preset merging rules; when a conflict or inconsistency in subtask results is detected, triggering an exception handling process or feeding back to the relevant agent for re-execution to ensure the consistency and accuracy of the output results.

[0023] When a main task is broken down into several subtasks executed by different agents, the system collects the results returned by each agent after all subtasks are completed. Using predefined result integration rules or algorithms (e.g., merging data according to task dependencies, or verifying and summarizing partial results from different sources), the scattered subtask results are synthesized into a complete output. For example, if the task is "warehouse order processing," with one agent responsible for inventory counting and another for scheduling shipments, the system will obtain the inventory list and shipping plan separately, and integrate them into a comprehensive report of completed order processing during the result merging phase. This result merging strategy ensures that multi-agent collaborative tasks produce consistent and comprehensive final results and provide users with unified task completion feedback.

[0024] Furthermore, the multiple intelligent agents 104-1 to 104-n are entities that specifically perform tasks; they are robots, software services, or other units with autonomous decision-making and execution capabilities. Each intelligent agent communicates with the scheduling decision module through an interface, receives task instructions, and reports execution progress and results.

[0025] The present invention also provides a multi-agent scheduling system based on task intent matching, comprising:

[0026] The task management module is used to receive task requests, parse task intents, and generate structured task tags.

[0027] The agent resource library stores the capability tags, status information, and historical performance data of multiple agents;

[0028] The scheduling decision module matches candidate agents in the agent resource library according to the task tags, and selects the optimal agent to execute the task through a built-in scoring mechanism. It monitors the task execution process and performs reflection and evaluation after the task is completed to update the scheduling strategy.

[0029] The result merging module is used to integrate the task execution results of multiple agents when they work together to complete a task and output the final result.

[0030] The entire system architecture supports modular expansion and information sharing. The modules interact through well-defined interface protocols and execute the multi-agent scheduling method based on task intent matching, thus enabling closed-loop management of tasks from receipt to completion.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] (1) Accurate Task Intent Analysis and Efficient Matching. By combining natural language processing with a domain knowledge base, deep semantic analysis of unstructured task descriptions is achieved, generating a multi-dimensional task tag system, which effectively improves the accuracy and standardization of task requirement understanding. Through the matching of task tags with agent capabilities and the optimized selection of scoring mechanisms, efficient matching between tasks and executing agents is achieved.

[0033] (2) Multi-dimensional intelligent evaluation system. By introducing post-task reflection evaluation, a closed-loop improvement system for scheduling strategies is established, which can continuously optimize based on historical experience.

[0034] (3) By merging the results of multiple tasks, it supports the collaborative work of intelligent agents in complex tasks and ensures the consistency of results. It improves the task response capability, resource utilization and collaborative work efficiency of multi-agent systems in dynamic environments, and has broad application prospects. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the multi-agent scheduling system based on task intent matching in Example 1;

[0036] Figure 2 This is a flowchart of the multi-agent scheduling method based on task intent matching in Example 1;

[0037] Figure 3 This is a schematic diagram illustrating the merging of the results of multiple intelligent agents collaborating to complete a task in Example 1; Detailed Implementation

[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0039] Example 1

[0040] This embodiment provides a multi-agent scheduling system based on task intent matching, such as... Figure 1 As shown, it includes:

[0041] The task management module is used to receive task requests, parse task intents, and generate structured task tags.

[0042] The agent resource library stores the capability tags, status information, and historical performance data of multiple agents;

[0043] The scheduling decision module matches candidate agents in the agent resource library according to the task tags, and selects the optimal agent to execute the task through a built-in scoring mechanism. It monitors the task execution process and performs reflection and evaluation after the task is completed to update the scheduling strategy.

[0044] The result merging module is used to integrate the task execution results of multiple agents when they work together to complete a task and output the final result.

[0045] The entire system architecture supports modular expansion and information sharing. The modules interact through well-defined interface protocols and execute a multi-agent scheduling method based on task intent matching, so that the task can be managed in a closed loop from receipt to completion.

[0046] like Figure 2 As shown, the system executes the multi-agent scheduling method based on task intent matching in the following steps:

[0047] S1: Task Reception and Parsing: Obtain the description information of the task to be executed, the task's objective and implicit intent, perform natural language processing on the task description information, and generate structured task tags containing the task intent; structured task tags include environmental monitoring, temperature monitoring, alarm capabilities, etc. Structured task tags are generated through a predefined domain knowledge base.

[0048] In a specific implementation, when a new task arrives, the task management module first receives the task request. The task request may include a task description, objective requirements, timeframe, and other constraints. The module parses the task description, for example, using natural language processing algorithms to transform the text description into a structured task representation. By accessing a domain knowledge base or rule base, it identifies the task's intent and key parameters, outputting the task objective and preliminary task tags. For example, for an input task "Please send a robot to check the temperature in warehouse A and issue an alarm if it exceeds the limit," after parsing, the intent can be extracted as "environmental monitoring," the location as "warehouse A," and the operations as "temperature detection" and "alarm." Therefore, task tags such as {environmental monitoring, temperature detection, alarm, warehouse A} are generated.

[0049] S2: Agent matching candidate acquisition: Based on the structured task labels, select candidate agents with corresponding capabilities or attributes from the pre-registered set of multiple agents;

[0050] In a specific implementation, the scheduling decision module queries the agent resource library for a list of agents with relevant capabilities based on the task tags passed by the task management module. The agent resource library pre-stores a set of capability tags for each agent; for example, an agent might have capability tags such as {environmental sensing, temperature measurement, and audible / visual alarm}. The system compares the task tags with the agent capability tags, filtering out all agents containing the key capabilities required for the task as a candidate agent set. In the above example, if a robot possesses an environmental temperature sensor and an alarm device, and its capability tags match the task requirements, then that robot enters the candidate list. If the task allows multiple agents to collaborate (e.g., the task can be divided), this step can also determine the candidate agent set for each sub-task based on the task division.

[0051] S3: Scoring mechanism to select the optimal agent: For the selected candidate agents, their ability to perform the task is scored based on a predetermined scoring mechanism. The scoring mechanism considers historical task performance, current state, response ability, and skill fit, and selects the agent with the highest score as the optimal agent.

[0052] In a specific implementation, for the obtained set of candidate agents, the scheduling decision module uses a built-in scoring mechanism to evaluate and score each candidate agent. The scoring mechanism considers the following factors:

[0053] Historical performance: Indicators such as the success rate, accuracy, and time taken by the agent in performing similar tasks in the past. If an agent has extensive experience and excellent performance in related tasks, it will receive a higher score.

[0054] Current status: The agent's current workload, remaining resources (such as battery power and memory), and task-related readiness. If the agent is currently idle and in good condition, the score increases; conversely, if it is busy or has limited resources, the score decreases.

[0055] Response capability: Estimate the speed of response time based on the distance between the agent and the task location or data source (for physical entities) or communication latency (for software agents). The faster the response, the higher the score.

[0056] Skill fit: The degree to which the agent's set of abilities matches the task requirements. If the agent's skills fully cover the task requirements and have surplus abilities, it indicates a high fit, which can increase the score.

[0057] The scheduling and decision-making module calculates a comprehensive score for each candidate agent based on the above factors according to certain weights. Then, the candidate list is sorted from highest to lowest score, and the agent with the highest score is selected as the executor of the task. In some cases, if the score results are not significantly different, the system can also consider redundancy or fault-tolerance strategies, such as selecting the top two ranked agents to execute the task together or using the second-highest ranked agent as a backup in case the primary agent fails midway. However, usually, the agent with the highest score is considered the optimal executor.

[0058] S4: Task Assignment, Execution Monitoring and Result Acquisition: The selected optimal agent executes the task, and monitors the task progress and important status data during the execution process; after the task is completed, the execution results and process data are collected.

[0059] In this specific implementation, after determining the optimal agent, the scheduling decision module sends task instructions and relevant parameters to that agent (e.g., agent 104-2) via a communication interface. Upon receiving the task, the agent enters the execution state and performs tasks according to the requirements. During this process, the agent periodically or when critical events occur, reports its progress to the scheduling module, including start time, interim results, and difficulties encountered, so that the system can understand the task's running status. If an abnormal situation occurs during execution (e.g., equipment failure, inability to complete a sub-step), the agent can request support or task reassignment. In this case, the scheduling module can select a backup agent to take over the remaining tasks or take other remedial measures as needed. This monitoring and feedback design of the execution process ensures the smooth progress of the task and enables timely response to unexpected situations.

[0060] Once the agent completes its task, it submits the final result or execution status report to the scheduling decision module. The result may take the form of sensor data, confirmation of operation completion, generated files, or any output related to the task objective. After receiving the task result, the scheduling module forwards it to the task management module for recording and triggers the next evaluation process.

[0061] S5: Reflection Evaluation and Strategy Update: Based on the collected execution results and process data, reflect on and evaluate the performance of the agent, and feed the evaluation results back to the scoring mechanism and scheduling strategy to dynamically update the agent score, adjust the task label matching rules or other scheduling parameters, thereby optimizing the allocation decision of subsequent tasks.

[0062] In a specific implementation, the task management module and the scheduling decision module collaboratively analyze the following:

[0063] Result Achievement: Compare the final result submitted by the agent with the initial objective requirements of the task to check whether the task intent has been met. For example, does the alarm task correctly trigger when the temperature exceeds the limit, and does the result meet expectations?

[0064] Efficiency and Quality: Assess whether the actual time taken for the statistical task is within the required range, whether resource consumption is reasonable, and whether there are any deviations or errors in the execution process. Record any deviations (e.g., how much time was delayed, how much difference was made in data accuracy).

[0065] Agent performance: Evaluating the agent's behavior during execution, such as whether autonomous decisions are appropriate, whether cooperation with other system components is smooth, and whether predetermined procedures are followed.

[0066] Based on the above assessment, the system updates the agent's historical performance data. For example, if the agent completes the task successfully and efficiently, a positive case is recorded in its capability profile, increasing its reputation score for the relevant task type. If a problem is found, such as the agent failing to detect excessive temperature leading to an alarm delay, a negative case is recorded, potentially reducing its scoring weight in emergency monitoring tasks. In addition to updating the agent's own data, the scheduling decision module also checks the applicability of the current scheduling strategy: Did the problems in this task flow stem from imperfections in task matching or the scoring process? If so, the algorithm parameters or rules are adjusted accordingly, such as enriching task labels to more accurately describe the task scenario, modifying the scoring formula to improve sensitivity to certain key factors, or introducing more monitoring nodes in future similar tasks. Through this iterative reflective learning process, the system's task scheduling strategy will gradually optimize, becoming increasingly reliable and efficient.

[0067] S6: Merge and output the results: as shown in Figure 3 As shown, when a task contains multiple subtasks executed by multiple agents, the execution results of each subtask are collected, and the results are verified, organized, and merged to obtain a complete task output result.

[0068] In a specific implementation, within a nuclear industry safety monitoring task, the system receives real-time data (such as radiation levels, temperature, and pressure) from various monitoring sensors within the nuclear facility. When the monitored data exceeds a preset safety threshold, the system automatically triggers a series of subsequent agents to execute corresponding tasks.

[0069] Data Analysis Agent: First, the central scheduling module matches an Agent with data analysis capabilities based on the task intent ("anomaly monitoring") to further verify and analyze the triggered monitoring data to confirm whether there are any potential security risks.

[0070] Report Generation Agent: After the data analysis agent completes the data processing, its results will be passed to the report generation agent. This agent is responsible for compiling the monitoring data and analysis results into a detailed security monitoring report, which includes specific indicators of abnormal data, analysis results, and recommended measures.

[0071] Notification Agent: At the same time, to ensure that safety hazards can be dealt with in a timely manner, the system also triggers a notification agent. Based on the results generated by the report, the agent sends an early warning notification to the nuclear facility safety control center, operation and maintenance personnel, etc., to ensure that relevant personnel can take emergency measures quickly.

[0072] Results Merging: Finally, the system integrates the data analysis results, generated safety reports, and notification status information through the results merging module to form a complete monitoring report. This report not only displays detailed information on monitoring data and anomalies but also includes the execution status of report generation and early warning notifications, providing comprehensive information support for the safety management of nuclear facilities.

[0073] Through the above process, the system realizes closed-loop management from automatic triggering of monitoring data, intelligent agent matching and division of labor, to reflection and evaluation after task execution and result merging, ensuring timely and accurate detection of anomalies in nuclear industry scenarios, and rapid generation of reports and notifications, thereby improving the overall effectiveness of nuclear safety management.

[0074] Tests were conducted in a simulated environment. Results show that after introducing task intent matching and reflective optimization mechanisms, the system's task allocation accuracy and efficiency are significantly improved. Compared to traditional static rule allocation, this invention can select more suitable agents based on task requirements, reducing task completion time by an average of approximately 20% and increasing task success rate by approximately 15%. In continuously executed task sequences, the system exhibits a gradual optimization trend because the scheduling strategy adjusts based on feedback from previous tasks: for example, for specific types of tasks, it can adaptively identify which agent is more proficient, thus increasingly assigning such tasks to that agent while reducing task assignments to unsuitable agents. This self-learning characteristic enables the system to maintain high efficiency over long-term operation.

[0075] Meanwhile, the architecture of this invention exhibits excellent scalability and robustness. When a new agent joins the system, its capability tags and initial scoring information are simply added to the agent resource library, and the scheduling module automatically includes it in the candidate pool for task allocation. The modular design of the scheduling algorithm also facilitates parameter adjustment or functional expansion according to different application scenarios; for example, it can be integrated with machine learning models to further optimize the scoring mechanism. In practical applications, this invention can be widely used in scenarios requiring multi-agent collaborative work, including multi-robot scheduling in industrial production lines, task allocation for unmanned vehicle fleets in warehousing and logistics, distributed intelligent inspection in smart cities, and task scheduling in large-scale distributed computing systems. Through a deep understanding of task intent and efficient utilization of agent resources, this invention helps achieve more intelligent automated system management, improves the execution efficiency and quality of complex tasks, and provides strong support for the development of related technical fields.

[0076] Components not described in detail in this embodiment are all existing components that can be purchased through public channels.

[0077] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A multi-agent scheduling method based on task intent matching, characterized in that, Includes the following steps: S1: Task Reception and Parsing: Obtain the description information of the task to be executed, perform natural language processing on the task description information, and generate a structured task tag containing the task intent; S2: Agent matching candidate acquisition: Based on the structured task labels, select candidate agents with corresponding capabilities or attributes from the pre-registered set of multiple agents; S3: Scoring mechanism to select the optimal agent: For the selected candidate agents, their ability to perform the task is scored based on a predetermined scoring mechanism. The scoring mechanism considers historical task performance, current state, response ability, and skill fit, and selects the agent with the highest score as the optimal agent. S4: Task Assignment, Execution Monitoring and Result Acquisition: The selected optimal agent executes the task, and monitors the task progress and important status data during the execution process; after the task is completed, the execution results and process data are collected. S5: Reflection Evaluation and Strategy Update: Based on the collected execution results and process data, reflect on and evaluate the performance of the agent, and feed the evaluation results back to the scoring mechanism and scheduling strategy to dynamically update the agent score, adjust the task label matching rules or other scheduling parameters, thereby optimizing the allocation decision of subsequent tasks. S6: Result merging and output: When a task contains multiple subtasks executed by multiple agents, the execution results of each subtask are collected, and the results are verified, organized, and merged to obtain the complete task output result.

2. The multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, In S1, semantic analysis technology is used to extract task target keywords. Through knowledge base mapping, the task description is converted into structured task information, and structured task tags that can represent task requirements are extracted. The structured task tags include: task type, required skills or resources, urgency level, and accuracy requirements.

3. The multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, In S2, in the pre-registered set of multiple agents, each agent is pre-associated with ability tags, types of tasks they are good at, and current state information. The filtering process includes comparing task tags with agent ability tags to select candidate agents that meet the task requirements.

4. The multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, In S3, the scoring mechanism comprehensively considers factors including: the success rate and quality of each agent in completing similar tasks in the past, the current workload and availability, response speed, and context-related capability indicators. By scoring each candidate agent, the system can quantitatively evaluate the expected performance of each agent in performing the task and select the agent with the highest score as the optimal agent.

5. The multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, In S3, the scoring mechanism includes the following sub-modules: a performance evaluation module for recording the historical task success rate and quality evaluation of each agent, a load monitoring module for monitoring the real-time status of agents, and a scoring calculation module for calculating the agent task adaptation score by comprehensively considering the above information; the scoring mechanism sorts the candidate agents according to their scores and prioritizes the agent with the highest score to execute the task.

6. The multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, In S5, the reflection and evaluation content includes: result achievement, efficiency and quality, and agent performance; specifically, it includes: comparing the degree of conformity between the actual results and the expected goals after the task is completed, recording the key performance indicators and anomalies that occurred during the execution process, and analyzing and providing feedback on the execution process, including: whether the task was completed on time and with high quality, whether errors or delays occurred during the process, and whether the agent's actions effectively met the task intent; the feedback is used to update the agent's performance records and scoring parameters, and to adjust and optimize future scheduling strategies.

7. A multi-agent scheduling method based on task intent matching according to claim 6, characterized in that, Adjusting and optimizing future scheduling strategies includes: updating the scoring parameters of agents, refining task label definitions or weights, and improving the parameter configuration of task matching and scoring algorithms; when consecutive task failures or inefficient executions are detected, adjusting the rules for matching task labels with agents or increasing the scoring weights of other candidate agents in similar tasks to achieve adaptive optimization of the scheduling strategy.

8. A multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, In S6, the result merging output includes: for multiple subtasks executed in parallel or step-by-step, collecting the results of all subtasks according to task dependencies; performing format standardization and integrity verification on the result data, and integrating the results of multiple subtasks into the final result using preset merging rules; when a conflict or inconsistency in the subtask results is detected, triggering an exception handling process or feeding back to the relevant agent for re-execution to ensure the consistency and accuracy of the output results.

9. A multi-agent scheduling method based on task intent matching according to claim 1, characterized in that, Multiple intelligent agents are entities that specifically perform tasks; they are robots, software services, or other units with autonomous decision-making and execution capabilities. Each intelligent agent communicates with the scheduling decision module through an interface, receives task instructions, and reports execution progress and results.

10. A multi-agent scheduling system based on task intent matching, characterized in that, include: The task management module is used to receive task requests, parse task intents, and generate structured task tags. The agent resource library stores the capability tags, status information, and historical performance data of multiple agents; The scheduling decision module matches candidate agents in the agent resource library according to the task tags, and selects the optimal agent to execute the task through a built-in scoring mechanism. It monitors the task execution process and performs reflection and evaluation after the task is completed to update the scheduling strategy. The result merging module is used to integrate the task execution results of multiple agents when they work together to complete a task and output the final result. The entire system architecture supports modular expansion and information sharing. The modules interact through well-defined interface protocols and execute the multi-agent scheduling method based on task intent matching as described in any one of claims 1-9, so that the task is managed in a closed loop from receipt to completion.

Images