Multi-agent dynamic scheduling and skill sharing system and method
By leveraging the semantic understanding and dynamic scheduling of the master scheduling agent, combined with global skill registration and selective context passing, the scheduling and resource utilization problems of multi-agent systems in complex task scenarios are solved, achieving efficient and secure enterprise-level multi-agent collaboration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AUTOCORE INTELLIGENT TECH (NANJING) CO LTD
- Filing Date
- 2026-05-29
- Publication Date
- 2026-06-30
AI Technical Summary
Existing multi-agent systems suffer from poor scheduling flexibility, low resource utilization, poor collaboration accuracy, and difficulty in ensuring security in practical engineering deployments, especially in complex task scenarios where they fail to meet enterprise-level requirements.
The system employs a master scheduling agent based on a large language model for semantic understanding, dynamically selects the optimal combination of execution agents, realizes skill sharing and on-demand mounting through a global skill registry, constructs a task dependency graph for serial-parallel strategy selection, and adopts selective context passing and global monitoring mechanisms to ensure the transparency and security of scheduling decisions.
It improves task matching accuracy and scheduling flexibility, reduces resource consumption and capacity redundancy, optimizes task execution efficiency, enhances system security and auditability, and meets enterprise-level compliance requirements.
Smart Images
Figure CN122309089A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to multi-agent system technology, and more particularly to a multi-agent dynamic scheduling and skill sharing system and method. Background Technology
[0002] With the rapid development of Large Language Model (LLM) technology, LLM-based agent systems are widely used in complex task automation scenarios and are deeply applied in high-complexity, multi-stage collaborative scenarios such as enterprise process automation, autonomous driving software development, content creation, and code engineering. However, existing mainstream multi-agent frameworks still have a series of core technical defects that are difficult to adapt to enterprise-level production needs during actual engineering deployment and large-scale implementation, which seriously restrict system scheduling efficiency, resource utilization, collaboration accuracy, and compliance security.
[0003] Existing multi-agent systems suffer from the following core technical problems in practical engineering deployments: Static routing: Existing multi-agent frameworks (such as AutoGen and LangChain) generally use predefined rules for task routing, failing to dynamically select the optimal agent combination based on task semantics, resulting in poor scheduling flexibility and high task mismatch rates. Skill coupling: In existing systems, skill modules are strongly bound to specific agents, preventing different agents from sharing skills, leading to capability redundancy and resource waste. Adding new skills requires modifying the implementations of each agent, resulting in high expansion costs. Lack of serial / parallel decision-making: Existing systems typically use fixed serial or parallel modes, failing to automatically select the optimal execution strategy based on the actual dependencies of subtasks, leading to unnecessary waiting delays or dependency conflicts. Mixed scheduling and execution: Existing orchestrator agents are responsible for both scheduling decisions and directly execute tool calls, resulting in unauditable operation chains, blurred security boundaries, and difficulty meeting enterprise-level compliance requirements. Context pollution: During cross-agent task collaboration, existing systems typically perform full context transmission, causing irrelevant information to interfere with downstream agent decisions, reducing collaboration efficiency and accuracy. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the present invention aims to provide a multi-agent dynamic scheduling and skill sharing system and method.
[0005] To achieve the objectives of this invention, the technical solution adopted is as follows: A multi-agent dynamic scheduling and skill sharing system includes a user interaction layer, a main scheduling layer, an execution layer, and a skill registration layer; The user interaction layer is used to receive the user's natural language task description; The main scheduling layer consists of a single main scheduling agent. The main scheduling agent uses a large language model to perform semantic understanding of natural language tasks, extracts a capability requirement matrix, performs semantic matching between the capability requirement matrix and the capability descriptions of registered execution agents, and dynamically selects the optimal execution agent. The execution layer consists of multiple domain-specific execution agents, which are used to execute specific sub-tasks. Each execution agent supports differentiated underlying model configurations. The skill registration layer maintains a global skill registration center, enabling the decoupling of skills from execution agents, cross-agent sharing, and on-demand mounting and releasing of skills.
[0006] Furthermore, the main scheduling agent is configured to: decompose complex tasks, analyze the dependencies between subtasks and construct a directed acyclic graph (DAG) of task dependencies; automatically select a serial or parallel execution strategy based on the DAG, schedule dependent subtasks in parallel, and schedule dependent subtasks serially in topological order.
[0007] Furthermore, when the agent is registered, it declares its domain of responsibility and task complexity level. Based on this, the system allocates a lightweight and fast model, a high-inference, high-parameter model, or a language-specific model, and supports dynamic adjustment at runtime.
[0008] Furthermore, the global skill registration center adopts a skill-as-a-service mechanism, with skill modules stored independently. Executing agents apply for and attach skills as needed, and the skills are automatically released after the task is completed. New skills support hot-swappable registration.
[0009] Furthermore, the master scheduling agent implements selective context passing: extracts upstream key execution summaries, filters them by relevance score, passes only contexts above a threshold to downstream execution agents, and persists key contexts to long-term memory.
[0010] Furthermore, the main scheduling agent has global status monitoring capabilities, can obtain the progress, resource usage and anomaly information of the executing agents in real time, handle anomalies in a graded manner and provide user decision options, and generate a structured summary report after the task is completed.
[0011] Furthermore, the main scheduling agent adopts a lightweight large language model to complete the task semantic understanding and matching of the execution agent's capabilities. It prioritizes querying the execution agent's capability registry for fast matching. When the matching confidence is lower than the threshold, it calls the large language model for deep inference matching to achieve dynamic matching of the execution agent.
[0012] A method for dynamic scheduling and skill sharing among multiple agents includes the following steps: (1) Users input natural language tasks through the interaction layer, and the system verifies and pushes them to the main scheduling agent; (2) The main scheduling agent completes the semantic understanding of the task through a lightweight large language model, extracts the capability requirement matrix, decomposes sub-tasks and constructs DAG, determines the serial parallel strategy, and dynamically matches the optimal combination of execution agents. (3) The executing agent applies to the global skill registry center to attach skills according to the task requirements. The registry center verifies and dynamically attaches the skills. The skills are released after the task is completed. (4) The executing agent executes subtasks according to the DAG topology, and the main scheduling agent does not participate in the execution; (5) When collaborating across agents, the main scheduling agent performs selective context passing to filter irrelevant information; (6) The main scheduling agent monitors the execution status globally, handles exceptions and summarizes the results, and outputs a structured task report.
[0013] Furthermore, the selective context passing includes: extracting upstream task summaries, calculating the relevance to downstream tasks, filtering the passing content according to a threshold, and persistently storing key information.
[0014] Furthermore, the global monitoring includes: real-time progress acquisition, overall view generation, anomaly impact assessment, hierarchical handling and user decision-making closed loop, and structured report output.
[0015] The beneficial effects of this invention are that, compared with the prior art, the main scheduling agent of this invention dynamically selects the optimal combination of agents based on semantic understanding of a large language model. Compared with static rule routing, the task matching accuracy and adaptability are greatly improved, the scheduling flexibility is significantly improved, and it can handle complex and fuzzy task descriptions that existing systems cannot handle.
[0016] This invention enables cross-agent skill sharing and on-demand mounting through a global skill registry center, eliminating capability redundancy, reducing redundant development, significantly improving skill reuse rate, and reducing overall system resource consumption. Furthermore, by automatically constructing a Directed Axis (DAG) and intelligently selecting a serial-parallel strategy, this invention parallelizes the processing of independent subtasks. Compared to a fixed serial approach, this significantly shortens the overall task execution time and optimizes task execution efficiency.
[0017] This invention features a strict separation of scheduling and execution. The main scheduling agent is only responsible for decision-making and does not execute operations. All side-effect operations are clearly recorded by the executing agent, ensuring a clear and traceable operation chain that meets enterprise-level compliance requirements and enhances system security and auditability. The invention's selective context inheritance mechanism filters irrelevant information, reduces contextual noise in downstream agents, improves the accuracy of cross-agent collaboration, and lowers the decision-making error rate caused by context contamination.
[0018] The main scheduling agent of this invention adopts a lightweight model, which only undertakes the responsibilities of task understanding and distribution. Compared with the traditional solution that uses a flagship model throughout the entire chain, the overall API call cost of the system is greatly reduced while ensuring scheduling accuracy.
[0019] The model resources of this invention are precisely allocated on demand. Each executing agent independently configures the most suitable underlying model according to its responsibility scope and task complexity, avoiding the waste of resources caused by heavy models handling simple tasks, and also avoiding the quality degradation caused by light models handling complex tasks.
[0020] This invention offers strong global observability. The master scheduling agent has the ability to observe the real-time status of all executing agents and detect anomalies. Combined with structured progress reporting and user decision support, the execution process of complex multi-agent tasks is made completely transparent and controllable to the user, significantly improving the reliability and maintainability of the system in the production environment.
[0021] The method of this invention is applicable to various complex task automation scenarios such as autonomous driving software development, content creation, and enterprise process automation, and has broad engineering application value. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of a multi-agent dynamic scheduling and skill sharing system according to the present invention; Figure 2 This is a flowchart of a multi-agent dynamic scheduling and skill sharing method according to the present invention. Detailed Implementation
[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of this application.
[0024] like Figure 1 As shown, the multi-agent dynamic scheduling and skill sharing system of the present invention includes a user interaction layer, a main scheduling layer, an execution layer, and a skill registration layer.
[0025] The user interaction layer receives the user's natural language task description.
[0026] The main scheduling layer consists of a single main scheduling agent (Dispatcher Agent), which is responsible for task understanding and scheduling decisions, and is prohibited from directly executing any side-effect operations.
[0027] After receiving a user task, the master scheduling agent uses a large language model to perform semantic understanding of the task, extracting the task type, domain keywords, and capability requirement matrix. Based on the capability requirement matrix, it performs semantic matching with the capability descriptions of registered execution agents, dynamically selecting one or more optimal execution agents. The master scheduling agent is strictly limited to pure scheduling functions, and is prohibited from directly calling any tools or executing any operations that produce side effects. All execution operations must be completed through execution agents, forming an auditable operation chain. It supports displaying scheduling decisions and waiting for user confirmation before execution, ensuring human supervision (Human-in-the-Loop).
[0028] The execution layer consists of multiple domain-specific execution agents (Sub Agents), each responsible for the specific execution of tasks in a particular domain.
[0029] This invention proposes differentiated configuration of execution agent models, differing from the existing system approach of using the same model for all agents. Each execution agent declares its responsibilities (e.g., driver development, control algorithms, copywriting, data analysis, etc.) and task complexity level (simple / medium / complex) upon registration. The system assigns the most suitable underlying model to each execution agent based on its responsibilities and complexity level: simple tasks (e.g., shell script generation): lightweight and fast models are configured to reduce latency and cost; complex reasoning tasks (e.g., control algorithm design, code migration): high-parameter models with strong reasoning capabilities are configured; creative generation tasks (e.g., content creation, copywriting generation): specialized models with excellent language expression are configured. Model configurations are stored in the configuration files of each execution agent, supporting dynamic adjustments at runtime without requiring a system restart. When allocating tasks, the master scheduling agent can perceive the current model capabilities of each execution agent, comprehensively considering capability matching and cost to make the optimal scheduling decision.
[0030] The skill registration layer maintains the global skill registry and provides on-demand skill services to all executing agents.
[0031] This invention proposes a Skill-as-a-Service (SAS) mechanism, where all skill modules are independently stored in a global skill registry, decoupled from specific execution agents. When executing a task, the execution agent requests the required skills from the global skill registry based on the current task requirements, and releases them upon task completion. The same skill can be reused by multiple execution agents, eliminating capability redundancy. Adding a new skill only requires registration with the registry center; no modification to any existing agent implementation is needed, supporting hot-swappable expansion.
[0032] The main scheduling agent uses a lightweight large language model (such as a small-parameter instruction fine-tuning model) as the scheduling engine, rather than a high-cost flagship model. The core mechanism is as follows: the scheduler only performs two core functions: task semantic understanding and matching the capabilities of the executing agents. It does not participate in any specific computation or code execution, and therefore has relatively low requirements for model inference capabilities. Thus, a lightweight model can be used to reduce the cost per scheduling cycle. The system maintains a dynamic registry of executing agent capabilities, recording metadata such as each agent's domain scope, proficient task types, and historical success rate. When the scheduler receives a task, it prioritizes querying the registry for quick matching. If the matching confidence is below a threshold, it calls the large language model for deep inference matching, achieving dynamic matching of the executing agents. The scheduler and executing agents are independent in model selection: the scheduler uses a lightweight model to ensure response speed, while each executing agent independently selects a suitable model based on its responsibility scope and task complexity. This lightweight scheduling and differentiated execution model allocation strategy significantly reduces the overall call cost while ensuring overall system capability.
[0033] like Figure 2 As shown, the multi-agent dynamic scheduling and skill sharing method of the present invention includes the following steps: (1) User input natural language task; Users input tasks described in natural language through the system's user interaction layer. The system performs preliminary verification of the task description and pushes valid task instructions to the main scheduling agent as the starting point for the entire process.
[0034] (2) After receiving the task, the main scheduling agent completes the semantic understanding of the task through a lightweight large language model and extracts the task domain, target and capability requirement matrix; it decomposes the complex task into multiple independent executable subtasks, automatically analyzes the dependencies between subtasks, and constructs a task dependency directed acyclic graph (DAG); based on the DAG structure, it automatically determines the serial parallel execution strategy, combines the global execution agent capability registry, dynamically selects the optimal combination of execution agents through semantic matching, and completes the scheduling decision. The master scheduling agent processes complex tasks as follows: it decomposes the complex task into multiple subtasks; it automatically analyzes the dependencies between subtasks and constructs a Directed Acyclic Graph (DAG); for subtask nodes without dependencies in the graph, it uses parallel scheduling and simultaneously launches the corresponding execution agents; for subtask nodes with dependencies in the graph, it uses serial scheduling and executes them sequentially according to the topological order. The above serial-parallel strategy is automatically determined by the large language model based on the task semantics, without requiring manual configuration of routing rules.
[0035] (3) The agent attaches skills to the global skill registry; After receiving a subtask from the main scheduling agent, the executing agent initiates a skill mounting request to the global skill registry center based on the capabilities required for task execution. After verifying the availability of the skill, the registry center dynamically mounts the corresponding skill module to the current executing agent. The skill is automatically released after the task is completed, realizing skill sharing and reuse across agents.
[0036] (4) The executing agent executes subtasks according to the DAG topology, and executes them in DAG serial / parallel manner. Only the executing agent is responsible for the specific operations throughout the process, and the main scheduling agent does not participate in the execution. (5) When cooperating across agent tasks, the main scheduling agent selectively passes context; When collaborating across agents, a selective context inheritance mechanism based on task relevance is adopted: after the upstream agent completes its task, the main scheduling agent extracts a key execution summary (including task objectives, main operations, key decisions, and output results); the summary is then scored for relevance with the task requirements of the downstream agent; only contexts with relevance higher than the threshold are passed to the downstream agent, filtering out irrelevant information; key contexts are synchronously persisted to long-term memory storage, supporting cross-session reuse.
[0037] (6) The main scheduling agent monitors the global status of all executing agents in real time. When anomalies such as timeout, error, or blockage occur, the agent assesses the scope of impact and handles them in a tiered manner. If there is no impact, the agent marks and records the event. If there is a dependent impact, the agent pauses the associated task and pushes the user's decision options. After all subtasks are completed, the main scheduling agent collects the execution results, time consumption, status and other information to complete the summary. The master scheduling agent performs real-time global observation and overall state management of all executing agents: 1. The main scheduling agent continuously monitors the running status of all started execution agents, including task execution progress (percentage), current execution stage, estimated remaining time, resource usage, and whether any anomalies have occurred; 2. For scenarios where multiple execution agents are running in parallel, the main scheduling agent will summarize the progress information of each execution agent and generate an overall task progress view, which will be presented to the user in a structured manner. 3. When any executing agent experiences a timeout, error, or blockage, the main scheduling agent immediately detects and assesses the scope of the impact: If the anomaly does not affect other executing agents: mark the anomaly and record it, and report it to the user after all tasks are completed; If the anomaly affects downstream dependent executing agents: immediately suspend the relevant downstream tasks, provide the user with feedback on the current status and decision options (retry / skip / terminate). 4. The main scheduling agent presents the observed running status, progress information, anomaly details, and optional handling solutions to the user in a structured manner, assisting the user in making quick decisions and forming a closed-loop collaborative mode of AI execution and human supervision; 5. After all executing agents have completed their tasks, the main scheduling agent generates a structured task report, including the execution results, time taken, success / failure status, and key outputs of each executing agent, and provides unified feedback to the user.
[0038] (7) Output a structured report.
[0039] Based on the aggregated information, the main scheduling agent generates a structured task report that includes the overall task completion status, subtask execution details, exception records, key outputs, and time consumption statistics. This report is then fed back to the user through the user interaction layer, completing the entire process loop.
[0040] Taking autonomous driving software development as an example: User input: "Help me migrate the camera driver from ROS2 to acos middleware, and update the control algorithm to adapt to the new interface."
[0041] After analysis, the main scheduling agent identified two sub-tasks: (1) camera driver migration (domain: driver development + middleware migration); (2) control algorithm adaptation (domain: control algorithm).
[0042] Dependency analysis: Subtask (2) depends on the output interface definition of subtask (1), so it is executed sequentially.
[0043] Scheduling decision: First, schedule the ros22acos_migration Agent to execute subtask (1). After completion, extract the interface definition summary, and then schedule the control_algo Agent to execute subtask (2).
[0044] Skill mounting: The ros22acos_migration Agent dynamically mounts the "Code Analysis Skill" and "Interface Conversion Skill"; the control_algo Agent mounts the "PID Parameter Adjustment Skill".
[0045] The main scheduling agent does not perform any code modification operations throughout the process; it is only responsible for coordination.
[0046] The applicant of this invention has provided a detailed description of the embodiments of the invention in conjunction with the accompanying drawings. However, those skilled in the art should understand that the above embodiments are merely preferred embodiments of the invention. The detailed description is only intended to help readers better understand the spirit of the invention and is not intended to limit the scope of protection of the invention. On the contrary, any improvements or modifications made based on the inventive spirit of the invention should fall within the scope of protection of the invention.
Claims
1. A multi-agent dynamic scheduling and skill sharing system, characterized in that, It includes the user interaction layer, the main scheduling layer, the execution layer, and the skill registration layer; The user interaction layer is used to receive the user's natural language task description; The main scheduling layer consists of a single main scheduling agent. The main scheduling agent uses a large language model to perform semantic understanding of natural language tasks, extracts a capability requirement matrix, performs semantic matching between the capability requirement matrix and the capability descriptions of registered execution agents, and dynamically selects the optimal execution agent. The execution layer consists of multiple domain-specific execution agents, which are used to execute specific sub-tasks. Each execution agent supports differentiated underlying model configurations. The skill registration layer maintains a global skill registration center, enabling the decoupling of skills from execution agents, cross-agent sharing, and on-demand mounting and releasing of skills.
2. The multi-agent dynamic scheduling and skill sharing system according to claim 1, characterized in that, The main scheduling agent is configured to: decompose complex tasks, analyze the dependencies between subtasks and construct a directed acyclic graph (DAG) of task dependencies; automatically select serial or parallel execution strategies based on the DAG, schedule dependent subtasks in parallel, and schedule dependent subtasks serially in topological order.
3. The multi-agent dynamic scheduling and skill sharing system according to claim 1, characterized in that, When registering an intelligent agent, the system declares the domain of responsibility and the level of task complexity. Based on this, the system allocates a lightweight and fast model, a model with strong inference and a large number of parameters, or a language expression-specific model, and supports dynamic adjustment at runtime.
4. The multi-agent dynamic scheduling and skill sharing system according to claim 1, characterized in that, The global skill registration center adopts a skill-as-a-service mechanism, with skill modules stored independently. Executing agents can apply for and attach skills as needed, and these skills are automatically released after the task is completed. New skills support hot-swappable registration.
5. The multi-agent dynamic scheduling and skill sharing system according to claim 1, characterized in that, The master scheduling agent implements selective context passing: extracts upstream key execution summaries, filters them according to relevance scores, passes only contexts above a threshold to downstream execution agents, and persists key contexts to long-term memory.
6. The multi-agent dynamic scheduling and skill sharing system according to claim 1, characterized in that, The main scheduling agent has global status monitoring capabilities, can obtain the progress, resource usage and abnormal information of the executing agents in real time, handle abnormalities in a hierarchical manner and provide user decision options, and generate a structured summary report after the task is completed.
7. The multi-agent dynamic scheduling and skill sharing system according to claim 1, characterized in that, The main scheduling agent uses a lightweight large language model to complete the semantic understanding of the task and the matching of the execution agent's capabilities. It prioritizes querying the execution agent's capability registry for fast matching. When the matching confidence is lower than the threshold, it calls the large language model for deep inference matching to achieve dynamic matching of the execution agent.
8. A method for dynamic scheduling and skill sharing among multiple agents, characterized in that, Includes the following steps: (1) Users input natural language tasks through the interaction layer, and the system verifies and pushes them to the main scheduling agent; (2) The main scheduling agent completes the semantic understanding of the task through a lightweight large language model, extracts the capability requirement matrix, decomposes sub-tasks and constructs DAG, determines the serial parallel strategy, and dynamically matches the optimal combination of execution agents. (3) The executing agent applies to the global skill registry center to attach skills according to the task requirements. The registry center verifies and dynamically attaches the skills. The skills are released after the task is completed. (4) The executing agent executes subtasks according to the DAG topology, and the main scheduling agent does not participate in the execution; (5) When collaborating across agents, the main scheduling agent performs selective context passing to filter irrelevant information; (6) The main scheduling agent monitors the execution status globally, handles exceptions and summarizes the results, and outputs a structured task report.
9. The method for dynamic scheduling and skill sharing of multiple agents according to claim 8, characterized in that, The selective context passing includes: extracting upstream task summaries, calculating the relevance to downstream tasks, filtering the passing content according to a threshold, and persistently storing key information.
10. The method for dynamic scheduling and skill sharing of multiple agents according to claim 8, characterized in that, The global monitoring includes: real-time progress acquisition, overall view generation, anomaly impact assessment, hierarchical handling and user decision-making closed loop, and structured report output.