An operating system interaction method and system based on multi-agent cooperation

Abstract

This application discloses an operating system interaction method and system based on multi-agent collaboration, belonging to the field of operating system interaction technology. First, a multi-level agent collaboration system is constructed; then, the coordinating agent performs semantic parsing and complexity assessment on user requests, and routes tasks to corresponding agents based on the assessment results; for complex tasks, the planning agent generates a task flowchart, identifies composite sub-task nodes, and assigns them to corresponding composite domain agents for recursive task decomposition; the supervising agent performs task scheduling and status monitoring based on the task flowchart, and, upon receiving intervention instructions, determines the influence domain through dependency path propagation analysis to perform local adjustments, sub-task replanning, or global task replanning; after task completion, execution data is collected, agent collaboration patterns are identified, and a virtual team template is generated. This method can be used for operating system interaction and task processing in complex task scenarios.

Description

Technical Field

[0001] This application relates to the field of operating system interaction technology, and in particular to an operating system interaction method and system based on multi-agent cooperation. Background Technology

[0002] With the development of computer technology, automated interaction technology in operating systems has been widely applied. Currently, technical solutions for task automation mainly include macros and automation tools based on fixed scripts, integrated intelligent assistants capable of executing simple questions and commands, and multi-agent systems that handle tasks by configuring multiple agents. These solutions can liberate users from repetitive manual operations in given scenarios, automating the execution of basic tasks and improving the interaction efficiency of the operating system to a certain extent.

[0003] However, existing operating system interaction technologies still have limitations in complex task scenarios. For example, when a task involves multi-step processing, multi-capability collaboration, and dynamic task adjustment, existing systems typically rely on preset processes or fixed scripts, making it difficult to flexibly decompose and dynamically plan tasks based on task semantics and execution context. Furthermore, users cannot easily intervene and adjust the executing task flow in real time during task execution. In addition, existing multi-agent collaboration methods often rely on static configuration, lacking mechanisms for utilizing historical execution data, making it difficult to optimize agent collaboration based on historical task experience.

[0004] Therefore, it is necessary to provide an operating system interaction method that can support complex task parsing, recursive task decomposition, and dynamic task intervention, so as to improve the processing capability of the operating system interaction system in complex task scenarios. Summary of the Invention

[0005] To address the shortcomings of existing operating system interaction technologies, such as difficulty in achieving recursive specialization and dynamic planning for complex tasks, inability to perform precise real-time human-machine intervention during task execution, lack of reusability of agent collaboration modes, and low system processing capacity and collaboration efficiency for complex open tasks, this invention provides an operating system interaction method based on multi-agent collaboration. The technical solution is as follows: On the one hand, a method for operating system interaction based on multi-agent cooperation is provided, the method comprising: Step 101: Construct a multi-level intelligent agent collaboration system. The system includes a management intelligent agent and a domain execution intelligent agent. The domain execution intelligent agent includes a basic function intelligent agent and a composite domain intelligent agent. The composite domain intelligent agent includes an internal task planning module and a set of sub-intelligent agents, which is used to start the internal task planning module and schedule the set of sub-intelligent agents to execute the sub-task when receiving a composite sub-task. Step 102: Receive the user's task request, perform semantic parsing and complexity evaluation on the task request through a coordinating agent, and route the task to the corresponding domain execution agent or planning agent based on the evaluation results. Step 103: In response to complex tasks, the planning agent parses the task and generates a structured task flowchart, identifies composite subtask nodes in the task flowchart, and assigns the composite subtask nodes to the corresponding composite domain agent, which then performs recursive task decomposition through the internal task planning module. Step 104: The supervisory agent performs task scheduling and status monitoring according to the task flowchart. During task execution, it receives user intervention instructions. Based on the node dependency relationship and execution stack information in the task flowchart, it determines the intervention impact domain through dependency path propagation analysis. Based on the impact domain, it selects to perform local task adjustment, subtask replanning or global task replanning so that the task flow continues to execute after adjustment. Step 105: After completing the task, summarize the results and deliver them to the user. At the same time, collect the execution data of the entire task lifecycle, identify the collaboration mode of multiple agents from the execution data, and encapsulate the collaboration mode into a virtual team template containing preset agent combinations and collaboration process information for the formation and planning of agent teams for similar tasks.

[0006] Optionally, the management agent includes a coordinating agent, a planning agent, and a supervising agent. The multi-level agent collaboration system is equipped with a unified agent registration and discovery mechanism, a standardized message communication mechanism, and a shared context storage module. Various agents interact collaboratively based on the mechanisms and modules, and new agents can be connected to the multi-level agent collaboration system.

[0007] Optionally, when the planning agent identifies that the skill complexity required for a subtask exceeds a preset threshold, the subtask is marked as a composite subtask node.

[0008] Optionally, step 102 includes: S21, the coordinating agent integrates a lightweight semantic analysis engine, and performs real-time semantic parsing of the task request by fusing rule matching based on keywords and templates and fine-tuning the intent classification model through the semantic analysis engine; S22, the coordinating agent generates a comprehensive score by comprehensively considering the number of verb actions in the instruction, the complexity of the object entity, and the implicit contextual dependencies through a complex quantification evaluation algorithm, and determines whether the task is a simple atomic task or a complex composite task based on the comprehensive score. S23. If the task is determined to be a simple atomic task, the task is routed to the corresponding domain execution agent. If the task is determined to be a complex composite task, the task is routed to the planning agent.

[0009] Optionally, the composite domain agent further includes an internal coordination module, which is used to receive composite subtasks and schedule the internal task planning module and the set of sub-agents. The internal task planning module is used to initiate the planning execution process, perform secondary decomposition of the composite subtasks, and generate sub-level task flowcharts. This recursive decomposition process can be carried out in multiple levels until the leaf tasks at the bottom level can be directly processed by the basic functional agents.

[0010] Optionally, step 103 includes: S31, the planning agent deconstructs the complex task description through the deep semantic understanding module and extracts the core objectives, constraints and implicit skill requirement map of the task; S32, the planning agent selects agents from the global agent pool to form an execution team through a skill matching algorithm based on vector similarity; S33, the planning agent uses graph planning technology to generate a directed acyclic task flowchart that includes the execution order of subtasks, dependencies and data flow; S34, the planning agent identifies the composite subtask nodes in the task flowchart and assigns the composite subtask nodes to the corresponding composite domain agent, which then executes recursive task decomposition through the internal task planning module.

[0011] Optionally, step 104 includes: S41, the supervisory agent maintains a dynamic priority scheduling queue according to the task flowchart, assigns tasks in real time according to the readiness status of sub-tasks, and continuously receives status feedback and resource consumption data from each agent through the full-process monitoring center. S42, the supervisory agent intercepts the user's intervention command through the high-priority intervention command channel, and determines the intervention impact domain based on the node dependency relationship and execution stack information in the task flowchart through dependency path propagation analysis. S43, If it is a local parameter adjustment, the supervisory agent notifies the corresponding execution agent to update the task context so that the task flow can continue to be executed after the adjustment; S44, If it is an internal adjustment of a composite subtask, the supervising agent entrusts the corresponding composite domain agent to perform internal replanning so that the task flow continues to be executed after the adjustment. S45, if the global target change is involved, the supervisory agent will send the task back to the planning agent to initiate an overall replanning, so that the task process can continue to be executed after the adjustment.

[0012] Optionally, step 105 includes: S51. After the task is completed, the results are summarized and delivered to the user. At the same time, the execution data of the entire task lifecycle is collected. The task execution data includes the original task request, task flowchart, execution logs of each node, agent collaboration records, user intervention records and performance indicators. S52, the execution data is anonymized and structured, and then persistently stored in the system experience knowledge base; S53, the system analyzes the execution data through machine learning algorithms to explore the combination methods, cooperation order and data exchange patterns of intelligent agents under different task types; S54, the identified collaboration patterns of multiple agents are encapsulated into a virtual team template containing preset agent combinations and collaboration process information, and the preset agent combinations and collaboration process information of the virtual team template are improved for the formation and planning of agent teams for similar tasks.

[0013] Optionally, the system utilizes the execution data in the system's experience knowledge base, using task completion time, resource utilization efficiency, and user satisfaction as reward signals, to optimize the task decomposition strategy model of the planning agent and the intervention influence domain judgment model of the supervised agent through reinforcement learning algorithms. At the same time, it optimizes the parsing and evaluation accuracy of the semantic parsing model and the complexity evaluation model through supervised learning.

[0014] On the other hand, an operating system interaction system based on multi-agent cooperation is provided. The system is used to implement the operating system interaction method based on multi-agent cooperation described above, including a management and coordination agent, a domain execution agent, an agent registration and discovery module, a standardized communication module, a shared context storage module, a system experience knowledge base, and a learning and optimization module. Each module interacts with data through an asynchronous message communication mechanism to support decoupled communication between the agents; The management and coordination intelligent agent includes a coordination intelligent agent, a planning intelligent agent, and a supervisory intelligent agent, which are respectively used to perform semantic parsing and routing of task requests, planning and decomposition of complex tasks, and task scheduling and intervention processing. The domain execution agent includes a basic functional agent and a composite domain agent. The composite domain agent has a built-in internal coordination module, an internal task planning module, and a set of sub-agents for the recursive management and execution of sub-tasks. The learning optimization module is used to extract task execution data from the system experience knowledge base, mine the agent collaboration mode and generate a virtual team template containing preset agent combinations and collaboration process information, and at the same time optimize the decision model of each agent. The supervisory agent incorporates an influence domain analysis engine and a dynamic scheduling engine; wherein... The influence domain assessment engine is configured to determine the intervention influence domain based on the node dependency relationship of the task flowchart and real-time execution stack information, through dependency path propagation analysis, analyze the influence range and degree of user intervention instructions, and output the assessment results of local adjustment, subtask replanning or global replanning. The dynamic scheduling engine is used to maintain the task ready queue, perform intelligent scheduling operations based on subtask dependencies and priorities, and the system sets up intervention transaction logs and consistency guarantee protocols to achieve the traceability of intervention operations and the consistency of the global state.

[0015] The purpose of this invention is to provide an operating system interaction method and system based on multi-agent cooperation, so as to solve the problem that existing operating system automation technology is difficult to support the recursive decomposition of complex tasks, dynamic task intervention, and reuse of agent cooperation modes. To achieve the above objectives, this invention provides an operating system interaction method based on multi-agent collaboration, comprising: constructing a multi-level agent collaboration system, the system including a management agent and a domain execution agent, wherein the domain execution agent includes a basic function agent and a composite domain agent; receiving user task requests, and performing semantic parsing and complexity assessment on the task requests through a coordinating agent, and routing the task to the corresponding domain execution agent or planning agent based on the assessment results; for complex tasks, generating a structured task flowchart through a planning agent, identifying composite sub-task nodes in the task flowchart, assigning the composite sub-task nodes to the corresponding composite domain agents, and having the composite domain agents perform recursive task decomposition through their internal task planning modules; during task execution, performing task scheduling and status monitoring based on the task flowchart through a supervisory agent, and determining the intervention impact domain through dependency path propagation analysis when receiving user intervention instructions, and performing local task adjustment, sub-task replanning, or global task replanning based on the impact domain; collecting task execution data after task completion, and generating a virtual team template by analyzing the collaboration patterns of multiple agents for subsequent team building and task planning for similar tasks.

[0016] Through the above technical solutions, the present invention realizes the recursive decomposition of complex tasks, dynamic intervention and control during task execution, and learning and reuse of intelligent agent collaboration modes, thereby improving the automated processing capability of the operating system in complex task scenarios. Attached Figure Description

[0017] Figure 1 A flowchart illustrating an operating system interaction method based on multi-agent collaboration is shown. Figure 2 This diagram illustrates a multi-level intelligent agent collaborative system architecture. Figure 3This diagram illustrates a task planning and recursive decomposition process. Figure 4 This diagram illustrates a dynamic intervention approach for a task. Figure 5 This diagram illustrates a system's continuous optimization process. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0019] In this article, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. Example 1

[0020] Figure 1 The flowchart illustrates an operating system interaction method based on multi-agent cooperation, which includes: Step 101: Construct a multi-level intelligent agent collaborative system.

[0021] The overall architecture of a multi-level intelligent agent collaborative system is shown in the diagram below. Figure 2 The diagram illustrates the collaborative relationship between the managerial intelligent agent and the domain execution intelligent agent in this invention, as well as the system infrastructure consisting of the intelligent agent registration and discovery module, the standardized communication module, and the shared context storage module.

[0022] The system comprises managerial agents and domain execution agents. Domain execution agents include basic functional agents and composite domain agents. Composite domain agents consist of an internal task planning module and a set of sub-agents, used to activate the internal task planning module and schedule the set of sub-agents to execute the sub-tasks when a composite sub-task is received.

[0023] In this embodiment, the multi-level intelligent agent collaboration system is a loosely coupled, dynamically expandable system architecture. The management intelligent agent serves as the system's management and coordination core, undertaking management tasks such as task allocation, planning, scheduling, and monitoring. The domain execution intelligent agent serves as the system's execution core, responsible for the implementation of specific tasks. Basic functional intelligent agents handle atomic operation tasks that do not require decomposition, such as file reading and writing, network requests, calculator startup, and single data calculation, directly completing the corresponding operations. Composite domain intelligent agents are designed for composite sub-tasks within a professional domain, serving as independent units for handling such tasks. Their integrated internal task planning module possesses task planning and decomposition capabilities. The sub-agent set contains several basic functional intelligent agents or secondary composite domain intelligent agents adapted to the professional domain. When a composite domain intelligent agent receives a composite sub-task assigned by a higher-level planning intelligent agent, it automatically activates its internal task planning module and, based on the task decomposition results, schedules the corresponding intelligent agents in the sub-agent set to execute the sub-task, achieving specialized and refined processing of composite sub-tasks.

[0024] Step 102: Receive the user's task request, perform semantic parsing and complexity evaluation on the task request through a coordinating agent, and route the task to the corresponding domain execution agent or planning agent based on the evaluation results.

[0025] In this embodiment, users can send task requests to the system through various means such as natural language and standardized interaction commands. After receiving the request, the system uniformly transmits it to the coordinating agent. The coordinating agent acts as the system's "intelligent gateway," undertaking the initial parsing and routing of task requests. First, it performs semantic parsing of the task request through its built-in parsing engine to accurately uncover the user's core intent. Then, based on the semantic parsing results, it performs a quantitative assessment of the task complexity to distinguish between simple atomic tasks and complex composite tasks. For simple atomic tasks that do not require task decomposition, the coordinating agent directly routes them to the corresponding domain execution agent to achieve rapid task response and execution. For complex composite tasks that require multi-step processing, cross-domain execution, or have implicit objectives, the coordinating agent routes them to the planning agent, which performs in-depth task planning and decomposition to ensure professional handling of complex tasks.

[0026] Step 103: In response to complex tasks, the planning agent parses the task and generates a structured task flowchart, identifies composite subtask nodes in the task flowchart, and assigns the composite subtask nodes to the corresponding composite domain agents. The composite domain agents then perform recursive task decomposition through their internal task planning modules.

[0027] Figure 3This diagram illustrates a task planning and recursive decomposition process, which shows how a planning agent, upon receiving a complex task, generates a task flowchart, identifies composite subtask nodes, assigns the composite subtasks to the corresponding composite domain agents, and performs recursive task decomposition.

[0028] In this embodiment, the planning agent acts as the "strategy center" for complex tasks, only coordinating the routing of complex composite tasks. It first deeply deconstructs the complex task, clarifying the core objective, explicit and implicit constraints, and the skill requirements for completing the task. Then, based on the skill requirements, it selects suitable agents from the global agent pool to generate a structured task flowchart containing the execution order and dependencies of subtasks. During the generation of the task flowchart, the planning agent identifies all decomposed subtasks, filters out composite subtask nodes that cannot be directly handled by the basic functional agents, and accurately assigns them to the corresponding composite domain agents based on their domain attributes. Upon receiving the composite subtask, the composite domain agent initiates the planning execution process through its internal task planning module, performing a secondary decomposition of the composite subtask. If the decomposed subtask is still a composite subtask, it is further decomposed by the corresponding secondary composite domain agent. This process can be recursively performed at multiple levels until all subtasks are leaf tasks that can be directly handled by the basic functional agents.

[0029] Step 104: The supervisory agent performs task scheduling and status monitoring based on the task flowchart. During task execution, it receives user intervention commands. Based on the node dependencies and execution stack information in the task flowchart, it determines the intervention impact domain through dependency path propagation analysis. Based on the impact domain, it selects to perform local task adjustment, subtask replanning, or global task replanning so that the task flow continues to execute after adjustment.

[0030] Figure 4 The diagram illustrates a dynamic task intervention mechanism, which shows how a supervisory agent receives user intervention instructions during task execution, determines the intervention impact domain through dependency path propagation analysis, and performs local task adjustments, subtask replanning, or global task replanning.

[0031] In this embodiment, the supervisory agent acts as the "execution commander" of the task, working entirely according to the task flowchart generated by the planning agent. On the one hand, it maintains the rhythm of task execution and performs intelligent scheduling based on the readiness status of sub-tasks; on the other hand, it continuously monitors the task execution status and resource consumption of each agent to keep abreast of the task execution progress. The system reserves a high-priority intervention command channel, and the supervisory agent is the sole listener of this channel, capable of intercepting user intervention commands in real time at any node of task execution. Upon receiving an intervention command, the supervisory agent first performs semantic parsing of the command to extract the core intervention intent. Then, combining the node dependencies in the task flowchart and the execution stack information of the current task, it traverses the entire task graph from the intervention target node through dependency path propagation analysis, marking all affected nodes and edges to accurately determine the scope and degree of the intervention impact domain. Finally, it executes the corresponding adjustment strategy based on the judgment results, and all adjustment operations do not interrupt the overall task flow, achieving uninterrupted task execution.

[0032] Step 105: After completing the task, summarize the results and deliver them to the user. At the same time, collect the execution data of the entire task lifecycle, identify the collaboration mode of multiple agents from the execution data, and encapsulate the collaboration mode into a virtual team template containing preset agent combinations and collaboration process information for the formation and planning of agent teams for similar tasks.

[0033] Figure 5 The diagram illustrates a continuous optimization process of a system, which shows how the system continuously learns and optimizes itself by collecting data from the entire lifecycle of the task after it is completed, analyzing the collaboration patterns of the agents, and generating a virtual team template.

[0034] In this embodiment, after all leaf tasks in the task flowchart have been completed, the supervising agent summarizes, integrates, and verifies the execution results of each agent to form a complete task result and delivers it to the user. At the same time, the system will automatically start the full-link data acquisition process to collect various types of execution data throughout the entire lifecycle of this task, from request reception to result delivery. The system analyzes the collected execution data through machine learning algorithms to discover efficient combination methods, collaboration sequences, data exchange modes, and other collaboration patterns of different agents during task execution. The system then encapsulates these collaboration patterns into standardized virtual team templates. The templates contain preset agent combinations and optimal collaboration process information suitable for the corresponding task types. When the system receives similar tasks in the future, it can directly call the template to form agent teams and plan tasks, thus achieving the reuse of collaboration patterns.

[0035] Therefore, the purpose of this invention is to provide an operating system interaction method and system based on multi-agent cooperation, so as to solve the problem that existing operating system automation technology is difficult to support the recursive decomposition of complex tasks, dynamic task intervention, and reuse of agent cooperation modes. The system comprises a multi-level intelligent agent collaboration system, including management agents and domain execution agents. The domain execution agents include basic functional agents and composite domain agents. The system receives user task requests and uses a coordinating agent to perform semantic parsing and complexity assessment. Based on the assessment results, the task is routed to the corresponding domain execution agent or planning agent. For complex tasks, the planning agent generates a structured task flowchart and identifies composite sub-task nodes within the flowchart. These composite sub-task nodes are then assigned to the corresponding composite domain agents, which perform recursive task decomposition through their internal task planning modules. During task execution, a supervisory agent performs task scheduling and status monitoring based on the task flowchart. Upon receiving user intervention commands, the system determines the intervention impact domain through dependency path propagation analysis and performs local task adjustments, sub-task replanning, or global task replanning based on the impact domain. After task completion, task execution data is collected, and a virtual team template is generated by analyzing the collaboration patterns of multiple agents for subsequent team building and task planning for similar tasks.

[0036] Through the above technical solutions, the present invention realizes the recursive decomposition of complex tasks, dynamic intervention and control during task execution, and learning and reuse of intelligent agent collaboration modes, thereby improving the automated processing capability of the operating system in complex task scenarios. Example 2

[0037] According to the operating system interaction method based on multi-agent cooperation in Embodiment 1, this embodiment defines the architectural details of the multi-level agent cooperation system. The management agents include coordinating agents, planning agents, and supervising agents. The multi-level agent cooperation system is equipped with a unified agent registration and discovery mechanism, a standardized message communication mechanism, and a shared context storage module. Various agents interact collaboratively based on the mechanism and module. New agents can be connected to the multi-level agent cooperation system.

[0038] In this embodiment, the three types of intelligent agents in the management system each perform their respective duties and cooperate to form a complete management and coordination link. The coordinating intelligent agent is responsible for the initial parsing and distribution of task requests, the planning intelligent agent is responsible for the planning and decomposition of complex tasks, and the supervising intelligent agent is responsible for the scheduling, monitoring, and intervention of task execution. The division of labor among the three makes the management work more professional. The intelligent agent registration and discovery mechanism is one of the basic core mechanisms of the system. It is used to uniformly manage and publish the capabilities, status, interfaces, and service endpoints of all intelligent agents in the system. All intelligent agents will automatically register with this mechanism when they start up and will automatically deregister when they go offline. The system can quickly search and match intelligent agents that are suitable for task requirements through this mechanism. The standardized message communication mechanism provides standardized message formats and reliable communication channels for interactions between various intelligent agents, enabling decoupled interaction and ensuring the accuracy and timeliness of message transmission. The shared context storage module is used to store and retrieve context information such as the global ID, stage parameters, intermediate results, and execution status of tasks, allowing cross-level and cross-type intelligent agents to access it on demand, ensuring the consistency of context information during task execution. Various intelligent agents carry out collaborative interaction based on the above unified mechanism and module, and newly developed intelligent agents can be directly connected to this multi-level intelligent agent collaboration system after completing development according to the system's unified interface specifications, realizing the dynamic expansion of the system's intelligent agent capabilities.

[0039] In this embodiment, by subdividing management agents into three categories—coordinating, planning, and supervising—a professional division of labor for task management and coordination is achieved, avoiding the limitations of managing a single agent and improving the system's management efficiency and professionalism. By configuring a unified agent registration and discovery mechanism, a standardized message communication mechanism, and a shared context storage module, a foundational support is provided for collaborative interaction between agents, enabling loosely coupled collaboration and ensuring efficient information transmission and consistent task context within the system. Furthermore, by supporting the convenient access of new agents, the system possesses elastic scalability, allowing for the continuous enrichment of agent types and capabilities based on business needs. This further enhances the system's generalization ability to handle complex and open tasks, enabling it to adapt to more application scenarios. Example 3

[0040] Optionally, when the planning agent identifies that the skill complexity required for a subtask exceeds a preset threshold, the subtask is marked as a composite subtask node.

[0041] In this embodiment, during the process of decomposing complex tasks and generating structured task flowcharts, the planning agent performs a quantitative assessment of the skill complexity of each subtask. The assessment indicators for skill complexity include the amount of domain knowledge required by the subtask, the number of independent operation steps, the difficulty of resource allocation, and the degree of correlation with other subtasks. The system sets corresponding weight coefficients for each assessment indicator based on the actual application scenario and task processing requirements, and presets a quantification threshold for skill complexity. The planning agent obtains a skill complexity score for each subtask through weighted calculation. When the score exceeds the preset threshold, the subtask is determined to be highly complex and cannot be directly processed by the basic functional agent. It is marked as a composite subtask node, and further matched with a corresponding composite domain agent based on the domain attributes of the subtask. If the skill complexity score does not exceed the preset threshold, the subtask is determined to be a leaf task and is directly matched with the corresponding basic functional agent for execution.

[0042] In this embodiment, by setting a skill complexity threshold and using it as an objective basis to identify composite subtask nodes, the planning agent can quickly and accurately distinguish between leaf tasks and composite subtasks. This provides a clear and unified judgment standard for subsequent agent allocation and task recursive decomposition, ensuring the scientific and rational nature of task decomposition. At the same time, this identification rule is an automated quantitative identification that does not require manual intervention, improving the system's task planning efficiency and enabling the system to adaptively handle subtasks with different skill complexities, further enhancing the system's intelligence level. Example 4

[0043] Optionally, step 102 includes: S21, the coordinating agent integrates a lightweight semantic analysis engine, which combines rule matching based on keywords and templates with a fine-tuned intent classification model to perform real-time semantic parsing of task requests. In this embodiment, the lightweight semantic analysis engine of the coordinating agent in S21 adopts a dual-channel design. The two channels run in parallel and are fused together using a confidence-weighted approach to output the parsing results. The rule matching channel based on keywords and templates has a built-in regular expression and keyword template library to capture high-frequency task requests with fixed patterns, such as "open the calculator" or "search for Excel files on the desktop," which is fast and efficient. The fine-tuned intent classification model channel is built based on a distilled Transformer model or a lightweight BERT model to handle diverse, colloquial, and ambiguous user expressions, such as "organize last week's project materials and save them to the root directory of drive D," which can accurately uncover the user's deep intent. The fusion of the two channels ensures that semantic parsing is both efficient and accurate.

[0044] S22, the coordinating agent uses a complex quantification evaluation algorithm to comprehensively consider the number of verb actions in the instruction, the complexity of the object entity, and the implicit contextual dependencies to generate a comprehensive score, and determines whether the task is a simple atomic task or a complex composite task based on the comprehensive score. The complexity quantification evaluation algorithm in S22 is based on the semantic parsing results of S21. It quantifies and weights the three core dimensions of the task request to generate a comprehensive score of 0-100. The verb action count refers to the number of independent operations contained in the task request. For example, "copy and rename the file" contains two verb actions and has a higher score. The object entity complexity refers to the complexity of the operation object in the task request. The complexity of a single file is lower than that of the entire file directory. The more complex the operation object, the higher the score. The implicit context dependency refers to the degree of referential expression and association with historical operations contained in the task request. For example, "process the file from before" has obvious context dependency. The higher the degree of dependency, the higher the score. The system presets a comprehensive score threshold to distinguish between simple atomic tasks and complex compound tasks.

[0045] S23. If the task is determined to be a simple atomic task, the task is routed to the corresponding domain execution agent. If the task is determined to be a complex composite task, the task is routed to the planning agent.

[0046] S23 is the routing execution step of the coordinating agent. It is a precise routing based on the complexity judgment result of S22. For requests judged as simple atomic tasks, the coordinating agent will quickly match the corresponding domain execution agent and directly dispatch the task through the agent registration and discovery mechanism, realizing a fast "request-response" closed loop. For requests judged as complex composite tasks, the coordinating agent will package the complete task request, semantic parsing result and complexity evaluation report, and submit them asynchronously to the waiting queue of the planning agent. At the same time, it will return task receipt confirmation information to the user to avoid long waiting time for the user.

[0047] In some embodiments, when a user inputs "delete empty folders on the desktop," the coordinating agent quickly parses the user's intent through a rule matching channel. This request contains only one verb action, the operation object is an empty folder on the desktop, there is no context dependency, and the overall score is 25 points, which is lower than the preset threshold. It is judged as a simple atomic task and directly routed to the file operation basic function agent for execution. In some embodiments, when a user inputs "organize sales data for the past month and generate a visualization report," the coordinating agent parses the user's core intent through an intent classification model channel. This request contains two verb actions: data organization and report generation. The operation object is sales data for the past month, it needs to call historical data resources, there is a context dependency, and the overall score is 80 points, which is higher than the preset threshold. It is judged as a complex composite task and routed to the planning agent for deep planning.

[0048] In this embodiment, by configuring a lightweight semantic analysis engine that integrates rule matching and intent classification models for the coordinating agent, the system can adapt to user task requests of different forms and expression styles. This ensures the parsing efficiency of high-frequency fixed requests and improves the parsing accuracy of diverse colloquial requests, laying a solid foundation for subsequent complexity assessment and routing. Through a complexity quantification evaluation algorithm, task requests are quantitatively scored in multiple dimensions, achieving an objective and accurate distinction between simple atomic tasks and complex composite tasks, avoiding biases in subjective judgments, and providing a quantitative basis for subsequent precise routing. Precise routing based on the judgment results ensures the system's ability to respond instantly to high-frequency simple requests and its professional processing capability for complex requests, improving the user's interactive experience and achieving efficient utilization of system resources. Example 5

[0049] Optionally, the composite domain agent also includes an internal coordination module. The internal coordination module is used to receive composite subtasks and schedule the internal task planning module and the set of sub-agents. The internal task planning module is used to initiate the planning execution process, perform secondary decomposition of composite subtasks and generate sub-level task flowcharts. This recursive decomposition process can be carried out in multiple levels until the leaf tasks at the bottom level can be directly processed by the basic functional agents.

[0050] In this embodiment, the composite domain agent is a miniaturized agent collaboration subsystem. The internal coordination module, internal task planning module, and sub-agent set form a complete "receive-plan-execute" closed loop, capable of independently processing composite sub-tasks within a specialized domain. The internal coordination module serves as the task entry point for the composite domain agent, receiving composite sub-tasks assigned by the superior planning agent or supervisory agent. Simultaneously, based on the decomposition results from the internal task planning module, it schedules the corresponding agents within the sub-agent set, coordinating the collaborative relationships and data flow between sub-agents. The internal task planning module functions identically to the core functions of the superior planning agent, initiating the planning and execution process. It performs secondary specialized decomposition of the received composite sub-tasks, clarifying the execution order and dependencies of the decomposed sub-tasks, and generating sub-level task flowcharts, providing a clear basis for scheduling sub-agents. The sub-agent set consists of multiple agents adapted to the corresponding specialized domain, and can include basic functional agents and secondary composite domain agents according to domain requirements. The recursive decomposition process of the composite domain agent is an adaptive multi-level process. After the internal task planning module completes the secondary decomposition, it will identify the complexity of the decomposed subtasks. If the decomposed subtasks are still composite subtasks, they will be assigned to the corresponding secondary composite domain agents, which will continue to execute the decomposition process. This process can be repeated until all subtasks are leaf tasks that can be directly processed by the basic functional agents, and then the corresponding basic functional agents will perform the specific operations.

[0051] In some embodiments, the user request is "to create a research PPT on the application of AI medical imaging." The planning agent decomposes this request into sub-tasks such as "data collection," "content organization," and "PPT creation." Among these, "PPT creation" is a composite sub-task node, assigned to a composite domain agent for PPT creation. After receiving the "PPT creation" task, the internal coordination module of this agent immediately schedules the internal task planning module to start the planning execution process, further decomposing it into sub-tasks such as "template matching," "material layout," "text editing," "chart insertion," and "export and save," generating a sub-level task flowchart. Based on the decomposition results, the agent schedules basic function agents such as template matching, layout design, and text processing in the sub-agent set to execute the corresponding sub-tasks, completing the execution of the entire composite sub-task.

[0052] In this embodiment, by configuring an internal coordination module, an internal task planning module, and a set of sub-agents for the composite domain agent, the composite domain agent acquires independent task reception, planning, decomposition, and execution capabilities, becoming an autonomous professional domain task processing unit capable of independently completing the refined processing of professional domain composite sub-tasks. By implementing a multi-level recursive decomposition process, the system can perform infinitely deep professional decomposition of complex tasks, transforming cross-domain, highly complex tasks into leaf tasks that can be processed by multiple basic functional agents. This solves the problem of insufficient task execution depth and inability to handle complex professional tasks in existing technologies, improving the system's ability to handle highly complex and professional tasks. At the same time, the recursive decomposition process is transparent to the external upper-level agents, which only need to focus on the overall execution result of the composite sub-tasks without interfering with the internal decomposition process, ensuring the simplicity and efficiency of the overall system architecture. Example 6

[0053] Optionally, step 103 includes: S31, the planning agent deconstructs complex task descriptions through a deep semantic understanding module, extracting the core objectives, constraints, and implicit skill requirement map of the task; In this embodiment, the deep semantic understanding module for planning intelligent agents in S31 is built based on a domain-fine-tuned large language model. It can perform deep and comprehensive semantic mining on complex task descriptions. It can not only identify the user's explicit core objectives, but also accurately mine the explicit and implicit constraints of the task, such as task completion time limits, output format requirements, data source restrictions, and result display requirements. At the same time, it represents the set of core capabilities required to complete the task in the form of multi-dimensional vectors, forming a structured skill requirement graph. The skill requirement graph provides a precise basis for subsequent intelligent agent selection.

[0054] S32, the planning agent selects agents from the global agent pool to form an execution team through a skill matching algorithm based on vector similarity; In S32, the planning agents are based on the skill requirement map generated in S31. Agents are selected through a skill matching algorithm based on vector similarity. This algorithm calculates the cosine similarity between each skill vector in the skill requirement map and the skill description vectors of each agent in the global agent pool, and selects a list of candidate agents whose similarity exceeds a preset threshold (such as 0.75). Then, a multi-objective optimization strategy is used to comprehensively rank the candidate agents, taking into account factors such as skill matching degree, historical execution performance, current load status, and historical cooperation tacitness between agents. Finally, the optimal agent is selected for each skill requirement to form an execution team adapted to the complex task.

[0055] S33, the planning agent uses graph planning technology to generate a directed acyclic task flowchart that includes the execution order of subtasks, dependencies and data flow; In S33, the planning agent uses graph planning technology. Taking the core objective, constraints, and capability boundaries of the execution team as input, it decomposes complex tasks into multiple logically clear subtasks, clarifies the execution order of each subtask, the dependencies between predecessors and successors, and the data exchange flow between subtasks, and generates a directed acyclic task flowchart. The flowchart clearly identifies subtask branches that can be executed in parallel and dependency chains that need to be executed serially. The generated task flowchart undergoes consistency verification to ensure that there are no circular dependencies, resource conflicts, or logical breaks. After the verification is passed, it is encapsulated into a standard task graph object.

[0056] S34, plan the composite sub-task nodes in the task flowchart of the intelligent agent recognition task, and assign the composite sub-task nodes to the corresponding composite domain intelligent agents, which then execute recursive task decomposition through the internal task planning module.

[0057] S34 is the step of identifying and allocating composite subtask nodes for the planning agent. The planning agent will traverse all subtask nodes in the task flowchart generated in S33, identify composite subtask nodes through the skill complexity threshold judgment rule, and accurately allocate them to the corresponding composite domain agent according to the professional domain attributes of the composite subtask. After receiving the composite subtask, the composite domain agent will start a recursive task decomposition process through its own internal task planning module to complete the multi-level decomposition of the composite subtask.

[0058] In this embodiment, a deep semantic understanding module performs a deep and comprehensive deconstruction of complex tasks, enabling the planning agent to accurately grasp the user's core intent and various constraints of the task. This avoids deviations between task planning and user needs, providing an accurate and comprehensive basis for subsequent task decomposition and agent formation. An execution team is formed using a skill matching algorithm based on vector similarity and a multi-objective optimization strategy, achieving precise matching between the agent and task requirements. This ensures the optimality and collaborative efficiency of the execution team, making task execution more professional. By using graph planning technology to generate a directed acyclic task flowchart, complex tasks are transformed into structured, executable subtask flows. The execution order, dependencies, and data flow of subtasks are clarified, providing a clear and standardized framework for subsequent task scheduling and improving the scientific rigor and rationality of complex task planning. Through the precise identification and allocation of composite subtask nodes, specialized division of labor for complex tasks is achieved, laying the foundation for subsequent recursive task decomposition and further enhancing the system's ability to handle complex tasks. Example 7

[0059] Optionally, step 104 includes: S41, the supervisory agent maintains a dynamic priority scheduling queue based on the task flowchart, assigns tasks in real time according to the readiness status of sub-tasks, and continuously receives status feedback and resource consumption data from each agent through the full-process monitoring center. In this embodiment, the dynamic priority scheduling queue of the supervisory agent in S41 is constructed based on the subtask dependencies in the task flowchart and the preset priority strategy. The preset priority strategy may include shortest expected execution time priority, critical path priority, etc. The supervisory agent will calculate the readiness status of each subtask in real time. When all the predecessor tasks of a subtask are completed and the required resources are available, it will be moved into the readiness queue and assigned to the corresponding execution agent according to priority. The full-process monitoring center continuously receives the heartbeat signals, status change notifications, and resource consumption data such as CPU / memory / IO periodically reported by each execution agent through the asynchronous communication bus. It aggregates these data into a global task view and synchronizes them to the shared context storage module, providing real-time and accurate data support for scheduling decisions and intervention analysis.

[0060] S42, the supervisory agent intercepts user intervention commands through a high-priority intervention command channel, and determines the intervention impact domain based on the node dependency relationships and execution stack information in the task flowchart through dependency path propagation analysis; In S42, the high-priority intervention command channel is a dedicated communication channel for user intervention commands, possessing the highest message priority. This ensures that user intervention commands can be quickly and accurately intercepted by the supervisory agent. After intercepting the intervention command, the supervisory agent first parses the command through a lightweight semantic analysis module to extract core information such as operation type, target object, and changed content. Then, combined with the node dependency relationships in the task flowchart and the execution stack information of the current task, it initiates dependency path propagation analysis. Starting from the intervention target node, it traverses forward and backward along the data dependency edges and control dependency edges of the task graph, marking all potentially affected nodes and edges, quantifying the scope (local / global) and degree of influence, and accurately determining the intervention influence domain.

[0061] S43, If it is a local parameter adjustment, the supervisory agent notifies the corresponding execution agent to update the task context so that the task flow can continue to execute after the adjustment; S44. If it is an internal adjustment of a composite subtask, the supervising agent entrusts the corresponding composite domain agent to perform internal replanning so that the task process continues to execute after the adjustment. S45, if the global goal is changed, the supervisory agent will send the task back to the planning agent to initiate a global replanning, so that the task process can continue to be executed after the adjustment.

[0062] S43, S44, and S45 are three-level hierarchical adjustment strategies for the supervisory agent based on the influence domain analysis results of S42, all with "uninterrupted task execution" as the core principle.

[0063] S43's local parameter adjustment is suitable for situations where only local information such as the output path and parameter settings of the currently executing task are modified. The supervisory agent will directly send the adjustment instruction to the corresponding executing agent. After receiving the instruction, the agent updates the task context in place and continues to execute the task without interrupting the process. S44's internal adjustment of a composite subtask is suitable for situations where the internal requirements and execution steps of a composite subtask are modified. The supervisory agent encapsulates the adjustment requirement as a "replanning request" and delegates it to the corresponding composite domain agent, which initiates a local replanning process internally, adjusts the flow of lower-level subtasks, and continues execution. The supervisory agent only needs to receive the final adjustment completion notification and does not affect the overall task flow. S45's global goal change is suitable for situations where the core goal, overall output requirements, and main execution direction of the task are modified. The supervisory agent will immediately broadcast a "safe pause" signal to all executing agents. After the current atomic operation is completed, the entire process is paused. The original task object, the state snapshot of the executed part, and the user intervention instruction are packaged together and sent back to the planning agent to initiate overall replanning. After the replanning is completed, the new task graph is submitted to the supervisory agent again, which seamlessly resumes execution at the breakpoint.

[0064] In this embodiment, by supervising the intelligent agent to maintain a dynamic priority scheduling queue and assign tasks in real time according to the ready status of subtasks, intelligent and efficient task scheduling is achieved, ensuring the rhythm and efficiency of task execution. Simultaneously, real-time monitoring by the full-process monitoring center allows the system to promptly grasp the task execution status and resource consumption, providing data support for anomaly handling and scheduling decisions. By setting up a high-priority intervention command channel, the system ensures the rapid and accurate interception of user intervention commands, allowing users to intervene in the task execution process at any node. Dependency path propagation analysis enables accurate determination of the intervention impact domain, providing a scientific basis for hierarchical adjustments and avoiding the impact of blind adjustments on task execution. Through a three-level hierarchical adjustment strategy, a "surgical" precise intervention in the task execution process is achieved, ensuring user intervention needs while minimizing the impact of intervention on the overall task flow, achieving uninterrupted task execution, and significantly improving the flexibility and efficiency of human-machine collaboration. Example 8

[0065] Optionally, step 105 includes: S51 summarizes the results and delivers them to the user after the task is completed. At the same time, it collects the execution data of the entire task lifecycle. The task execution data includes the original task request, task flowchart, execution logs of each node, agent collaboration records, user intervention records and performance indicators. In this embodiment, after all leaf tasks in the task flowchart have been executed in S51, the supervising agent will summarize, integrate, and verify the execution results of each agent to form a complete task result and deliver it to the user through the system interface. At the same time, the system will start the full-link data acquisition process to collect multi-dimensional execution data of the entire lifecycle of this task. This data covers the entire process of task processing. The original task request includes information such as the natural language / instructions input by the user and the request time. The task flowchart includes the standard task graph object generated by the planning agent and the sub-level task graphs of recursion. The execution log of each node includes information such as the execution time, execution result, exception information, and processing agent of each sub-task. The agent collaboration record includes information such as the combination method of agents, collaboration order, data exchange content and timing. The user intervention record includes information such as intervention instructions, intervention time, impact domain judgment results, and adjustment strategies. The performance indicators include information such as task completion time, resource utilization efficiency, agent response speed, and task success rate.

[0066] S52, after anonymizing and structuring the execution data, persists it to the system's experience knowledge base; In S52, the system will preprocess the execution data collected by S51. First, it will perform anonymization to remove privacy information such as user identification and device information. Then, it will perform structured transformation according to a predefined data pattern to transform unstructured data into analyzable structured data. After preprocessing, the data will be persistently stored in the system's experience knowledge base through a batch write interface. This knowledge base adopts a distributed storage architecture and supports fast retrieval of multiple conditions based on time, task type, and agent dimensions, providing a data foundation for subsequent data analysis.

[0067] S53: The system analyzes the execution data through machine learning algorithms to explore the combination methods, cooperation order and data exchange patterns of agents under different task types. In S53, the system will use machine learning algorithms to perform in-depth analysis of the structured execution data in the system's experience knowledge base. First, it will use clustering algorithms to classify tasks according to semantic features, skill requirement maps, and execution process structures to identify different task types. Then, it will use frequent pattern mining algorithms to mine the execution data of each type of task and extract core collaborative information such as the optimal combination of agents, collaboration order, and data exchange patterns under different task types.

[0068] S54 encapsulates the identified collaboration patterns of multiple agents into a virtual team template containing preset agent combinations and collaboration process information, and improves the preset agent combinations and collaboration process information of the virtual team template for the formation and planning of agent teams for similar tasks.

[0069] In S54, the system will encapsulate the agent collaboration patterns discovered in S53 into standardized virtual team templates. The templates contain information such as preset agent combinations adapted to the corresponding task types, default collaboration processes, and historically optimal resource configuration parameters. At the same time, the system will integrate newly discovered collaboration patterns with existing virtual team templates, improving the preset agent combinations and collaboration process information in the templates to make the templates more adaptable to task processing needs. The encapsulated virtual team templates will be stored in a special namespace of the agent registration and discovery mechanism. When a planning agent receives a similar task in the future, it can prioritize retrieving the matching virtual team template, directly load the predefined team structure and collaboration framework, and only need to make minor adjustments for task-specific details.

[0070] In this embodiment, by collecting multi-dimensional execution data throughout the entire task lifecycle, a rich, comprehensive, and accurate data source is provided for the system's learning and optimization. This data covers the entire task processing process, enabling the system to extract effective information from practice. By anonymizing and structuring the execution data and storing it in a distributed system experience knowledge base, user privacy and security are protected, while efficient data storage and rapid retrieval are achieved, laying a solid foundation for subsequent data analysis. Machine learning algorithms are used to mine agent collaboration patterns under different task types, allowing the system to extract valuable information from a large amount of execution data, realizing the transformation from "data" to "knowledge." By generating and improving virtual team templates, the reuse of agent collaboration patterns is realized, enabling planning agents to quickly form execution teams and formulate execution processes when handling similar tasks. This significantly shortens planning time, improves task planning efficiency and agent collaboration efficiency, and also gives the system the ability to accumulate knowledge. Example 9

[0071] Optionally, the system utilizes execution data from the system's experience knowledge base, using task completion time, resource utilization efficiency, and user satisfaction as reward signals. It optimizes the task decomposition strategy model of the planning agent and the intervention influence domain judgment model of the supervised agent through reinforcement learning algorithms. At the same time, it optimizes the parsing and evaluation accuracy of the semantic parsing model and the complexity evaluation model through supervised learning.

[0072] In this embodiment, the system employs reinforcement learning algorithms to iteratively optimize the task decomposition strategy model of the planning agent and the intervention influence domain judgment model of the supervisory agent. The decision-making processes of the two models are modeled as Markov decision processes. The state space of the task decomposition strategy model of the planning agent includes task characteristics, available agent states, and resource states, while the action space includes task decomposition strategies and agent selection methods. The state space of the intervention influence domain judgment model of the supervisory agent includes task execution states, intervention instruction characteristics, and task graph structure, while the action space includes influence domain judgment results and intervention strategy selection. The reward signal of reinforcement learning is composed of multiple indicators such as task completion time, resource utilization efficiency, and user satisfaction. The shorter the task completion time, the higher the resource utilization efficiency, and the higher the user satisfaction, the higher the corresponding reward value, and vice versa. The system uses reinforcement learning algorithms such as policy gradient and Q-learning, and utilizes the execution data in the system's experience knowledge base to iteratively train and update the two models, enabling the models to continuously optimize decision-making strategies and improve the accuracy and scientific nature of task decomposition and influence domain judgment.

[0073] Meanwhile, the system employs supervised learning algorithms to optimize the semantic parsing model and complexity evaluation model of the coordinating agent. High-quality execution data from the system's experience knowledge base, which has been manually or automatically labeled, is used as training samples. The task requests in the samples are taken as input, and the labeled intent results and the true complexity values ​​are taken as output. The semantic parsing model and complexity evaluation model are continuously trained, and the model parameters are continuously optimized to improve the model's accuracy in parsing task requests and the precision in evaluating complexity.

[0074] In some embodiments, after the system optimizes the task decomposition strategy model of the planning agent using reinforcement learning algorithms, for complex tasks such as "data analysis and report generation," the model can more reasonably decompose sub-tasks and select more suitable agent combinations, effectively shortening the task completion time by more than 20%. After optimizing the intervention influence domain judgment model of the supervised agent using reinforcement learning algorithms, the model can more accurately determine the intervention influence domain, reducing unnecessary global replanning by more than 40% and improving the efficiency of human-machine intervention. After optimizing the semantic parsing model through supervised learning, the model's parsing accuracy for colloquial and ambiguous task requests is improved by more than 25%.

[0075] In this embodiment, reinforcement learning algorithms are used to iteratively optimize the core decision-making models of the planning agent and the supervised agent using multi-dimensional quantitative indicators as reward signals. This allows the models to learn autonomously and evolve continuously from a large number of task execution practices, constantly improving the scientificity and accuracy of task decomposition and influence domain judgment, and further enhancing the system's task processing efficiency and human-machine collaboration efficiency. Supervised learning algorithms are used to continuously optimize the semantic parsing model and complexity evaluation model, improving the model's parsing and evaluation accuracy, ensuring the accuracy of task request parsing and routing distribution, and laying a solid foundation for subsequent task processing. Overall, through continuous optimization of each core model, the system forms a complete closed loop of autonomous learning and evolution, continuously improving its intelligence level in handling complex and open tasks through ongoing practice, and continuously optimizing the system's performance. Example 10

[0076] On the other hand, an operating system interaction system based on multi-agent cooperation is provided, which is used to implement the above-mentioned operating system interaction method based on multi-agent cooperation.

[0077] In this embodiment, the management and coordination agent and the domain execution agent are the core functional entities of the system, undertaking core tasks such as task parsing, planning, execution, and monitoring; the agent registration and discovery module, the standardized communication module, and the shared context storage module are the basic support modules of the system, providing basic guarantees for the collaborative interaction of agents; the system experience knowledge base is the data source support module of the system, used to store the execution data of the entire task lifecycle; the learning and optimization module is the core module of the system's autonomous learning and evolution, undertaking tasks such as collaborative pattern mining, template generation, and model optimization.

[0078] Each module and agent interacts with each other through an asynchronous message communication mechanism. This mechanism defines a standardized message envelope format that encapsulates instructions, data, status and control signals, thereby achieving decoupled communication between agents and enabling each agent to be functionally cohesive and autonomous.

[0079] Management and coordination agents include coordination agents, planning agents, and supervisory agents, which are used to perform semantic parsing and routing of task requests, planning and decomposition of complex tasks, and task scheduling and intervention processing, respectively. Domain-specific intelligent agents include basic functional intelligent agents and composite domain intelligent agents. Composite domain intelligent agents have built-in internal coordination modules, internal task planning modules, and a set of sub-intelligent agents for the recursive management and execution of sub-tasks. The learning optimization module is used to extract task execution data from the system experience knowledge base, mine the agent collaboration patterns and generate a virtual team template containing preset agent combinations and collaboration process information, and is also used to optimize the decision models of each agent. The supervisory agent incorporates an influence domain assessment engine and a dynamic scheduling engine; among them... The impact domain assessment engine is configured to determine the intervention impact domain based on the node dependency relationship of the task flowchart and real-time execution stack information, through dependency path propagation analysis, analyze the impact range and degree of user intervention commands, and output the assessment results of local adjustment, subtask replanning or global replanning. The dynamic scheduling engine is used to maintain the task ready queue and perform intelligent scheduling operations based on subtask dependencies and priorities. The system also sets up intervention transaction logs and consistency guarantee protocols to ensure the traceability of intervention operations and the consistency of the global state.

[0080] In the management and coordination type of intelligent agent, the coordinating agent, planning agent, and supervisory agent respectively undertake the functions of task parsing and routing, planning decomposition, and scheduling intervention, forming a complete management and coordination link. In the domain execution type of intelligent agent, the basic functional agent undertakes the execution of atomic tasks, and the composite domain agent realizes the recursive management and execution of composite sub-tasks through its internal coordination, planning, and execution modules. The learning and optimization module extracts execution data from the system's experience knowledge base. On the one hand, it uses machine learning algorithms to mine the agent collaboration patterns and generate and improve the virtual team template. On the other hand, it uses the execution data to iteratively optimize the core decision-making models of each agent, realizing the system's autonomous learning and evolution. The supervisory agent internally comprises two core engines: an influence domain assessment engine and a dynamic scheduling engine. The influence domain assessment engine includes semantic parsing, dependency analysis, and path propagation modules, responsible for parsing intervention commands, assessing the influence domain, and outputting adjustment strategies. The dynamic scheduling engine includes queue management and intelligent scheduling modules, responsible for maintaining the task-ready queue and intelligently scheduling subtasks. The system's intervention transaction log records all intervention commands, assessment processes, and execution operations, ensuring auditability and traceability of intervention operations. The consistency guarantee protocol employs a two-phase commit or distributed transaction coordination mechanism to ensure the atomicity of subtask flow adjustments and the consistency of the global task state when delegating internal replanning to the composite agent. Simultaneously, the system sets timeout and rollback mechanisms for each intervention operation to ensure the reliability of intervention processing.

[0081] In this embodiment, by constructing a complete system architecture encompassing management coordination, execution, basic support, data storage, and learning optimization, the methods of each embodiment are implemented, enabling the system to possess full-process capabilities for complex task recursive decomposition, dynamic planning, real-time human-machine intervention, and autonomous learning evolution, thus overcoming many shortcomings of existing technologies. Through asynchronous message communication mechanisms for data interaction among modules and intelligent agents, decoupled communication between agents is achieved, ensuring the loose coupling characteristics of the system and giving it good scalability and maintainability; new intelligent agents can be plugged and played. Furthermore, by configuring internal coordination and planning modules and sub-intelligent modules for composite domain intelligent agents... The aggregation of agents enables the system to recursively manage and execute complex sub-tasks within a specialized domain, enhancing its ability to handle intricate and specialized tasks. By incorporating a learning optimization module, the system can learn collaborative patterns and optimize decision-making models from task execution practices, achieving autonomous learning and evolution, and continuously improving its intelligence level and task processing efficiency. Furthermore, by configuring a dual-core engine for the supervisory agent and establishing intervention transaction logs and consistency guarantee protocols, the system achieves professional and precise task scheduling and intervention processing, ensuring the stability and efficiency of human-machine collaboration, while also ensuring the traceability of intervention operations and the consistency of the global state.

[0082] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0083] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for operating system interaction based on multi-agent cooperation, characterized in that, The method includes: Step 101: Construct a multi-level intelligent agent collaboration system. The system includes a management intelligent agent and a domain execution intelligent agent. The domain execution intelligent agent includes a basic function intelligent agent and a composite domain intelligent agent. The composite domain intelligent agent includes an internal task planning module and a set of sub-intelligent agents, which is used to start the internal task planning module and schedule the set of sub-intelligent agents to execute the sub-task when receiving a composite sub-task. Step 102: Receive the user's task request, perform semantic parsing and complexity evaluation on the task request through a coordinating agent, and route the task to the corresponding domain execution agent or planning agent based on the evaluation results. Step 103: In response to complex tasks, the planning agent parses the task and generates a structured task flowchart, identifies composite subtask nodes in the task flowchart, and assigns the composite subtask nodes to the corresponding composite domain agent, which then performs recursive task decomposition through the internal task planning module. Step 104: The supervisory agent performs task scheduling and status monitoring according to the task flowchart. During task execution, it receives user intervention instructions. Based on the node dependency relationship and execution stack information in the task flowchart, it determines the intervention impact domain through dependency path propagation analysis. Based on the impact domain, it selects to perform local task adjustment, subtask replanning or global task replanning so that the task flow continues to execute after adjustment. Step 105: After completing the task, summarize the results and deliver them to the user. At the same time, collect the execution data of the entire task lifecycle, identify the collaboration mode of multiple agents from the execution data, and encapsulate the collaboration mode into a virtual team template containing preset agent combinations and collaboration process information for the formation and planning of agent teams for similar tasks.

2. The method of claim 1, wherein, The management agent includes a coordinating agent, a planning agent, and a supervising agent. The multi-level agent collaboration system is equipped with a unified agent registration and discovery mechanism, a standardized message communication mechanism, and a shared context storage module. Various agents interact collaboratively based on the mechanisms and modules, and new agents can be connected to the multi-level agent collaboration system. 3.The multi-agent collaboration based operating system interaction method of claim 1, wherein, When the planning agent identifies that the skill complexity required for a subtask exceeds a preset threshold, the subtask is marked as a composite subtask node.

4. The method of claim 1, wherein, Step 102 includes: S21, the coordinating agent integrates a lightweight semantic analysis engine, and performs real-time semantic parsing of the task request by fusing rule matching based on keywords and templates and fine-tuning the intent classification model through the semantic analysis engine; S22, the coordinating agent generates a comprehensive score by comprehensively considering the number of verb actions in the instruction, the complexity of the object entity, and the implicit contextual dependencies through a complex quantification evaluation algorithm, and determines whether the task is a simple atomic task or a complex composite task based on the comprehensive score. S23. If the task is determined to be a simple atomic task, the task is routed to the corresponding domain execution agent. If the task is determined to be a complex composite task, the task is routed to the planning agent.

5. The method of claim 1, wherein, The composite domain agent also includes an internal coordination module, which is used to receive composite subtasks and schedule the internal task planning module and the set of sub-agents. The internal task planning module is used to initiate the planning execution process, perform secondary decomposition of composite subtasks and generate sub-level task flowcharts. This recursive decomposition process can be carried out in multiple levels until the leaf tasks at the bottom level can be directly processed by the basic functional agents.

6. The method of claim 1, wherein, Step 103 includes: S31, the planning agent deconstructs the complex task description through the deep semantic understanding module and extracts the core objectives, constraints and implicit skill requirement map of the task; S32, the planning agent selects agents from the global agent pool to form an execution team through a skill matching algorithm based on vector similarity; S33, the planning agent uses graph planning technology to generate a directed acyclic task flowchart that includes the execution order of subtasks, dependencies and data flow; S34, the planning agent identifies the composite subtask nodes in the task flowchart and assigns the composite subtask nodes to the corresponding composite domain agent, which then executes recursive task decomposition through the internal task planning module.

7. The method of claim 1, wherein the plurality of agents are configured to interact with the operating system based on a plurality of operating system interaction rules. Step 104 includes: S41, the supervisory agent maintains a dynamic priority scheduling queue according to the task flowchart, assigns tasks in real time according to the readiness status of sub-tasks, and continuously receives status feedback and resource consumption data from each agent through the full-process monitoring center. S42, the supervisory agent intercepts the user's intervention command through the high-priority intervention command channel, and determines the intervention impact domain based on the node dependency relationship and execution stack information in the task flowchart through dependency path propagation analysis. S43, If it is a local parameter adjustment, the supervisory agent notifies the corresponding execution agent to update the task context so that the task flow can continue to be executed after the adjustment; S44, If it is an internal adjustment of a composite subtask, the supervising agent entrusts the corresponding composite domain agent to perform internal replanning so that the task flow continues to be executed after the adjustment. S45, if the global target change is involved, the supervisory agent will send the task back to the planning agent to initiate an overall replanning, so that the task process can continue to be executed after the adjustment.

8. The operating system interaction method based on multi-agent cooperation according to claim 1, characterized in that, Step 105 includes: S51. After the task is completed, the results are summarized and delivered to the user. At the same time, the execution data of the entire task lifecycle is collected. The task execution data includes the original task request, task flowchart, execution logs of each node, agent collaboration records, user intervention records and performance indicators. S52, the execution data is anonymized and structured, and then persistently stored in the system experience knowledge base; S53, the system analyzes the execution data through machine learning algorithms to explore the combination methods, cooperation order and data exchange patterns of intelligent agents under different task types; S54, the identified collaboration patterns of multiple agents are encapsulated into a virtual team template containing preset agent combinations and collaboration process information, and the preset agent combinations and collaboration process information of the virtual team template are improved for the formation and planning of agent teams for similar tasks.

9. The operating system interaction method based on multi-agent cooperation according to claim 8, characterized in that, The system utilizes execution data from the system's experience knowledge base, using task completion time, resource utilization efficiency, and user satisfaction as reward signals. It optimizes the task decomposition strategy model of the planning agent and the intervention influence domain judgment model of the supervised agent through reinforcement learning algorithms. At the same time, it optimizes the parsing and evaluation accuracy of the semantic parsing model and the complexity evaluation model through supervised learning.

10. An operating system interaction system based on multi-agent cooperation, characterized in that, The system is used to implement the operating system interaction method based on multi-agent cooperation as described in any one of claims 1-9, including a management and coordination agent, a domain execution agent, an agent registration and discovery module, a standardized communication module, a shared context storage module, a system experience knowledge base, and a learning and optimization module; Each module interacts with data through an asynchronous message communication mechanism to support decoupled communication between the agents; The management and coordination intelligent agent includes a coordination intelligent agent, a planning intelligent agent, and a supervisory intelligent agent, which are respectively used to perform semantic parsing and routing of task requests, planning and decomposition of complex tasks, and task scheduling and intervention processing. The domain execution agent includes a basic functional agent and a composite domain agent. The composite domain agent has a built-in internal coordination module, an internal task planning module, and a set of sub-agents for the recursive management and execution of sub-tasks. The learning optimization module is used to extract task execution data from the system experience knowledge base, mine the agent collaboration mode and generate a virtual team template containing preset agent combinations and collaboration process information, and at the same time optimize the decision model of each agent. The supervisory agent incorporates an influence domain analysis engine and a dynamic scheduling engine; wherein... The influence domain assessment engine is configured to determine the intervention influence domain based on the node dependency relationship of the task flowchart and real-time execution stack information, through dependency path propagation analysis, analyze the influence range and degree of user intervention instructions, and output the assessment results of local adjustment, subtask replanning or global replanning. The dynamic scheduling engine is used to maintain the task ready queue, perform intelligent scheduling operations based on subtask dependencies and priorities, and the system sets up intervention transaction logs and consistency guarantee protocols to achieve the traceability of intervention operations and the consistency of the global state.

Images