Intelligent agent autonomous cooperation method and system based on group hierarchical cooperation

Abstract

The invention relates to the technical field of artificial intelligence and multi-agent cooperative control, in particular to an agent autonomous cooperation method and system based on group hierarchical cooperation, and the method comprises the steps: responding to a natural language task request input by a user; generating a preliminary task execution plan based on the task description data, selecting expert agents matched with sub-tasks in the task execution plan from the group, and generating a task allocation instruction to perform task cooperation; and each expert agent receives the task allocation instruction to execute the corresponding sub-task, writes an execution state, a stage result and intermediate data into a group memory space in real time in an execution process, and performs semantic integration and induction after all the sub-tasks are executed to generate a unified task output result. Under the unified command of the super agent, each expert agent can realize autonomous cooperation and real-time information sharing based on the group memory space, and realize adaptive dynamic re-planning when the task is abnormal or the environment is changed.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and multi-agent cooperative control technology, and in particular to an agent autonomous cooperation method and system based on group hierarchical cooperation. Background Technology

[0002] Against the backdrop of rapid development in artificial intelligence technology, multi-agent systems have become an important research direction for achieving division of labor and collaboration in complex tasks. Existing multi-agent collaborative systems are mostly based on rule-driven workflows or centralized scheduling architectures. Their core idea is to statically orchestrate the roles, responsibilities, and task execution paths of each agent during the system design phase. These systems typically use task graphs or state machine models to achieve task decomposition and process control, allocating and coordinating sub-tasks through a single scheduling center. While this system can achieve controllable parallel execution in stable task environments with clear objectives, its adaptability and flexibility are clearly insufficient in complex and dynamic open environments.

[0003] However, with the increasing reasoning and generation capabilities of large language models, traditional multi-agent architectures based on fixed rules and centralized control are gradually revealing systemic bottlenecks. On the one hand, static task planning patterns struggle to cope with environmental changes and task uncertainties. When external inputs or operational states change, the system cannot dynamically adjust based on real-time information, leading to rigid execution processes and increased task failure rates. On the other hand, centralized scheduling mechanisms make the entire system highly dependent on a single core agent or central node. If this node fails, it will directly trigger a cascading failure of the system. Furthermore, communication and context synchronization between multiple agents still primarily rely on global broadcasting or full history sharing mechanisms, resulting in high bandwidth overhead, severe information redundancy, and low inference efficiency. Summary of the Invention

[0004] The purpose of this invention is to address the problems of insufficient task execution flexibility, inadequate environmental awareness, and large amounts of repetitive development work in existing intelligent agent cooperative technologies. It proposes an autonomous cooperative method and system based on hierarchical group collaboration, aiming to achieve dynamic task allocation, autonomous planning, and efficient collaborative execution among multiple intelligent agents. This objective can be achieved through the following technical solutions: This invention provides a method for autonomous agent cooperation based on hierarchical group collaboration, comprising the following steps: In response to user input of natural language task requests, the task semantics are parsed to extract the task objectives, constraints and contextual semantic information, forming structured task description data; Based on task description data, the super agent retrieves task-related knowledge nodes and execution state paths from the knowledge graph and state transition graph, generates a preliminary task execution plan containing execution order and dependencies, selects expert agents that match the subtasks in the task execution plan from the group, and generates task allocation instructions for task collaboration. Each expert agent receives task assignment instructions and executes the corresponding subtasks. During the execution process, the execution status, stage results and intermediate data are written to the group memory space in real time. The super agent monitors the execution progress of expert agents based on real-time state data in the group memory space. When anomalies, failures, or changes in the external environment are detected, the task execution plan is dynamically updated and unfinished sub-tasks are reassigned. After all sub-tasks are completed, the super agent performs semantic integration and summarization of the task results in the group memory space to generate a unified task output result.

[0005] Furthermore, in response to the user's input of a natural language task request, the task semantics are parsed to extract the task's objectives, constraints, and contextual semantic information, forming structured task description data, including: Responding to user input natural language task requests, the system acquires the task text and performs preprocessing, including word segmentation, part-of-speech tagging, and sentence normalization. Based on the semantic parsing model, the preprocessed task text and intent are identified and semantic roles are labeled to extract the task objectives, constraints and key operation objects. The extracted task objectives, constraints, and key operational objects are semantically associated and matched with the contextual information of historical tasks in the group memory space to form a context-enhanced representation of the task. Based on the context-enhanced representation, structured task description data is generated and mapped to a standard task data format for the super agent to perform knowledge retrieval and task planning.

[0006] Furthermore, based on task description data, the superagent retrieves task-related knowledge nodes and execution state paths from the knowledge graph and state transition graph, generating a preliminary task execution plan that includes execution order and dependencies, including: The super intelligent agent retrieves entity nodes, relation edges, and semantic tags related to the task from the knowledge graph based on the task objectives and key operational objects to determine the executable set of knowledge associations. The knowledge graph is constructed based on a semantic network structure and is a multi-dimensional knowledge graph formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model. It is used to represent the semantic associations and causal relationships between key operational objects. Based on constraints and contextual semantic information, the initial state, target state, and execution state path related to the task objective are identified in the state transition graph. The state transition graph is constructed based on task execution records and operation procedures. Through a knowledge extraction model, each state node and its corresponding transition conditions during task execution are extracted, and the transition relationship between state nodes is established to describe the state change path of the task at different execution stages. The knowledge association set obtained from the knowledge graph is semantically fused with the state path identified in the state transition graph to construct the state dependency chain of the task. Based on the state dependency chain and the logical association between each subtask, a preliminary task execution plan containing execution order, resource dependencies and condition constraints is generated.

[0007] Furthermore, based on the construction of a semantic network structure, a multi-dimensional knowledge graph is formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model, including: The response to the file upload interface of the knowledge platform retrieves multi-source knowledge files, including technical manuals, operation and maintenance procedures, and log records. The document parsing tool performs structured parsing of multi-source knowledge files, performs OCR text recognition and layout analysis to identify the correspondence between title levels, body paragraphs and pure image areas, form hierarchical text structure data and store it according to segmented text, which is used as the input set for the knowledge extraction model. Based on the knowledge extraction model, semantic parsing and knowledge recognition are performed on segmented text to extract corresponding entity, relation and attribute information, and to establish semantic triples of entity-relation-entity; where relation represents the operation behavior in task execution and corresponds to the action node of the task. Semantic triples are transformed into a relational structure between multi-source knowledge nodes, forming a multi-dimensional knowledge graph in the form of a semantic network as a knowledge graph.

[0008] Furthermore, based on task execution records and operating procedures, a knowledge extraction model is used to extract each state node and its corresponding transition conditions during task execution, establishing transition relationships between state nodes, including: The response knowledge platform receives task execution records and standard processing flow documents, which include emergency manuals, operating procedures, and equipment maintenance records. Preprocess the task execution records and standard processing flow documents, perform OCR recognition and text structure analysis, and extract semantic fragments containing operation steps, event triggering conditions and result descriptions; Based on the knowledge extraction model, semantic fragments are identified and classified for event recognition, key event nodes and their dependencies in the task execution process are determined, and state recognition model is used to identify the state nodes and corresponding transition conditions of each event node, and state-event-state transition triples are constructed. Transition triples are transformed into a multidimensional directed graph structure between state nodes, forming a set of state transition paths as a state transition graph.

[0009] Furthermore, the knowledge association set obtained from the knowledge graph is semantically fused with the state paths identified in the state transition graph to construct the task's state dependency chain. Based on the state dependency chain and the logical relationships between subtasks, a preliminary task execution plan is generated, including execution order, resource dependencies, and condition constraints, including: Obtain the knowledge association set in the knowledge graph and the state path set in the state transition graph and perform semantic alignment. Based on the semantic similarity calculation results between the task objective and the context, perform semantic pairing between state nodes and knowledge nodes to form a state-knowledge mapping relationship. Based on the mapping relationship, establish the sequential dependencies between each action node and state node in the task, and construct the state dependency chain of the task. Based on the logical order and constraints of nodes in the state dependency chain, parallelizable subtasks and dependent subtasks in the task are identified, and the execution topology diagram of the task is generated. Based on the topology diagram, combined with resource dependencies, execution constraints and task priorities, the execution order, parallel relationship and triggering conditions of each subtask are determined, a preliminary task execution plan is generated and stored in the group memory space.

[0010] Furthermore, each expert agent receives task assignment instructions and executes the corresponding subtasks. During execution, the execution status, stage results, and intermediate data are written to the group memory space in real time, including: In response to the task allocation instructions, each expert agent parses the corresponding subtask content and execution parameters; Based on the execution order, resource dependencies, and condition constraints in the preliminary task execution plan, the corresponding sub-tasks are executed. During the task execution process, the execution status is monitored in real time, and the execution status, stage results, and intermediate data are written into the group memory space in a structured form, along with task identifiers, timestamps, and execution context information, for the super agent to perform global task status perception and progress tracking.

[0011] Furthermore, the super agent monitors the execution progress of the expert agent based on real-time state data from the group memory space. When anomalies, failures, or changes in the external environment are detected, the super agent dynamically updates the task execution plan and reallocates unfinished sub-tasks, including: Periodically read the task status data in the group memory space to obtain the execution progress, resource utilization and stage results of each expert agent; When an execution exception, task timeout, or change in external environment parameters is detected, a new task planning request is generated based on context information and dependencies. Based on the replanning request, adjust the state dependency chain and task priority in the task execution plan, and reallocate incomplete or failed subtasks to available expert agents.

[0012] Based on the same inventive concept, this invention provides an intelligent agent autonomous cooperation system based on group hierarchical cooperation, employing the intelligent agent autonomous cooperation method described above, including: The response module is used to respond to the natural language task requests input by the user, parse the task semantics to extract the task objectives, constraints and contextual semantic information, and form structured task description data; The data processing module is used to retrieve task-related knowledge nodes and execution state paths from the knowledge graph and state transition graph based on task description data, generate a preliminary task execution plan including execution order and dependencies, select expert agents that match the subtasks in the task execution plan from the group, and generate task allocation instructions for task collaboration; each expert agent receives the task allocation instructions and executes the corresponding subtask, and writes the execution status, stage results and intermediate data into the group memory space in real time during the execution process; The results output module is used by the super agent to monitor the execution progress of the expert agent based on real-time state data in the group memory space. When an anomaly, failure, or change in the external environment is detected, the task execution plan is dynamically updated and unfinished sub-tasks are reallocated. After all sub-tasks are completed, the super agent performs semantic integration and summarization of the task results in the group memory space to generate a unified task output result.

[0013] Furthermore, the data processing module includes: The knowledge retrieval unit is used by the super intelligent agent to retrieve entity nodes, relation edges, and semantic tags related to the task from the knowledge graph based on the task objectives and key operation objects, so as to determine the executable set of knowledge associations. The knowledge graph is constructed based on a semantic network structure and is a multi-dimensional knowledge graph formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model. It is used to represent the semantic associations and causal relationships between key operation objects. The state recognition unit is used to identify the initial state, target state, and execution state path related to the task objective in the state transition graph based on constraints and contextual semantic information. The state transition graph is constructed based on task execution records and operation procedures. It extracts each state node and its corresponding transition conditions during task execution through a knowledge extraction model, establishes the transition relationship between state nodes, and describes the state change path of the task at different execution stages. The task planning unit is used to semantically fuse the set of knowledge associations obtained from the knowledge graph with the state paths identified in the state transition graph to construct the state dependency chain of the task; based on the state dependency chain and the logical associations between each subtask, it generates a preliminary task execution plan that includes execution order, resource dependencies and condition constraints.

[0014] Compared with the prior art, the present invention has at least one of the following technical advantages: This invention achieves efficient, flexible, and intelligent task execution in multi-agent systems by constructing an autonomous collaboration mechanism based on hierarchical group cooperation. Unlike traditional multi-agent systems that rely on fixed processes and centralized scheduling, this invention, under the unified command of a super-agent, enables individual expert agents to autonomously collaborate and share information in real-time based on a group memory space. This allows for adaptive dynamic replanning in the event of task anomalies or environmental changes. It enhances environmental awareness and task robustness, reduces the workload of manual orchestration and repetitive development, improves the reusability of agent modules, and enhances the accuracy of task execution and the overall system reliability through a semantically consistent collaboration protocol. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below: Figure 1 This is a flowchart illustrating the autonomous collaboration method of intelligent agents based on group hierarchical collaboration of the present invention. Figure 2 This is a flowchart illustrating the task state transition and knowledge fusion of the intelligent agent in an embodiment of the present invention; Figure 3 This is the semantic relation structure of equipment fault knowledge in the embodiments of the present invention; Figure 4 This is a flowchart illustrating the fault handling state transition in an embodiment of the present invention; Figure 5 This is a visual view of the task output results in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0017] First Embodiment As Large Language Models (LLMs) enter the era of trillions of parameters (such as cutting-edge models like GPT-5 and Claude 3), the intelligent collaboration capabilities of Multi-Agent Systems (MAS) face new technical bottlenecks. Existing agent collaboration systems generally adopt a task orchestration paradigm of "predefined processes + fixed roles," meaning that task execution is driven by manually set rules or static graph structures. This centralized architecture is controllable in small-scale, deterministic scenarios, but it gradually reveals the following limitations in dynamic and complex task environments: (1) As pointed out in the Nature Machine Intelligence (2024) study, existing rule-driven LLM task planning systems rely on static decomposition graphs to perform tasks. The functions of the agent are strictly limited to the scope of preset sub-tasks (such as navigation, grasping, recognition, etc.). When the task requirements or environmental conditions change, the system cannot adaptively reconstruct the task flow and requires a lot of resources to retrain the module or manually adjust the rules.

[0018] (2) Meta AI’s multi-agent collaboration platform released in 2025 adopts a single-point centralized decision-making architecture, with the master control model undertaking all reasoning and scheduling tasks, resulting in a system communication bandwidth occupancy rate of up to 80%. Furthermore, when a decision node fails, the entire system experiences cascading failures, severely impacting task stability and real-time response capabilities.

[0019] (3) Based on the experimental results of Google DeepMind (2024) and Stanford HAI (2025), the computational cost of the "memory-enhanced agent" with the full history sharing mechanism in a 50-node cluster is O(n). 2 The growth trend shows that model inference only accounts for 30% of computing power utilization, with the remaining 70% consumed by redundant protocols and context synchronization operations, which greatly limits the realization of the potential of large models. Addressing the current situation where intelligent agent collaboration technology in the industry still primarily relies on traditional workflow methods for task orchestration, the inventors, through in-depth research, discovered that this model often requires manual pre-setting of complete task steps and execution sequences. Once the external environment or task conditions change, the system cannot adjust the process in real time through self-awareness mechanisms, resulting in poor task execution flexibility, low efficiency in exception handling, and the need for repeated development and rule reconstruction in new task scenarios, leading to extremely high maintenance costs. The inventors proposed an intelligent agent autonomous collaboration method based on group hierarchical collaboration. Through intelligent agent group conversations, multiple intelligent agents are organized into a chat, with a super intelligent agent within the group directing the collaborative work among multiple agents. Ultimately, the user-input task requirements are broken down into objectives, planned step-by-step, and executed, moving away from fixed-scheduling workflows and greatly improving the reusability of intelligent agents. Furthermore, the super intelligent agent can flexibly implement task execution according to different demand scenarios.

[0020] This method draws inspiration from human team collaboration mechanisms, dividing the intelligent agent system into a two-tiered structure of a super-intelligent agent layer and an expert intelligent agent layer. It constructs a dynamic collaboration framework supported by knowledge graphs and state transition graphs. The super-intelligent agent is responsible for semantic parsing and hierarchical planning of the user's natural language tasks, generating an adaptive task execution plan through knowledge retrieval, state awareness, and dependency reasoning. Expert intelligent agents autonomously collaborate to execute sub-tasks based on task allocation instructions, and synchronize their execution status in real time to the group memory space, allowing the super-intelligent agent to dynamically adjust task paths and scheduling strategies. The specific implementation is as follows: like Figure 1 As shown, this invention provides an agent autonomous cooperation method based on group hierarchical cooperation, comprising the following steps: In response to user input of natural language task requests, the task semantics are parsed to extract the task objectives, constraints and contextual semantic information, forming structured task description data; Based on task description data, the super agent retrieves task-related knowledge nodes and execution state paths from the knowledge graph and state transition graph, generates a preliminary task execution plan containing execution order and dependencies, selects expert agents that match the subtasks in the task execution plan from the group, and generates task allocation instructions for task collaboration. Each expert agent receives task assignment instructions and executes the corresponding subtasks. During the execution process, the execution status, stage results and intermediate data are written to the group memory space in real time. The super agent monitors the execution progress of expert agents based on real-time state data in the group memory space. When anomalies, failures, or changes in the external environment are detected, the task execution plan is dynamically updated and unfinished sub-tasks are reassigned. After all sub-tasks are completed, the super agent performs semantic integration and summarization of the task results in the group memory space to generate a unified task output result.

[0021] Specifically, when a user submits a task request in natural language, the input natural language task request is parsed by a multi-agent platform. The super agent automatically understands the task semantics, decomposes the objectives, and generates a task planning graph. Then, expert agents from various fields are dynamically assigned to work together to achieve automatic execution and intelligent analysis of the task.

[0022] In this embodiment, expert agents include, but are not limited to, the following types: Security protection intelligent agent: responsible for monitoring the system's security status, performing risk identification and emergency response strategy formulation; Equipment operation intelligent agent: used to collect and analyze equipment operation data, and perform status diagnosis and health assessment; Operation and maintenance task intelligent agent: responsible for task planning, work order generation and execution scheduling; Personnel Management Intelligent Agent: Used to identify personnel operation status and perform permission matching and task assignment; Intelligent environmental monitoring agent: Real-time acquisition and analysis of environmental parameters such as temperature, humidity, and voltage; Energy management intelligent agent: responsible for power consumption monitoring, energy efficiency analysis and energy dispatch optimization; Intelligent logistics and warehousing agents: perform material flow monitoring, inventory analysis, and scheduling optimization; Anomaly detection agent: performs multimodal anomaly identification on sensor data and logs; Knowledge-based question-answering intelligent agents: providing semantic interpretation and technical support for tasks based on knowledge graphs; Data statistical analysis intelligent agent: performs multidimensional data aggregation, visualization, and report generation; Code-generating intelligent agents: automatically generate or optimize scripts or interface logic for analysis and control; Model optimization agent: performs algorithm model training, parameter optimization, and performance evaluation; Predictive maintenance agent: Enables early warning of equipment failures through trend modeling.

[0023] Each expert agent shares information and synchronizes its state through a unified conversation protocol. The results are integrated and analyzed by the super agent and output to the intelligent data analysis module on the right, enabling multi-dimensional intelligent data processing such as trend diagnosis, fault prediction, performance evaluation, and report generation.

[0024] The operation of this method resembles the working mode of a "multi-layered intelligent team"—when a user inputs a natural language task request, the super agent first assumes the role of command and understanding. Like a project manager, it analyzes the user's intent, identifies the task objective, constraints, and contextual information from the input, and transforms this information into a structured data description, laying the foundation for subsequent knowledge retrieval and task planning. After obtaining the task description, the super agent further invokes the knowledge graph and state transition graph to retrieve the required knowledge content and execution path. The knowledge graph acts as a knowledge brain, used to understand the objects, relationships, and behavioral logic involved in the task; the state transition graph reflects the possible state changes and condition transitions during task execution. Through comprehensive analysis of both, the super agent can automatically generate a preliminary task execution plan containing the execution order, dependencies, and resource conditions. Subsequently, the system selects suitable expert agents within the group, assigns them their respective sub-tasks, and initiates collaborative work through standardized instruction protocols. During the task execution phase, each expert agent independently completes its assigned sub-task and synchronizes its execution status, stage results, and intermediate data to the group's memory space in real time. The super agent continuously monitors the data flow in this shared space. When it detects anomalies, failures, or changes in the external environment, it can instantly adjust the task execution plan and reallocate sub-tasks, thus forming a closed-loop collaborative mechanism of "real-time perception—dynamic decision-making—adaptive reconfiguration." This mechanism enables the system to maintain continuous operation even under unexpected events (such as equipment failure or task blockage), significantly reducing the need for human intervention. Therefore, three innovations have been achieved in this process: First, in terms of environmental adaptability, by introducing group negotiation and state feedback, it can autonomously reconfigure strategies, possessing the flexible decision-making capabilities of a human-like team; second, in terms of development efficiency, a flexible architecture of "main agent + functional modules" is adopted, encapsulating core functions as standard components, and new tasks can be quickly adapted simply by replacing the dialogue template; finally, in terms of collaboration mode, a hierarchical decision-making system has been constructed—the super agent is responsible for task decomposition and planning, while expert agents collaborate through standardized conversation protocols, truly realizing autonomous collaboration and knowledge sharing among multiple agents.

[0025] Furthermore, in response to the user's input of a natural language task request, the task semantics are parsed to extract the task's objectives, constraints, and contextual semantic information, forming structured task description data, including: Responding to user input natural language task requests, the system acquires the task text and performs preprocessing, including word segmentation, part-of-speech tagging, and sentence normalization. Based on the semantic parsing model, the preprocessed task text and intent are identified and semantic roles are labeled to extract the task objectives, constraints and key operation objects. The extracted task objectives, constraints, and key operational objects are semantically associated and matched with the contextual information of historical tasks in the group memory space to form a context-enhanced representation of the task. Based on the context-enhanced representation, structured task description data is generated and mapped to a standard task data format for the super agent to perform knowledge retrieval and task planning.

[0026] Specifically, the system receives natural language task text input from users and preprocesses it, including word segmentation, part-of-speech tagging, and sentence normalization, to eliminate semantic ambiguity and standardize the expression format. For example, when a user inputs "Help me analyze the reasons for production line downtime over the past three months," the system decomposes the sentence into several semantic units, such as "analyze" (verb, task type), "past three months" (time constraint), and "reasons for production line downtime" (analysis object). Subsequently, based on a semantic parsing model, the preprocessed task text undergoes intent recognition and semantic role labeling to extract the task objective, constraints, and key operational objects. The task objective determines the core purpose the user expects to achieve (e.g., "generate a downtime reason report"), the constraints limit the task scope (e.g., "the time period is the past three months," "the data source is production logs"), and the key operational objects clarify the entities that need to be processed when performing the task (e.g., "production line A," "downtime event"). Next, the extracted task objective, constraints, and key operational objects are semantically matched with the historical task context in the group memory space to generate a context-enhanced representation of the task. For example, if a similar task such as "analyzing the reasons for downtime in the past month" has been processed in the memory space, the corresponding analysis template, commonly used indicators and knowledge nodes (such as "equipment health" and "process anomaly distribution") will be automatically retrieved and integrated into the semantic representation of the current task to improve planning accuracy and execution efficiency.

[0027] Finally, based on this context-enhanced representation, structured task description data is generated and mapped to a standardized task data format for use by the super agent for knowledge retrieval and task planning.

[0028] For example, the above natural language task will eventually be structured as follows: Task type: Analysis task Objective: Generate a production line downtime cause report. Constraints: Time range = past three months; Data source = production logs; Output format = visual report Key objects: Production line A, downtime events, equipment operation records After this transformation, the super agent can quickly access the corresponding knowledge graph and state transition path to generate a reasonable task decomposition and execution plan, thereby achieving intelligent, semantic-driven task collaboration.

[0029] Furthermore, based on task description data, the superagent retrieves task-related knowledge nodes and execution state paths from the knowledge graph and state transition graph, generating a preliminary task execution plan that includes execution order and dependencies, including: The super intelligent agent retrieves entity nodes, relation edges, and semantic tags related to the task from the knowledge graph based on the task objectives and key operational objects to determine the executable set of knowledge associations. The knowledge graph is constructed based on a semantic network structure and is a multi-dimensional knowledge graph formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model. It is used to represent the semantic associations and causal relationships between key operational objects. Based on constraints and contextual semantic information, the initial state, target state, and execution state path related to the task objective are identified in the state transition graph. The state transition graph is constructed based on task execution records and operation procedures. Through a knowledge extraction model, each state node and its corresponding transition conditions during task execution are extracted, and the transition relationship between state nodes is established to describe the state change path of the task at different execution stages. The knowledge association set obtained from the knowledge graph is semantically fused with the state path identified in the state transition graph to construct the state dependency chain of the task. Based on the state dependency chain and the logical association between each subtask, a preliminary task execution plan containing execution order, resource dependencies and condition constraints is generated.

[0030] Specifically, the super agent first retrieves task-related entity nodes, relational edges, and semantic tags from the knowledge graph based on the task objective and key operational objects. The knowledge graph acts as the system's knowledge repository, extracting and associating "entity-relationship-attribute" relationships from multi-source knowledge documents through a semantic network structure. For example, in a device maintenance scenario, if the task objective is "diagnosing the cause of UPS device failure," it will retrieve UPS-related knowledge nodes from the knowledge graph, such as "UPS action," "voltage anomaly," and "device alarm code," and identify their relational edges (e.g., "caused by," "leads to"), forming an executable set of knowledge associations to guide subsequent task planning. Next, based on the task's constraints and contextual semantic information, the super agent identifies the initial state, target state, and execution path related to the task objective in the state transition graph. The state transition graph is constructed from task execution records and operating procedures, describing the state changes and triggering conditions of the task at different stages. For example, in the above scenario, the system may identify the state sequence: "alarm code detected → manual confirmation → generate maintenance work order → perform repair → state restored", and mark the transition conditions between each node (such as "voltage restored to normal" or "device restarted successfully").

[0031] Furthermore, based on the construction of a semantic network structure, a multi-dimensional knowledge graph is formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model, including: The response to the file upload interface of the knowledge platform retrieves multi-source knowledge files, including technical manuals, operation and maintenance procedures, and log records. The document parsing tool performs structured parsing of multi-source knowledge files, performs OCR text recognition and layout analysis to identify the correspondence between title levels, body paragraphs and pure image areas, form hierarchical text structure data and store it according to segmented text, which is used as the input set for the knowledge extraction model. Based on the knowledge extraction model, semantic parsing and knowledge recognition are performed on segmented text to extract corresponding entity, relation and attribute information, and to establish semantic triples of entity-relation-entity; where relation represents the operation behavior in task execution and corresponds to the action node of the task. Semantic triples are transformed into a relational structure between multi-source knowledge nodes, forming a multi-dimensional knowledge graph in the form of a semantic network as a knowledge graph.

[0032] Furthermore, based on task execution records and operating procedures, a knowledge extraction model is used to extract each state node and its corresponding transition conditions during task execution, establishing transition relationships between state nodes, including: The response knowledge platform receives task execution records and standard processing flow documents, which include emergency manuals, operating procedures, and equipment maintenance records. Preprocess the task execution records and standard processing flow documents, perform OCR recognition and text structure analysis, and extract semantic fragments containing operation steps, event triggering conditions and result descriptions; Based on the knowledge extraction model, semantic fragments are identified and classified for event recognition, key event nodes and their dependencies in the task execution process are determined, and the state recognition model is used to identify the state nodes and corresponding transition conditions of each event node, and state-event-state transition triples are constructed. Transition triples are transformed into a multidimensional directed graph structure between state nodes, forming a set of state transition paths as a state transition graph.

[0033] This solution utilizes a self-built IE (Internet Information Center) large-scale model, i.e., a knowledge extraction model, to structurally extract textual knowledge from documents, forming key information elements such as entities, relationships, actions, text segments, and events, thus providing high-quality data for the knowledge platform. Specifically, when the super-intelligent agent performs knowledge retrieval and task planning, it first needs to rely on the knowledge platform to construct a knowledge graph and a state transition graph to support subsequent semantic reasoning and state planning.

[0034] The construction process of a knowledge graph includes the following steps: Upload the original text document to the knowledge platform via file upload; Using a self-developed PDF document parsing tool, OCR text recognition is performed on the images in the document, and layout analysis is conducted to identify the correspondence between heading levels, body paragraphs, and pure image areas, forming hierarchical structured data. The parsed text data is segmented and stored according to semantic paragraphs for subsequent knowledge extraction. The system uses an IE big model to parse and extract knowledge from knowledge documents, identifying various types of knowledge units, including entities, relationships, states, actions, and text segments. For example, in an equipment operation and maintenance manual, the system can identify "fault code," "fault phenomenon," and "fault cause" as entities. When these entities belong to the same equipment, the system establishes relationships between entities centered around the equipment. Finally, the aforementioned entities and relationships are structured and stored as a knowledge graph database to represent the semantic and causal relationships between devices, components, and events.

[0035] Furthermore, to support dynamic execution and status monitoring of tasks, the system also constructs a state transition graph through a similar process: 6. Employee emergency handling manuals and standard handling procedures are also input into the knowledge extraction model for parsing, identifying state nodes and transition conditions during task execution to form a state transition graph. For example, in equipment maintenance tasks, the state path of "fault viewing status → manual confirmation status → generating maintenance work order → performing repair → status recovery" can be identified, and the triggering conditions between each state can be marked, thus reflecting the complete state change process of task execution.

[0036] By integrating static knowledge structures with dynamic task states in the above manner, super agents can achieve joint decision-making based on "knowledge-driven + state reasoning" in subsequent task planning, providing a unified knowledge foundation for autonomous collaboration among multiple agents.

[0037] like Figure 2As shown, the knowledge graph layer is used to store the static knowledge structure of devices and their operational logic. First, the original equipment manual, maintenance manual, and historical repair records are received via the file upload module. Then, the knowledge extraction model (IE large model) automatically parses the document content, performs OCR text recognition on the image portions, and analyzes the layout to identify the correspondence between titles, levels, tables, and pure images. The parsed text data is segmented and stored before entering the knowledge extraction stage. The model identifies and extracts elements such as entities, relationships, states, and actions, ultimately forming the knowledge graph structure. Entities in the knowledge graph include device nodes, component nodes, alarm item nodes, fault phenomenon nodes, fault cause nodes, and handling action nodes. Relationship types include component relationships, conditional relationships, action relationships, and causal relationships. For example, in the subgraph of the UPS equipment, the system uses the UPS host as the core node and establishes "component relationships" with rectifier modules, battery packs, inverter modules, etc.; it describes operating condition constraints through "conditional relationships," such as "triggering alarm 001 when low voltage and excessive temperature are detected"; it indicates the actions the system should take after an alarm through "action relationships," such as "UPS action → switching bypass power supply" or "UPS action → shutting down output"; and it connects fault phenomena and fault causes through "causal relationships," such as "excessive battery voltage difference → output voltage fluctuation → UPS action triggered." Each relationship edge is accompanied by semantic weight, source document, and update time tag, supporting subsequent dynamic updates and knowledge self-learning. When the system receives "alarm 001," the super agent first searches the knowledge graph for all related nodes and causal links related to the alarm item, thereby quickly determining the set of possible fault causes and priority handling suggestions.

[0038] The state transition layer is used to express the dynamic process of fault handling and task execution. Internally, it consists of a state transition graph composed of multiple state nodes and state transition edges. State nodes include, but are not limited to: "Alarm Status," "Located," "Analyzed," "Confirmed," "Work Order Created," "Maintenance Performed," "Retesting," and "Closed." Each state node is associated with corresponding triggering conditions, inputs / outputs, and fallback paths, reflecting the task execution lifecycle. When an alarm is triggered, the system automatically enters the state transition process: First, in the "Located" stage, a preliminary diagnosis is performed using the knowledge graph and real-time data; in the "Analyzed" stage, the RAG retrieval module is called to integrate historical similar cases and equipment data to confirm the cause; in the "Confirmed" stage, a specific repair plan is generated and the interface is called to query spare parts information; the system completes spare parts inventory retrieval through the WMS warehouse system interface, and then automatically creates a repair work order via API and pushes it to the work order system; after the repair task is executed, it enters the "Retesting" state. The system collects retest data and compares it with a threshold model. If it meets the qualification standard, the state transitions to "Closed"; otherwise, it falls back to the "Analyzed" stage for reassessment.

[0039] Example of execution process: When the system receives "Alarm 001" input, the super agent first retrieves the fault phenomenon node "rectifier module output voltage exceeded limit detected" and the potential cause "excessive battery voltage difference" associated with the alarm in the knowledge graph. Then, based on the state transition graph, it enters the "Located → Analyzed → Confirmed" stage, automatically generating a maintenance plan and creating a work order. After maintenance is completed, the system enters the "Retesting" state to perform automatic verification. If verification passes, it transitions to the "Closed" state; if verification fails, it reverts to the "Analyzed" stage for reassessment. This execution process simultaneously writes the processing path, verification data, and operation logs back to the knowledge graph, enabling continuous optimization and experience accumulation.

[0040] like Figure 3 As shown, after identifying entities, relationships, and attributes in documents such as equipment manuals and maintenance records during the knowledge extraction phase, a knowledge subgraph centered on the UPS equipment is generated. In this knowledge graph subgraph, "shutdown via bypass" is the core equipment operation node, around which multiple semantically related entities and corresponding causal relationship chains are established. The system uses "UPS action" as an action relationship node, connecting it to multiple alarm items (such as 004, 005, 015, 016, 022, and 023); each alarm item is further associated with its specific "fault cause" entity, such as "fault warning that the effective value of the connection voltage is lower than the set value," "detection that the output voltage of the autotransformer exceeds the operating range," and "bus voltage lower than the allowable value alarm." Through this three-layer semantic chain structure of "equipment action—alarm item—fault cause," the system realizes a traceable reasoning path from fault phenomenon to root cause. When a user inputs the natural language task "diagnose the cause of UPS shutdown", the super agent can quickly locate the relevant alarm nodes and their corresponding set of fault causes based on the semantic relationships in the knowledge graph, and infer the task's status (such as "alarm triggered" → "location in progress" → "analysis completed") by combining the state transition graph, and then automatically generate targeted handling strategies and task execution plans.

[0041] Furthermore, the knowledge association set obtained from the knowledge graph is semantically fused with the state paths identified in the state transition graph to construct the task's state dependency chain. Based on the state dependency chain and the logical relationships between subtasks, a preliminary task execution plan is generated, including execution order, resource dependencies, and condition constraints, including: Obtain the knowledge association set in the knowledge graph and the state path set in the state transition graph and perform semantic alignment. Based on the semantic similarity calculation results between the task objective and the context, perform semantic pairing between state nodes and knowledge nodes to form a state-knowledge mapping relationship. Based on the mapping relationship, establish the sequential dependencies between each action node and state node in the task, and construct the state dependency chain of the task. Based on the logical order and constraints of nodes in the state dependency chain, parallelizable subtasks and dependent subtasks in the task are identified, and the execution topology diagram of the task is generated. Based on the topology diagram, combined with resource dependencies, execution constraints and task priorities, the execution order, parallel relationship and triggering conditions of each subtask are determined, a preliminary task execution plan is generated and stored in the group memory space.

[0042] Furthermore, each expert agent receives task assignment instructions and executes the corresponding subtasks. During execution, the execution status, stage results, and intermediate data are written to the group memory space in real time, including: In response to the task allocation instructions, each expert agent parses the corresponding subtask content and execution parameters; Based on the execution order, resource dependencies, and condition constraints in the preliminary task execution plan, the corresponding sub-tasks are executed. During the task execution process, the execution status is monitored in real time, and the execution status, stage results, and intermediate data are written into the group memory space in a structured form, along with task identifiers, timestamps, and execution context information, for the super agent to perform global task status perception and progress tracking.

[0043] Furthermore, the super agent monitors the execution progress of the expert agent based on real-time state data from the group memory space. When anomalies, failures, or changes in the external environment are detected, the super agent dynamically updates the task execution plan and reallocates unfinished sub-tasks, including: Periodically read the task status data in the group memory space to obtain the execution progress, resource utilization and stage results of each expert agent; When an execution exception, task timeout, or change in external environment parameters is detected, a new task planning request is generated based on context information and dependencies. Based on the replanning request, adjust the state dependency chain and task priority in the task execution plan, and reallocate incomplete or failed subtasks to available expert agents.

[0044] Specifically, the fault handling process in this embodiment of the invention includes the following main steps: When the system detects a device fault, the super agent first triggers a fault analysis task. Historical maintenance data is read from the group memory space to identify the fault type and analyze its causes, generating fault repair information. Subsequently, the system automatically performs a spare parts inventory verification operation. If sufficient spare parts are detected, a repair work order is created, and material retrieval instructions and operation prompts are generated. During the repair execution phase, the expert agent sequentially executes disassembly suggestions, installation suggestions, and live-line testing. After repair, the system initiates a fault recovery detection process to automatically detect whether the fault point status has returned to normal: if the detection result is normal, the system records the manual confirmation result and closes the repair work order, ending the task execution; if the detection result is abnormal, the system automatically regenerates the work order and triggers a new round of detection until verification is passed. Through the above-mentioned cyclic detection and closed-loop verification mechanism, a task flow structure of "fault detection → solution generation → work order execution → verification write-back" is formed, realizing fully automated management and dynamic self-repair throughout the process. (The textual description of this process is missing.) Figure 4 As shown.

[0045] Specifically, the super agent is responsible for understanding user needs and retrieving task-related knowledge from the knowledge graph. For example, if a device is in a faulty state, fault attribution analysis is needed before fault suggestions can be provided, thus enabling task breakdown and route planning. The super agent searches the current agent group for the expert agent to be assigned to each task and sends task instructions. Each expert agent performs the task sequentially, and after completion, feeds back the task results to the super agent. If an expert agent fails, the super agent detects this and replans, assigning a new task to continue the attempt until the task can proceed normally. Finally, after each expert agent completes its sub-task, the processing information is stored in the group's memory space. The super agent then performs a unified summary, as follows: Figure 5 To provide a visual presentation.

[0046] Second Embodiment Based on the same inventive concept, this invention provides an intelligent agent autonomous cooperation system based on group hierarchical cooperation, employing the intelligent agent autonomous cooperation method described above, including: The response module is used to respond to the natural language task requests input by the user, parse the task semantics to extract the task objectives, constraints and contextual semantic information, and form structured task description data; The data processing module is used to retrieve task-related knowledge nodes and execution state paths from the knowledge graph and state transition graph based on task description data, generate a preliminary task execution plan including execution order and dependencies, select expert agents that match the subtasks in the task execution plan from the group, and generate task allocation instructions for task collaboration; each expert agent receives the task allocation instructions and executes the corresponding subtask, and writes the execution status, stage results and intermediate data into the group memory space in real time during the execution process; The results output module is used by the super agent to monitor the execution progress of the expert agent based on real-time state data in the group memory space. When an anomaly, failure, or change in the external environment is detected, the task execution plan is dynamically updated and unfinished sub-tasks are reallocated. After all sub-tasks are completed, the super agent performs semantic integration and summarization of the task results in the group memory space to generate a unified task output result.

[0047] Furthermore, the data processing module includes: The knowledge retrieval unit is used by the super intelligent agent to retrieve entity nodes, relation edges, and semantic tags related to the task from the knowledge graph based on the task objectives and key operation objects, so as to determine the executable set of knowledge associations. The knowledge graph is constructed based on a semantic network structure and is a multi-dimensional knowledge graph formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model. It is used to represent the semantic associations and causal relationships between key operation objects. The state recognition unit is used to identify the initial state, target state, and execution state path related to the task objective in the state transition graph based on constraints and contextual semantic information. The state transition graph is constructed based on task execution records and operation procedures. It extracts each state node and its corresponding transition conditions during task execution through a knowledge extraction model, establishes the transition relationship between state nodes, and describes the state change path of the task at different execution stages. The task planning unit is used to semantically fuse the set of knowledge associations obtained from the knowledge graph with the state paths identified in the state transition graph to construct the state dependency chain of the task; based on the state dependency chain and the logical associations between each subtask, it generates a preliminary task execution plan that includes execution order, resource dependencies and condition constraints.

[0048] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A method for autonomous agent cooperation based on hierarchical group collaboration, characterized in that, Includes the following steps: In response to user input of natural language task requests, the task semantics are parsed to extract the task objectives, constraints and contextual semantic information, forming structured task description data; Based on the task description data, the super agent retrieves knowledge nodes and execution state paths related to the task from the knowledge graph and state transition graph, generates a preliminary task execution plan containing execution order and dependencies, selects expert agents that match the subtasks in the task execution plan from the group, and generates task allocation instructions for task collaboration. Each of the expert intelligent agents receives the task allocation instruction and executes the corresponding sub-task. During the execution process, the execution status, stage results and intermediate data are written into the group memory space in real time. The super agent monitors the execution progress of the expert agent based on the real-time status data of the group memory space. When an anomaly, failure, or change in the external environment is detected, the super agent dynamically updates the task execution plan and reallocates the unfinished sub-tasks. After all the sub-tasks are completed, the super agent performs semantic integration and summarization of the task results in the group memory space to generate a unified task output result.

2. The intelligent agent autonomous cooperation method according to claim 1, characterized in that, In response to user-input natural language task requests, the system parses task semantics to extract task objectives, constraints, and contextual semantic information, forming structured task description data, including: In response to the natural language task request input by the user, the task text is obtained and preprocessed, including word segmentation, part-of-speech tagging and sentence normalization; Based on the semantic parsing model, the preprocessed task text and intent recognition and semantic role labeling are used to extract the task objective, the constraints and key operation objects; The extracted task objectives, constraints, and key operation objects are semantically associated and matched with the context information of historical tasks in the group memory space to form a context-enhanced representation of the task. Based on the context-enhanced representation, structured task description data is generated and mapped to a standard task data format for the super agent to perform knowledge retrieval and task planning.

3. The intelligent agent autonomous cooperation method according to claim 2, characterized in that, Based on the task description data, the super agent retrieves knowledge nodes and execution state paths related to the task from the knowledge graph and state transition graph, and generates a preliminary task execution plan containing execution order and dependencies, including: The super intelligent agent, based on the task objective and the key operation objects, retrieves entity nodes, relation edges, and semantic tags related to the task from the knowledge graph to determine the executable knowledge association set; wherein, the knowledge graph is constructed based on a semantic network structure, and is a multi-dimensional knowledge graph formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model, and is used to represent the semantic associations and causal relationships between the key operation objects; Based on the constraints and the contextual semantic information, the initial state, target state, and execution state path related to the task objective are identified in the state transition graph; wherein, the state transition graph is constructed based on task execution records and operation procedures, and each state node and corresponding transition condition during the task execution process is extracted through the knowledge extraction model, and the transition relationship between the state nodes is established to describe the state change path of the task at different execution stages. The knowledge association set obtained from the knowledge graph is semantically fused with the state path identified in the state transition graph to construct a state dependency chain for the task; based on the state dependency chain and the logical associations between each subtask, a preliminary task execution plan is generated that includes the execution order, resource dependencies, and condition constraints.

4. The intelligent agent autonomous cooperation method according to claim 3, characterized in that, Based on the construction of a semantic network structure, a multi-dimensional knowledge graph is formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model, including: The response knowledge platform obtains multi-source knowledge files through its file upload interface, including technical manuals, operation and maintenance procedures, and log records. The document parsing tool performs structured parsing of the multi-source knowledge files, performs OCR text recognition and layout analysis to identify the correspondence between title levels, body paragraphs and pure image areas, forms hierarchical text structure data and stores it according to segmented text, which is used as the input set for the knowledge extraction model. Based on the knowledge extraction model, semantic parsing and knowledge recognition are performed on the segmented text to extract the corresponding entity, relation, and attribute information, and to establish semantic triples of entity-relation-entity; wherein, the relation represents the operation behavior in the task execution and corresponds to the action node of the task; The semantic triples are transformed into a relational structure between the multi-source knowledge nodes, forming a multi-dimensional knowledge graph in the form of a semantic network, which serves as the knowledge graph.

5. The intelligent agent autonomous cooperation method according to claim 3, characterized in that, Based on task execution records and operating procedures, the knowledge extraction model is used to extract each state node and its corresponding transition conditions during task execution, and to establish the transition relationships between the state nodes, including: The knowledge platform receives the task execution record and standard processing flow document, which includes an emergency manual, operating procedures and equipment maintenance records. The task execution record and standard processing flow document are preprocessed, and the OCR recognition and text structure analysis are performed to extract semantic fragments containing operation steps, event triggering conditions and result descriptions. Based on the knowledge extraction model, the semantic fragment is identified and classified to determine the key event nodes and their dependencies in the task execution process. The state recognition model is used to identify the state nodes and their corresponding transition conditions for each event node, and a state-event-state transition triplet is constructed. The transition triples are transformed into a multidimensional directed graph structure between the state nodes, forming a set of state transition paths as the state transition graph.

6. The intelligent agent autonomous cooperation method according to claim 4, characterized in that, The knowledge association set obtained from the knowledge graph is semantically fused with the state paths identified in the state transition graph to construct a task state dependency chain; based on the state dependency chain and the logical associations between each subtask, a preliminary task execution plan is generated, including the execution order, resource dependencies, and condition constraints, including: Obtain the knowledge association set in the knowledge graph and the state path set in the state transition graph and perform semantic alignment. Based on the semantic similarity calculation result between the task objective and the context, semantically pair the state node with the knowledge node to form a state-knowledge mapping relationship. Based on the mapping relationship, establish the sequential dependency relationship between each action node and the state node in the task, and construct the state dependency chain of the task; Based on the logical order of nodes in the state dependency chain and the constraints, parallelizable subtasks and dependent subtasks in the task are identified, and an execution topology diagram of the task is generated. Based on the topology diagram, combined with resource dependencies, execution constraints and task priorities, the execution order, parallel relationship and triggering conditions of each subtask are determined, and the preliminary task execution plan is generated and stored in the group memory space.

7. The intelligent agent autonomous cooperation method according to claim 5 or 6, characterized in that, Each of the expert agents receives the task allocation instruction and executes the corresponding subtask. During execution, the execution status, stage results, and intermediate data are written to the group memory space in real time, including: In response to the task allocation instruction, each of the expert agents parses the content and execution parameters of the corresponding sub-task; Based on the execution order, resource dependencies, and condition constraints in the preliminary task execution plan, the corresponding sub-tasks are executed. During task execution, the execution status is monitored in real time, and the execution status, stage results, and intermediate data are written into the group memory space in a structured form, with the addition of task identifiers, timestamps, and execution context information, for the super agent to perform global task status perception and progress tracking.

8. The method for autonomous cooperation of intelligent agents according to claim 7, characterized in that, The super agent monitors the execution progress of the expert agent based on real-time state data from the group memory space. When an anomaly, failure, or change in the external environment is detected, the super agent dynamically updates the task execution plan and reallocates the unfinished sub-tasks, including: The task status data in the group memory space is periodically read to obtain the execution progress, resource utilization rate and stage results of each expert agent; When an execution anomaly, task timeout, or change in external environment parameters is detected, a new task replanning request is generated based on the context information and the preceding and following dependencies. Based on the replanning request, the state dependency chain and task priority in the task execution plan are adjusted, and the incomplete or failed subtasks are reallocated to the available expert agents.

9. An agent-based autonomous cooperation system based on hierarchical group cooperation, employing the agent-based autonomous cooperation method as described in any one of claims 1 to 8, characterized in that, include: The response module is used to respond to the natural language task requests input by the user, parse the task semantics to extract the task objectives, constraints and contextual semantic information, and form structured task description data; The data processing module is used to retrieve knowledge nodes and execution state paths related to the task from the knowledge graph and state transition graph based on the task description data, generate a preliminary task execution plan containing execution order and dependencies, select expert agents that match the subtasks in the task execution plan from the group, and generate task allocation instructions for task collaboration. Each of the expert intelligent agents receives the task allocation instruction and executes the corresponding sub-task. During the execution process, the execution status, stage results and intermediate data are written into the group memory space in real time. The result output module is used by the super agent to monitor the execution progress of the expert agent based on the real-time status data of the group memory space. When an anomaly, failure, or change in the external environment is detected, the super agent dynamically updates the task execution plan and reallocates the unfinished sub-tasks. After all the sub-tasks are executed, the super agent performs semantic integration and summarization of the task results in the group memory space to generate a unified task output result.

10. The intelligent agent autonomous cooperation system according to claim 9, characterized in that, The data processing module includes: The knowledge retrieval unit is used by the super agent to retrieve entity nodes, relation edges, and semantic tags related to the task from the knowledge graph based on the task objective and the key operation objects, so as to determine the executable knowledge association set; wherein, the knowledge graph is constructed based on a semantic network structure, and is a multi-dimensional knowledge graph formed by identifying entities, relations, and attributes in multi-source knowledge documents through a knowledge extraction model, and is used to represent the semantic associations and causal relationships between the key operation objects; A state recognition unit is used to identify the initial state, target state, and execution state path related to the task objective in the state transition graph based on the constraints and the contextual semantic information; wherein, the state transition graph is constructed based on task execution records and operation procedures, and each state node and corresponding transition condition during the task execution process is extracted through the knowledge extraction model to establish the transition relationship between the state nodes, which is used to describe the state change path of the task at different execution stages; The task planning unit is used to semantically fuse the knowledge association set obtained from the knowledge graph with the state path identified in the state transition graph to construct a state dependency chain for the task; and to generate a preliminary task execution plan containing the execution order, resource dependencies and condition constraints based on the state dependency chain and the logical associations between the subtasks.

Images