A large model-based enterprise multi-agent task autonomous processing method and system

Abstract

This invention relates to the field of intelligent task processing technology, and more particularly to a method and system for autonomous multi-agent task processing in enterprises based on a large model. The method receives business instructions from an enterprise management system, generates a structured task tree with business domain labels, hierarchical structure, and dependencies using a natural language target parsing algorithm, maps sub-task nodes to candidate enterprise agents to obtain a first allocation result, and performs a second dynamic reallocation based on the load difference between the task load index and a preset load threshold to obtain a second allocation result. Each candidate enterprise agent asynchronously executes sub-task nodes according to dependency edges, achieving state persistence across steps, agents, and days. An exception self-healing algorithm automatically handles execution anomalies, and a compliance audit algorithm provides full-process traceability and access control. This invention realizes natural language-driven, load-balanced autonomous execution and full-process compliance auditing for complex enterprise business processes.

Description

Technical Field

[0001] This invention relates to the field of intelligent task processing technology, and in particular to an enterprise multi-agent task autonomous processing method and system based on a large model. Background Technology

[0002] In recent years, agent technology based on large language models has been widely applied in the fields of natural language understanding and task automation, achieving a certain degree of process automation through task decomposition and multi-model collaboration. However, existing technologies still have shortcomings in enterprise-level multi-agent autonomous execution scenarios: existing task allocation mechanisms mostly adopt static business domain mapping or fixed model matching, lacking the ability to dynamically reallocate tasks based on the real-time load status of agents, resulting in unbalanced loads and low resource utilization among multiple agents.

[0003] Meanwhile, existing systems lack mechanisms for persistently inheriting task states across steps, agents, and days, as well as automatic identification and self-healing capabilities for execution anomalies, making it difficult to guarantee the stable and continuous execution of complex, long-cycle tasks for enterprises. Furthermore, existing solutions fail to build fine-grained access control and end-to-end compliance audit mechanisms for enterprise management scenarios, thus failing to meet the requirements of enterprise businesses for security, traceability, and unattended operation.

[0004] Therefore, there is an urgent need for a multi-agent task autonomous processing method that is geared towards enterprise scenarios and has dynamic load balancing and full-process autonomous capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an enterprise multi-agent task autonomous processing method and system based on a large model, thereby solving the aforementioned problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides an enterprise multi-agent task autonomous processing method based on a large model, comprising the following steps: S1. Receive business instructions from the enterprise management system, perform semantic parsing and business domain identification on the business instructions through a natural language target parsing algorithm, and generate a structured task tree with business domain labels, hierarchical structure and dependency relationships. The structured task tree contains multiple sub-task nodes and dependency edges that indicate the execution order between nodes. S2. Based on the business domain labels, map each subtask node to the corresponding candidate enterprise intelligent agent set to obtain the first allocation result; S3. Based on the first allocation result, calculate the task load index of each candidate enterprise agent, perform a difference operation between the task load index and the preset load threshold to obtain the load difference value, and dynamically reallocate the first allocation result according to the load difference value to obtain the second allocation result. S4. Based on the second allocation result, each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the execution order represented by the dependency edge. During the execution process, the context inheritance algorithm is used to persist the task state and call the context, and the exception self-healing algorithm is used to automatically identify, retry, roll back and report tasks with execution exceptions. S5. Adopt a compliance audit algorithm to record and control access to the entire process, integrate the execution results of each sub-task node, and generate processing strategies for business instructions in the enterprise management system.

[0007] Preferably, in S1, the semantic parsing and business domain identification of business instructions are performed using a natural language target parsing algorithm to generate a structured task tree with business domain labels, hierarchical structure, and dependency relationships, including: The system performs word segmentation and semantic parsing on business instructions in the enterprise management system, extracts business objectives, constraints and key business entities, and identifies and generates business domain tags based on the domain to which the key business entities belong. A two-level splitting strategy is adopted to decompose business objectives into phased main tasks to form a task skeleton. Each phased main task is then further broken down into multiple executable sub-task nodes. Based on the logical sequence and data dependency relationships between each subtask node, dependency edges representing the execution order between nodes are constructed, and business domain labels, subtask nodes, and dependency edges are combined to generate a structured task tree.

[0008] Preferably, in S2, the candidate enterprise intelligent agent set includes financial intelligent agent, human resources intelligent agent, supply chain intelligent agent, compliance intelligent agent and operational intelligent agent; The financial intelligent agent is used to execute financial domain sub-task nodes, which include financial reconciliation, voucher generation, report output, and tax declaration. Human intelligence agents are used to execute human resources domain sub-task nodes, which include personnel entry, transfer, and departure, payroll calculation, and attendance approval. The supply chain intelligent agent is used to execute sub-task nodes in the supply chain domain, which include procurement fulfillment, inventory management, and logistics tracking. The compliance agent is used to execute compliance domain sub-task nodes, which include compliance review, risk identification, and audit recording. The operational intelligence agent is used to execute sub-task nodes in the operational domain, which include operational monitoring, data aggregation, and process supervision.

[0009] Preferably, in S2, based on the business domain label, each subtask node is mapped to the corresponding candidate enterprise agent set to obtain the first allocation result, including: Traverse each subtask node in the structured task tree, extract the business domain label of each subtask node, and compare and match the business domain label with the preset business domain identifier of each candidate enterprise intelligent agent in the candidate enterprise intelligent agent set. When the business domain label of a subtask node matches the preset business domain identifier of a candidate enterprise intelligent agent, the subtask node is mapped to the corresponding candidate enterprise intelligent agent, and the initial mapping relationship between the subtask node and the candidate enterprise intelligent agent is established. The initial mapping relationships between all subtask nodes and candidate enterprise agents are integrated to obtain the first allocation result.

[0010] Preferably, in S3, based on the first allocation result, the task load index of each candidate enterprise agent is calculated, and the difference between the task load index and a preset load threshold is calculated to obtain the load difference value, including: Based on the first allocation result, the number of subtask nodes mapped to each candidate enterprise agent and the preset complexity weight of each subtask node are extracted. The task load index of each candidate enterprise agent is calculated by weighting the number of subtask nodes with the preset complexity weight. ; in, Let be the task load index of the i-th candidate enterprise intelligent agent. This represents the number of subtask nodes mapped to the i-th candidate enterprise agent. This is the set of subtask nodes mapped to the i-th candidate enterprise agent. α and β are the preset complexity weights for the j-th subtask node, where α and β are both set weighting coefficients; The difference between the obtained task load index and the preset load threshold is calculated to obtain the load difference value, which provides a basis for dynamic redistribution.

[0011] Preferably, in S3, the second allocation result is obtained by dynamically reallocating the first allocation result based on the load difference value, including: Candidate enterprise agents with load difference values ​​greater than zero are identified as overloaded, and subtask nodes to be migrated are extracted from their mapped subtask nodes. Match the subtask nodes to be migrated to other candidate enterprise intelligent agents with load difference values ​​less than zero and the same business domain processing capabilities according to the business domain label, and re-establish the mapping relationship between subtask nodes and candidate enterprise intelligent agents. The mapping relationship between all subtask nodes after reallocation and candidate enterprise agents is integrated to obtain the second allocation result.

[0012] Preferably, in S4, based on the second allocation result, each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the execution order represented by the dependency edges, including: Based on the second allocation result, the sub-task nodes mapped by each candidate enterprise agent are extracted. The execution order of each sub-task node is determined according to the execution order represented by the dependency edges in the structured task tree. Sub-task nodes without dependencies are determined as nodes that can be executed in parallel, and sub-task nodes with dependencies are determined as nodes that can be executed sequentially. Each candidate enterprise agent executes its mapped sub-task nodes asynchronously according to the execution sequence. During execution, a context inheritance algorithm is used to persist task states and call contexts, ensuring that the output of the preceding sub-task node serves as the input for the subsequent sub-task node.

[0013] Preferably, an enterprise multi-agent task autonomous processing system based on a large model includes: The instruction receiving module is used to receive business instructions from the enterprise management system. It performs semantic parsing and business domain identification on the business instructions through a natural language target parsing algorithm, and generates a structured task tree with business domain labels, hierarchical structure and dependency relationships. The structured task tree contains multiple sub-task nodes and dependency edges that indicate the execution order between nodes. The first allocation module is used to map each subtask node to the corresponding candidate enterprise intelligent agent set according to the business domain label, so as to obtain the first allocation result; The second allocation module is used to calculate the task load index of each candidate enterprise agent based on the first allocation result, perform a difference operation between the task load index and a preset load threshold to obtain a load difference value, and dynamically reallocate the first allocation result according to the load difference value to obtain the second allocation result. The task execution module is used to asynchronously execute the mapped sub-task nodes according to the execution order represented by the dependency edges, based on the second allocation result. During the execution, the context inheritance algorithm is used to persist the task state and call the context, and the exception self-healing algorithm is used to automatically identify, retry, roll back and report tasks with execution exceptions. The strategy generation module is used to use compliance auditing algorithms to record and control permissions for the entire process, integrate the execution results of each sub-task node, and generate processing strategies for business instructions in the enterprise management system.

[0014] The advantages and beneficial effects of this invention compared to the prior art are: 1. This invention uses a natural language target parsing algorithm to perform semantic parsing and business domain identification on business instructions in an enterprise management system. It automatically generates a structured task tree with business domain labels, hierarchical structure, and dependency relationships, realizing automated understanding and task decomposition of business instructions without the need for manual pre-setting of processes, thus lowering the threshold for use.

[0015] 2. This invention combines primary task allocation with secondary task allocation based on load difference values ​​to achieve dynamic load balancing of subtask nodes among candidate enterprise agents, avoiding single-point overload and improving the efficiency of multi-agent collaborative execution and system resource utilization.

[0016] 3. This invention employs a context inheritance algorithm to persist task states and invoke context across steps, agents, and days, ensuring the continuous execution capability of long-cycle complex tasks for enterprises and supporting breakpoint resume.

[0017] 4. This invention employs an anomaly self-healing algorithm to automatically identify, retry, roll back, and report execution anomalies, and uses a compliance auditing algorithm to record and control access permissions for the entire process, significantly improving the system's robustness, security, and compliance auditing capabilities.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart of an enterprise multi-agent task autonomous processing method based on a large model, as described in an embodiment of the present invention. Figure 2 This is a structural diagram of an enterprise multi-agent task autonomous processing system based on a large model, according to an embodiment of the present invention. Detailed Implementation

[0020] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this invention provides an enterprise multi-agent task autonomous processing method based on a large model, comprising the following steps: S1. Receive business instructions from the enterprise management system, perform semantic parsing and business domain identification on the business instructions through a natural language target parsing algorithm, and generate a structured task tree with business domain labels, hierarchical structure and dependency relationships. The structured task tree contains multiple sub-task nodes and dependency edges that indicate the execution order between nodes. S2. Based on the business domain labels, map each subtask node to the corresponding candidate enterprise intelligent agent set to obtain the first allocation result; S3. Based on the first allocation result, calculate the task load index of each candidate enterprise agent, perform a difference operation between the task load index and the preset load threshold to obtain the load difference value, and dynamically reallocate the first allocation result according to the load difference value to obtain the second allocation result. S4. Based on the second allocation result, each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the execution order represented by the dependency edge. During the execution process, the context inheritance algorithm is used to persist the task state and call the context, and the exception self-healing algorithm is used to automatically identify, retry, roll back and report tasks with execution exceptions. S5. Adopt a compliance audit algorithm to record and control access to the entire process, integrate the execution results of each sub-task node, and generate processing strategies for business instructions in the enterprise management system.

[0023] Specifically, in actual deployment, enterprise users input business instructions into the enterprise management system via a management terminal. These instructions are business objectives described by the user in everyday language, such as "complete this month's financial closing and generate reports," which differs from traditional input methods that require selecting fixed menus or filling out structured forms. The natural language target parsing algorithm, based on a large language model, performs intent recognition and entity extraction on the instructions, transforming unstructured natural language into a machine-understandable structured task tree. Business domain labels are used to identify the professional field to which each subtask belongs, such as finance, human resources, or supply chain. The structured task tree organizes tasks in a tree-like hierarchy, with the root node corresponding to the overall business objective and leaf nodes being subtask nodes, i.e., the smallest indivisible execution unit. Dependency edges are used to represent the logical sequence constraints and data flow relationships between subtask nodes; for example, the "voucher review" node depends on the "voucher generation" node.

[0024] Candidate enterprise agents are AI proxy programs that encapsulate a specific domain knowledge base and system operation interface. The first allocation result is a preliminary draft of the correspondence between sub-task nodes and candidate enterprise agents based on business domain labels. The task load index is used to quantify the current workload saturation of each candidate enterprise agent. The load difference value is the deviation between this saturation level and the system's preset processing capacity limit; when the load difference value of an agent is positive, it indicates that it is in an overloaded state. The second allocation result is the final mapping scheme after load balancing adjustment. The context inheritance algorithm is used to save and transmit intermediate task states and data across steps, agents, and days during execution, ensuring that long-cycle tasks can be recovered from breakpoints after interruption; the anomaly self-healing algorithm is used to monitor errors and timeouts during execution and automatically trigger retries, rollbacks, or manual reporting; the compliance audit algorithm is used to record operation logs, verify user permissions, and generate audit trail records. The final generated processing strategy is a complete execution result or decision scheme oriented towards the original business instructions.

[0025] Preferably, in S1, the semantic parsing and business domain identification of business instructions are performed using a natural language target parsing algorithm to generate a structured task tree with business domain labels, hierarchical structure, and dependency relationships, including: The system performs word segmentation and semantic parsing on business instructions in the enterprise management system, extracts business objectives, constraints and key business entities, and identifies and generates business domain tags based on the domain to which the key business entities belong. A two-level splitting strategy is adopted to decompose business objectives into phased main tasks to form a task skeleton. Each phased main task is then further broken down into multiple executable sub-task nodes. Based on the logical sequence and data dependency relationships between each subtask node, dependency edges representing the execution order between nodes are constructed, and business domain labels, subtask nodes, and dependency edges are combined to generate a structured task tree.

[0026] Specifically, the natural language target parsing algorithm first performs word segmentation and semantic analysis on the business instructions of the enterprise management system, extracting business objectives, constraints, and key business entities. Key business entities refer to the specific business objects involved in the instructions, such as personnel names, document types, accounting subjects, or time periods. Based on the professional field to which the key business entities belong, the system automatically generates business domain tags.

[0027] It's important to understand that the two-level decomposition strategy refers to a two-stage approach to task decomposition, starting with a coarse-grained approach and gradually refining towards a fine-grained one. The first stage breaks down the overall business objective into phased main tasks, forming a task skeleton—a framework composed of several macro-level phases, such as "data preparation—business processing—result output." The second stage further refines each phased main task, generating specific sub-task nodes that can be directly executed. For example, "data preparation" is broken down into "retrieving bank statements" and "matching invoice information." Subsequently, the system determines the logical sequence between sub-task nodes based on business rules and identifies data dependencies based on data flow requirements—that is, whether a subsequent node requires the output of a preceding node as an input prerequisite. Based on these relationships, dependency edges are constructed, ultimately combining business domain labels, sub-task nodes, and dependency edges into a complete structured task tree.

[0028] Preferably, in S2, the candidate enterprise intelligent agent set includes financial intelligent agent, human resources intelligent agent, supply chain intelligent agent, compliance intelligent agent and operational intelligent agent; The financial intelligent agent is used to execute financial domain sub-task nodes, which include financial reconciliation, voucher generation, report output, and tax declaration. Human intelligence agents are used to execute human resources domain sub-task nodes, which include personnel entry, transfer, and departure, payroll calculation, and attendance approval. The supply chain intelligent agent is used to execute sub-task nodes in the supply chain domain, which include procurement fulfillment, inventory management, and logistics tracking. The compliance agent is used to execute compliance domain sub-task nodes, which include compliance review, risk identification, and audit recording. The operational intelligence agent is used to execute sub-task nodes in the operational domain, which include operational monitoring, data aggregation, and process supervision.

[0029] Specifically, the candidate enterprise intelligent agent set consists of five types of specialized intelligent agents, each pre-configured with corresponding domain business knowledge and system interfaces. The financial intelligent agent is dedicated to handling sub-tasks in the financial domain, including financial reconciliation, voucher generation, report output, and tax filing. It completes fund and accounting operations by connecting to the enterprise ERP or financial system. The human resources intelligent agent is dedicated to handling sub-tasks in the human resources domain, including employee onboarding, transfers, and departures, payroll calculation, and attendance approval. It completes employee lifecycle and payroll management by connecting to the HR system.

[0030] The supply chain intelligence agent is dedicated to handling sub-tasks within the supply chain domain, including procurement fulfillment, inventory management, and logistics tracking. It monitors the flow of materials by connecting to procurement and warehousing systems. The compliance intelligence agent is dedicated to handling sub-tasks within the compliance domain, including compliance review, risk identification, and audit documentation. It verifies compliance by connecting to risk control rule bases and audit systems. The operations intelligence agent is dedicated to handling sub-tasks within the operations domain, including operations monitoring, data aggregation, and process supervision. It achieves business situation awareness and process follow-up by connecting to BI and operations management systems.

[0031] Preferably, in S2, based on the business domain label, each subtask node is mapped to the corresponding candidate enterprise agent set to obtain the first allocation result, including: Traverse each subtask node in the structured task tree, extract the business domain label of each subtask node, and compare and match the business domain label with the preset business domain identifier of each candidate enterprise intelligent agent in the candidate enterprise intelligent agent set. When the business domain label of a subtask node matches the preset business domain identifier of a candidate enterprise intelligent agent, the subtask node is mapped to the corresponding candidate enterprise intelligent agent, and the initial mapping relationship between the subtask node and the candidate enterprise intelligent agent is established. The initial mapping relationships between all subtask nodes and candidate enterprise agents are integrated to obtain the first allocation result.

[0032] Specifically, the generation process of the first allocation result is as follows: The system traverses all sub-task nodes in the structured task tree, extracting the business domain label carried by each sub-task node one by one; candidate enterprise agents are pre-configured with preset business domain identifiers during registration, used to declare their areas of expertise. The system compares and matches the business domain labels of the sub-task nodes with the preset business domain identifiers of the candidate enterprise agents. When they match, an initial mapping relationship is established between the sub-task node and the corresponding candidate enterprise agent. This relationship is a preliminary correspondence table that only considers domain matching and does not consider load status. After the traversal is completed, the system integrates all initial mapping relationships and summarizes them to form the first allocation result.

[0033] Preferably, in S3, based on the first allocation result, the task load index of each candidate enterprise agent is calculated, and the difference between the task load index and a preset load threshold is calculated to obtain the load difference value, including: Based on the first allocation result, the number of subtask nodes mapped to each candidate enterprise agent and the preset complexity weight of each subtask node are extracted. The task load index of each candidate enterprise agent is calculated by weighting the number of subtask nodes with the preset complexity weight. ; in, Let be the task load index of the i-th candidate enterprise intelligent agent. This represents the number of subtask nodes mapped to the i-th candidate enterprise agent. This is the set of subtask nodes mapped to the i-th candidate enterprise agent. α and β are the preset complexity weights for the j-th subtask node, where α and β are both set weighting coefficients; The difference between the obtained task load index and the preset load threshold is calculated to obtain the load difference value, which provides a basis for dynamic redistribution.

[0034] Specifically, the task load index is used to comprehensively reflect the saturation level of the work currently undertaken by each candidate enterprise agent. Based on the first allocation result, the system counts the number of sub-task nodes mapped to each candidate enterprise agent and retrieves the preset complexity weights of each sub-task node. These weights are pre-configured by the system administrator based on business experience and are used to distinguish the execution difficulty of different tasks. For example, the complexity weight of cross-system data verification tasks is higher than that of single-table query tasks.

[0035] The system comprehensively considers the number of quantum task nodes and preset complexity weights to calculate the task load index of each candidate enterprise agent. Subsequently, the system performs a difference calculation between the obtained task load index and a pre-set load threshold to obtain a load difference value. This load difference value is used to measure the degree of surplus or overload of each agent's current load relative to its designed processing capacity, providing a quantitative basis for subsequent task migration.

[0036] Preferably, in S3, the second allocation result is obtained by dynamically reallocating the first allocation result based on the load difference value, including: Candidate enterprise agents with load difference values ​​greater than zero are identified as overloaded, and subtask nodes to be migrated are extracted from their mapped subtask nodes. Match the subtask nodes to be migrated to other candidate enterprise intelligent agents with load difference values ​​less than zero and the same business domain processing capabilities according to the business domain label, and re-establish the mapping relationship between subtask nodes and candidate enterprise intelligent agents. The mapping relationship between all subtask nodes after reallocation and candidate enterprise agents is integrated to obtain the second allocation result.

[0037] Specifically, the second allocation result is obtained through dynamic reallocation: The system first determines the load difference value of each candidate enterprise agent. Candidate enterprise agents with a load difference value greater than zero are classified as overloaded, meaning their actual load has exceeded the preset processing capacity limit. For candidate enterprise agents in an overloaded state, the system selects sub-task nodes to be migrated from their currently mapped sub-task nodes, prioritizing tasks that are not tightly coupled and can be executed independently. Subsequently, the system matches the sub-task nodes to be migrated to other candidate enterprise agents with a load difference value less than zero and the same business domain processing capability, based on their business domain labels. The same business domain processing capability means that the receiving agent must possess the knowledge and interface permissions to handle this type of business, ensuring that the task can still be executed correctly after migration. After re-establishing the mapping relationship, the system integrates the final correspondence between all sub-task nodes and candidate enterprise agents to obtain the second allocation result.

[0038] Preferably, in S4, based on the second allocation result, each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the execution order represented by the dependency edges, including: Based on the second allocation result, the sub-task nodes mapped by each candidate enterprise agent are extracted. The execution order of each sub-task node is determined according to the execution order represented by the dependency edges in the structured task tree. Sub-task nodes without dependencies are determined as nodes that can be executed in parallel, and sub-task nodes with dependencies are determined as nodes that can be executed sequentially. Each candidate enterprise agent executes its mapped sub-task nodes asynchronously according to the execution sequence. During execution, a context inheritance algorithm is used to persist task states and call contexts, ensuring that the output of the preceding sub-task node serves as the input for the subsequent sub-task node.

[0039] Specifically, the asynchronous execution process is as follows: Based on the second allocation result, the system extracts all sub-task nodes mapped to each candidate enterprise agent and determines the execution sequence of each sub-task node, i.e., the actual starting order of each task, according to the logical constraints represented by the dependency edges in the structured task tree. For sub-task nodes without dependency edges, the system determines them as parallel execution nodes, allowing the corresponding candidate enterprise agents to start processing simultaneously; for sub-task nodes with dependency edges, the system determines them as serial execution nodes, requiring them to wait for the completion of the preceding node before starting. Each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the determined execution sequence. During execution, a context inheritance algorithm is used for task state persistence and context invocation. Task state persistence refers to writing intermediate execution data and state information to the storage medium to ensure recovery after a day or interruption; context invocation refers to the subsequent sub-task node automatically reading the output results of the preceding sub-task node as its own input, achieving seamless data flow.

[0040] Preferred, such as Figure 2 As shown, an enterprise multi-agent task autonomous processing system based on a large model includes: The instruction receiving module is used to receive business instructions from the enterprise management system. It performs semantic parsing and business domain identification on the business instructions through a natural language target parsing algorithm, and generates a structured task tree with business domain labels, hierarchical structure and dependency relationships. The structured task tree contains multiple sub-task nodes and dependency edges that indicate the execution order between nodes. The first allocation module is used to map each subtask node to the corresponding candidate enterprise intelligent agent set according to the business domain label, so as to obtain the first allocation result; The second allocation module is used to calculate the task load index of each candidate enterprise agent based on the first allocation result, perform a difference operation between the task load index and a preset load threshold to obtain a load difference value, and dynamically reallocate the first allocation result according to the load difference value to obtain the second allocation result. The task execution module is used to asynchronously execute the mapped sub-task nodes according to the execution order represented by the dependency edges, based on the second allocation result. During the execution, the context inheritance algorithm is used to persist the task state and call the context, and the exception self-healing algorithm is used to automatically identify, retry, roll back and report tasks with execution exceptions. The strategy generation module is used to use compliance auditing algorithms to record and control permissions for the entire process, integrate the execution results of each sub-task node, and generate processing strategies for business instructions in the enterprise management system.

[0041] Specifically, in one particular implementation, the enterprise multi-agent task autonomous processing system based on a large model of the present invention implements the above method through the following functional modules: The instruction receiving module, deployed in the user interaction layer, receives business instructions from the enterprise management system and uses a natural language target parsing algorithm to generate a structured task tree. The first allocation module performs business domain label matching to generate a preliminary draft of candidate enterprise agent mappings. The second allocation module calculates task load indices and analyzes load differences, dynamically reallocates tasks based on the analysis results, and outputs the second allocation result after load balancing. The task execution module schedules each candidate enterprise agent to asynchronously execute sub-task nodes according to dependency edge constraints, and uses context inheritance and exception self-healing algorithms to ensure execution continuity during execution. The strategy generation module uses a compliance auditing algorithm to audit and control permissions throughout the entire process, integrates the execution results of each sub-task node, and finally generates a processing strategy for the enterprise management system's business instructions.

[0042] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for autonomous processing of enterprise multi-agent tasks based on a large model, characterized in that, Includes the following steps: S1. Receive business instructions from the enterprise management system, perform semantic parsing and business domain identification on the business instructions using a natural language target parsing algorithm, and generate a structured task tree with business domain labels, hierarchical structure and dependency relationships. The structured task tree contains multiple sub-task nodes and dependency edges that indicate the execution order between nodes. S2. Based on the business domain labels, map each subtask node to the corresponding candidate enterprise intelligent agent set to obtain the first allocation result; S3. Based on the first allocation result, calculate the task load index of each candidate enterprise agent, perform a difference operation between the task load index and the preset load threshold to obtain the load difference value, and dynamically reallocate the first allocation result according to the load difference value to obtain the second allocation result. S4. Based on the second allocation result, each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the execution order represented by the dependency edge. During the execution process, the context inheritance algorithm is used to persist the task state and call the context, and the exception self-healing algorithm is used to automatically identify, retry, roll back and report tasks with execution exceptions. S5. Adopt a compliance audit algorithm to record and control access to the entire process, integrate the execution results of each sub-task node, and generate processing strategies for business instructions in the enterprise management system.

2. The enterprise multi-agent task autonomous processing method based on a large model according to claim 1, characterized in that, In step S1, the business instructions are semantically parsed and their business domains are identified using a natural language target parsing algorithm, generating a structured task tree with business domain labels, hierarchical structure, and dependencies, including: The system performs word segmentation and semantic parsing on business instructions in the enterprise management system, extracts business objectives, constraints and key business entities, and identifies and generates business domain tags based on the domain to which the key business entities belong. A two-level splitting strategy is adopted to decompose business objectives into phased main tasks to form a task skeleton. Each phased main task is then further broken down into multiple executable sub-task nodes. Based on the logical sequence and data dependency relationships between each subtask node, dependency edges representing the execution order between nodes are constructed, and business domain labels, subtask nodes, and dependency edges are combined to generate a structured task tree.

3. The enterprise multi-agent task autonomous processing method based on a large model according to claim 2, characterized in that, In S2, the candidate enterprise intelligent agent set includes financial intelligent agent, human resources intelligent agent, supply chain intelligent agent, compliance intelligent agent and operation intelligent agent; The financial intelligent agent is used to execute financial domain sub-task nodes, which include financial reconciliation, voucher generation, report output, and tax declaration. The human intelligence agent is used to execute human resources domain sub-task nodes, which include personnel entry, transfer, and departure, payroll calculation, and attendance approval. The supply chain agent is used to execute supply chain domain sub-task nodes, which include procurement fulfillment, inventory management, and logistics tracking. The compliance agent is used to execute compliance domain sub-task nodes, which include compliance review, risk identification, and audit recording. The operational intelligence agent is used to execute operational domain sub-task nodes, which include operational monitoring, data aggregation, and process supervision.

4. The enterprise multi-agent task autonomous processing method based on a large model according to claim 3, characterized in that, In step S2, each subtask node is mapped to the corresponding candidate enterprise intelligent agent set according to the business domain label, and the first allocation result is obtained, including: Traverse each subtask node in the structured task tree, extract the business domain label of each subtask node, and compare and match the business domain label with the preset business domain identifier of each candidate enterprise intelligent agent in the candidate enterprise intelligent agent set. When the business domain label of a subtask node matches the preset business domain identifier of a candidate enterprise intelligent agent, the subtask node is mapped to the corresponding candidate enterprise intelligent agent, and the initial mapping relationship between the subtask node and the candidate enterprise intelligent agent is established. The initial mapping relationships between all subtask nodes and candidate enterprise agents are integrated to obtain the first allocation result.

5. The enterprise multi-agent task autonomous processing method based on a large model according to claim 4, characterized in that, In step S3, based on the first allocation result, the task load index of each candidate enterprise agent is calculated, and the difference between the task load index and a preset load threshold is calculated to obtain the load difference value, including: Based on the first allocation result, the number of subtask nodes mapped to each candidate enterprise agent and the preset complexity weight of each subtask node are extracted. The task load index of each candidate enterprise agent is calculated by weighting the number of subtask nodes with the preset complexity weight. ； in, Let be the task load index of the i-th candidate enterprise intelligent agent. This represents the number of subtask nodes mapped to the i-th candidate enterprise agent. This is the set of subtask nodes mapped to the i-th candidate enterprise agent. α and β are the preset complexity weights for the j-th subtask node, where α and β are both set weighting coefficients; The difference between the obtained task load index and the preset load threshold is calculated to obtain the load difference value, which provides a basis for dynamic redistribution.

6. The enterprise multi-agent task autonomous processing method based on a large model according to claim 5, characterized in that, In step S3, the first allocation result is dynamically reallocated based on the load difference value to obtain the second allocation result, including: Candidate enterprise agents with load difference values ​​greater than zero are identified as overloaded, and subtask nodes to be migrated are extracted from their mapped subtask nodes. Match the subtask nodes to be migrated to other candidate enterprise intelligent agents with load difference values ​​less than zero and the same business domain processing capabilities according to the business domain label, and re-establish the mapping relationship between subtask nodes and candidate enterprise intelligent agents. The mapping relationship between all subtask nodes after reallocation and candidate enterprise agents is integrated to obtain the second allocation result.

7. The enterprise multi-agent task autonomous processing method based on a large model according to claim 6, characterized in that, In step S4, based on the second allocation result, each candidate enterprise agent asynchronously executes the mapped sub-task nodes according to the execution order represented by the dependency edges, including: Based on the second allocation result, the sub-task nodes mapped by each candidate enterprise agent are extracted. The execution order of each sub-task node is determined according to the execution order represented by the dependency edges in the structured task tree. Sub-task nodes without dependencies are determined as nodes that can be executed in parallel, and sub-task nodes with dependencies are determined as nodes that can be executed sequentially. Each candidate enterprise intelligent agent asynchronously executes the mapped sub-task nodes according to the execution sequence. During the execution process, the context inheritance algorithm is used to persist the task state and call the context, ensuring that the output of the previous sub-task node serves as the input basis for the subsequent sub-task node.

8. An enterprise multi-agent task autonomous processing system based on a large model, characterized in that, include: The instruction receiving module is used to receive business instructions from the enterprise management system, perform semantic parsing and business domain identification on the business instructions through a natural language target parsing algorithm, and generate a structured task tree with business domain labels, hierarchical structure and dependency relationships. The structured task tree contains multiple sub-task nodes and dependency edges that indicate the execution order between nodes. The first allocation module is used to map each subtask node to the corresponding candidate enterprise intelligent agent set according to the business domain label, so as to obtain the first allocation result; The second allocation module is used to calculate the task load index of each candidate enterprise agent based on the first allocation result, perform a difference operation between the task load index and a preset load threshold to obtain a load difference value, and dynamically reallocate the first allocation result according to the load difference value to obtain the second allocation result. The task execution module is used to asynchronously execute the mapped sub-task nodes according to the execution order represented by the dependency edges, based on the second allocation result. During the execution, the context inheritance algorithm is used to persist the task state and call the context, and the exception self-healing algorithm is used to automatically identify, retry, roll back and report tasks with execution exceptions. The strategy generation module is used to use compliance auditing algorithms to record and control permissions for the entire process, integrate the execution results of each sub-task node, and generate processing strategies for business instructions in the enterprise management system.

Images