Management agent generation and scheduling method based on large model and dynamic workflow

Abstract

This invention discloses a method for generating and scheduling management agents based on a large model and dynamic workflow. The method includes: analyzing business requirements through a large model, automatically generating and instantiating specialized management agents according to capability matching; dynamically generating workflows based on a rule engine and node selection probability, and adjusting the process structure in real time according to the execution status; employing a centralized multi-agent scheduling algorithm to decompose tasks into atomic tasks, allocating and adapting agents according to comprehensive scheduling to achieve collaborative scheduling and load balancing; and establishing an execution feedback and optimization mechanism, collecting logs and analyzing them using a large model to update rule weights and scheduling parameters. This invention can automatically generate management agents, shorten deployment cycles, reduce costs, improve process flexibility and system stability, form a self-optimizing closed loop, and significantly improve the efficiency of multi-agent collaboration and the level of system intelligence.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and enterprise management technology, and in particular to a method for generating and scheduling management intelligent agents based on large models and dynamic workflows. Background Technology

[0002] With the rapid development of large-scale model technology, intelligent agents have become core software entities for perceiving the environment, making autonomous decisions, and collaboratively executing tasks, and are widely used in business scenarios such as enterprise management, customer service, project collaboration, and risk control. However, existing management-type intelligent agents have significant technical shortcomings in agent generation, workflow construction, multi-agent scheduling, and operational feedback optimization, making it difficult to adapt to the complex, dynamic, and ever-changing management needs of enterprises, as detailed below: (1) The generation of intelligent agents relies on manual configuration, and the automation and scenario adaptation capabilities are insufficient. The creation, capability definition, rule configuration and parameter optimization of existing intelligent agents all require manual development by technical personnel, and it is impossible to automatically parse the requirements and generate suitable management intelligent agents according to the business scenario. When enterprises add management scenarios such as prevention and control of illegal guarantees, pre-review of major issues, and review of entertainment expenses, it is necessary to redevelop, deploy and configure management intelligent agents, which has problems such as long development cycle, high implementation cost and weak scalability, and cannot quickly respond to dynamically changing business needs.

[0003] (2) The workflow adopts a static and fixed configuration, which has poor flexibility and adaptability. The task flow of the existing system is mostly a fixed structure. The process nodes, flow rules and approval logic are pre-set in the system deployment stage and cannot be dynamically adjusted according to business type, task attributes and risk level. Taking contract review as an example, the review rules, approval nodes and risk points of procurement, sales and financial contracts are significantly different. The static workflow is difficult to adapt to the dynamic review needs of multiple types, multiple scenarios and multiple rules, and its universality and adaptability are insufficient.

[0004] (3) Inefficient multi-agent collaborative scheduling, poor capability matching and load balancing. In scenarios where multiple agents collaborate to execute complex tasks, existing technologies lack efficient scheduling and allocation mechanisms. They typically use simple methods such as polling and random allocation to allocate tasks without comprehensively considering the capability adaptability of sub-agents, real-time load status, and historical execution efficiency. For example, in the financial leasing scenario, multiple sub-agents such as collateral screening, leased property verification, and lessee risk assessment need to be called in a coordinated manner. Unreasonable task allocation can easily lead to execution congestion, resource waste, and task queuing timeouts, resulting in low overall collaborative efficiency.

[0005] (4) Lack of operational closed-loop feedback and dynamic optimization mechanism. The existing system cannot collect operational status, execution results, time consumption indicators and abnormal information in real time during task execution. It cannot dynamically adjust the workflow structure, scheduling strategy and agent configuration parameters based on actual operation data, making it difficult to achieve continuous iterative optimization of processes and strategies. The overall stability, execution efficiency and intelligence level of the system are significantly limited. Summary of the Invention

[0006] This invention provides a method for generating and scheduling management agents based on a large model and dynamic workflow. It solves the technical problems of agent generation relying on manual configuration, insufficient automation and scenario adaptability, static and fixed configuration of workflow, poor flexibility and adaptability, inefficient multi-agent collaborative scheduling, poor capability matching and load balancing, and lack of runtime closed-loop feedback and dynamic optimization mechanism.

[0007] This invention provides a method for generating and scheduling management agents based on large models and dynamic workflows, including: Based on the large model, semantic parsing of user business requirements is performed to extract business elements. Through capability matching degree calculation, intelligent agent configuration information is automatically generated and instantiated to create a special management intelligent agent. A workflow rule base is pre-built; tasks to be executed are received, and a dynamic workflow orchestration method based on a rule engine is adopted. Adaptive workflow nodes are selected by calculating the node selection probability, and workflow instances that adapt to business types and task attributes are dynamically generated. During execution, the workflow structure is adjusted according to the real-time execution status and dynamic adjustment rules. A centralized multi-agent collaborative scheduling algorithm is adopted to decompose the workflow node tasks into atomic tasks. Through comprehensive scheduling and computation, an appropriate agent is assigned to each atomic task to achieve multi-agent collaborative scheduling and load balancing. Build an execution feedback and optimization iteration mechanism, collect log data during the execution process, analyze the log data using a large model, and update the rule weights and scheduling parameters in the workflow rule base.

[0008] Optionally, the process of automatically generating agent configuration information and completing instantiation includes: Input the user's business requirements into the large model and extract four business elements: business objectives, task types, input and output requirements, and constraints. The required capabilities are matched from the intelligent agent capability model based on business elements. The intelligent agent capability model includes document processing capability, data acquisition capability, natural language processing capability, business process processing capability, and system integration capability. Computational capability matching degree, automatically generate intelligent agent configuration information; The management agent is instantiated based on the agent configuration information. The management agent includes a perception module, a decision-making module, an execution module, and a feedback module.

[0009] Optionally, creating a specialized management agent includes: Configure multi-system data access capabilities, substantive guarantee clause identification capabilities, and multi-dimensional verification capabilities for illegal guarantees to generate an intelligent agent for illegal guarantees; Configure the ability to collect and analyze topic information, intelligently match the list of items, conduct compliance review of the initiating department, and conduct compliance review of the decision-making process to generate a pre-review intelligent agent for the three major issues; Configure automatic document recognition and intelligent compliance verification capabilities to generate an intelligent entity for pre-auditing entertainment expenses; Configure the ability to acquire and parse contract attachments, the ability to review text compliance, and the ability to review business compliance, and generate a smart contract review agent; Configure high-risk screening capabilities for collateral, leased assets, lessees / guarantors, and loan disbursement conditions to generate a financial leasing intelligent agent.

[0010] Optionally, the compliance score of the pre-audit agent for entertainment expenses is calculated according to the following formula: , In the above formula, For compliance scoring, To indicate whether the first condition is satisfied Indicator functions for compliance rules, This represents the total number of compliance rules.

[0011] Optionally, the contract risk score of the intelligent contract review agent is calculated according to the following formula: , In the above formula, Score the contract risk. For the first Risk factor of such clauses For the first The severity score of the violation of the clauses, This represents the total number of clause categories.

[0012] Optionally, the ability matching degree is calculated according to the following formula: , In the above formula, For capability matching, For the first The required score for each ability For the first The weighting coefficient of each ability, This represents the total number of capabilities.

[0013] Optionally, the node selection probability is calculated according to the following formula: , In the above formula, To select a node The probability, For tasks and nodes similarity, For temperature coefficient, This represents the total number of candidate nodes.

[0014] Optionally, the integrated scheduling score is calculated according to the following formula: , In the above formula, For integrated scheduling, For capability matching, The current load factor, Based on historical performance, As task priority weight, This is the weighting coefficient for ability matching. The current load factor weighting coefficient. Weighting coefficients for historical performance. This is the task priority weight coefficient, and .

[0015] Optionally, the steps for updating rule weights specifically include: Collect log data during workflow execution, including task execution time, resource consumption, and execution results; Analyze log data using large models to identify execution bottlenecks and update rule weights in the workflow rule base.

[0016] Optionally, the rule weights are updated according to the following formula: , In the above formula, For the updated rule weights, The original rule weights, For learning rate, Rate the current execution performance. The historical average performance score is used.

[0017] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. To address the shortcomings of existing technologies in generating intelligent agents, which rely on manual configuration and lack automation and scenario adaptability, this invention proposes a demand-driven method for automatically generating specialized intelligent agents based on semantic parsing of business requirements using a large model. By calculating the capability matching degree, the method automatically generates intelligent agent configurations and completes instantiation. It can quickly create specialized management intelligent agents for scenarios such as illegal guarantee prevention, preliminary review of major issues, and audit of entertainment expenses without manual development, significantly shortening the deployment cycle of intelligent agents, reducing implementation costs, and significantly improving the automation level and scenario adaptability of intelligent agent generation.

[0018] 2. To address the shortcomings of existing workflow technologies that employ static and fixed configurations, resulting in poor flexibility and adaptability, this invention innovatively proposes a dynamic workflow orchestration method based on a rule engine. This method dynamically generates workflow instances that adapt to business types and task attributes by calculating node selection probability. During execution, the workflow structure is adjusted based on real-time execution status and dynamic adjustment rules, effectively improving the flexibility, versatility, and dynamic adaptability of workflow orchestration.

[0019] 3. To address the shortcomings of existing technologies in multi-agent collaborative scheduling, such as inefficiency, poor capability matching, and poor load balancing, this invention innovatively proposes a centralized multi-agent collaborative scheduling algorithm. By comprehensively scheduling and calculating, the algorithm allocates tasks to agents, achieving efficient collaboration and load balancing among multiple agents, avoiding task congestion and resource waste, and significantly improving the parallel execution efficiency of multiple agents and the overall system stability under complex tasks.

[0020] 4. In view of the shortcomings of existing technologies that lack operational closed-loop feedback and dynamic optimization mechanisms, this invention innovatively constructs an execution feedback and optimization iteration mechanism to continuously update rule weights and scheduling parameters, forming a self-optimizing closed loop, enabling the system to have continuous self-optimization and self-evolution capabilities, and comprehensively improving the level of intelligence and execution effect.

[0021] In summary, existing technologies have significant shortcomings in four aspects: automation level, process flexibility, collaborative efficiency, and self-optimization, making it difficult to meet the dynamic management needs of enterprises. Attached Figure Description

[0022] Figure 1 This is a flowchart of a management agent generation and scheduling method based on a large model and dynamic workflow according to the present invention. Detailed Implementation

[0023] This invention provides a method for generating and scheduling management agents based on a large model and dynamic workflow. It solves the technical problems of agent generation relying on manual configuration, insufficient automation and scenario adaptability, static and fixed configuration of workflow, poor flexibility and adaptability, inefficient multi-agent collaborative scheduling, poor capability matching and load balancing, and lack of runtime closed-loop feedback and dynamic optimization mechanism.

[0024] To better understand, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described in this invention are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0025] I. Overall Methodology and Process like Figure 1 As shown, the present invention provides a method for generating and scheduling management agents based on large models and dynamic workflows (hereinafter referred to as the method), including steps S1 to S4.

[0026] S1: Based on the large model, semantic parsing of user business requirements is performed to extract business elements. Through capability matching degree calculation, intelligent agent configuration information is automatically generated and instantiated to create a special management intelligent agent.

[0027] In step S1, the process of automatically generating intelligent agent configuration information and completing instantiation includes: inputting the user's business requirements description into the large model, extracting four business elements: business objectives, task type, input / output requirements, and constraints; matching the required capabilities from the intelligent agent capability model based on the business elements, wherein the intelligent agent capability model includes document processing capability, data acquisition capability, natural language processing capability, business process processing capability, and system integration capability; calculating the capability matching degree, and automatically generating intelligent agent configuration information; and instantiating the management intelligent agent based on the intelligent agent configuration information, wherein the management intelligent agent includes a perception module, a decision-making module, an execution module, and a feedback module.

[0028] In step S1, creating a special management intelligent agent includes: configuring multi-system data access capabilities, substantive guarantee clause identification capabilities, and multi-dimensional verification capabilities for illegal guarantees (external group guarantees, guarantees exceeding shareholding ratios, unapproved guarantees, and guarantees without repayment capacity), generating an illegal guarantee intelligent agent; configuring issue information collection and parsing capabilities, intelligent matching capabilities for item lists, compliance review capabilities for initiating departments, and compliance review capabilities for decision-making procedures, generating a pre-review intelligent agent for major issues; configuring automatic document recognition capabilities and intelligent compliance verification capabilities, generating an entertainment expense pre-review intelligent agent; configuring contract attachment acquisition and parsing capabilities, text compliance review capabilities, and business compliance review capabilities, generating a contract intelligent review intelligent agent; configuring high-risk screening capabilities for collateral, leased assets, lessees / guarantors, and loan disbursement conditions fulfillment risk screening capabilities, generating a financial leasing intelligent agent.

[0029] The compliance score for the pre-audit agent of entertainment expenses is calculated according to the following formula: , In the above formula, For compliance scoring, To indicate whether the first condition is satisfied Indicator functions for compliance rules, This represents the total number of compliance rules.

[0030] The contract risk score of the intelligent contract review agent is calculated according to the following formula: , In the above formula, Score the contract risk. For the first Risk factor of such clauses For the first The severity score of the violation of the clauses, This represents the total number of clause categories.

[0031] In step S1, the capability matching degree is calculated according to the following formula: , In the above formula, For capability matching, For the first The required score for each ability For the first The weighting coefficient of each ability, This represents the total number of capabilities.

[0032] S2: Pre-build a workflow rule base, which includes task classification rules, node mapping rules, process combination rules, and dynamic adjustment rules; receive tasks to be executed, adopt a dynamic workflow orchestration method based on a rule engine, select suitable workflow nodes through node selection probability calculation, dynamically generate workflow instances that adapt to business types and task attributes, and adjust the workflow structure during execution based on real-time execution status and dynamic adjustment rules, including adding and deleting nodes, switching paths, and parallel adjustments.

[0033] In step S2, the node selection probability is calculated according to the following formula: , In the above formula, To select a node The probability, For tasks and nodes similarity, For temperature coefficient, This represents the total number of candidate nodes.

[0034] In step S2, the specific steps for selecting the appropriate workflow node are as follows: select the appropriate workflow node from the workflow node library according to the task type and node mapping rules.

[0035] In step S2, the specific steps for dynamically generating workflow instances are as follows: dynamically generate workflow instances according to the process combination rules, and determine the connection relationship and execution order between nodes.

[0036] S3: Employs a centralized multi-agent cooperative scheduling algorithm, decomposing workflow node tasks into atomic tasks. Through comprehensive scheduling and computation, an appropriate agent is assigned to each atomic task, achieving multi-agent cooperative scheduling and load balancing. The multi-agent cooperative scheduling includes agent registration management, task queue management, and scheduling decisions.

[0037] In step S3, the integrated scheduling score is calculated according to the following formula: , In the above formula, For integrated scheduling, For capability matching, The current load factor, Based on historical performance, As task priority weight, This is the weighting coefficient for ability matching. The current load factor weighting coefficient. Weighting coefficients for historical performance. This is the task priority weight coefficient, and .

[0038] In step S3, the process of assigning an adaptive agent to each atomic task specifically includes: the agent registering with the central scheduler and reporting its capability description and current status; the central scheduler receiving the workflow execution request, decomposing the workflow node task into atomic tasks and adding them to the task queue; calculating the comprehensive scheduling score for each candidate agent; selecting the agent with the highest comprehensive scheduling score to execute the task and updating the agent's status; receiving the agent's execution result feedback, updating the agent's historical performance record, and triggering subsequent task scheduling.

[0039] S4: Build an execution feedback and optimization iteration mechanism, collect log data during the execution process, analyze the log data using a large model, and update the rule weights and scheduling parameters in the workflow rule base.

[0040] In step S4, the specific steps for updating rule weights include: collecting log data during the workflow execution process, including task execution time, resource consumption, and execution results; analyzing the log data using a large model to identify execution bottlenecks and updating the rule weights in the workflow rule base.

[0041] In step S4, the rule weights are updated according to the following formula: , In the above formula, For the updated rule weights, The original rule weights, For learning rate, Rate the current execution performance. The historical average performance score is used.

[0042] This method also includes optional steps: unified management, configuration, and querying of various management agents, supporting view switching by type and name search. The specific implementation process includes: using the agent management unit as a centralized management entry point for various agents, supporting centralized management, information viewing, and initiation of dialogue operations; supporting view switching by "All / Intelligent Assistant / Business Expert / Risk Control" dimensions, and quickly locating target agents through agent name search; performing add, delete, and modify operations on agents within the agent management page, with configuration content including agent name, icon, description, workflow access address, and fixed input parameters.

[0043] This method also includes an optional step: maintaining the capability descriptions and state information of registered agents.

[0044] This method also includes optional steps: monitoring task execution status and agent load.

[0045] II. Irregular Guarantee Smart Entities Taking a non-compliant guarantor as an example, the specific implementation process of this method is explained as follows: Step P1: The user inputs their business requirements through the intelligent agent management module: "We need an intelligent agent that can automatically identify guarantee clauses in contracts, verify whether the guarantor is an external company, whether it exceeds the shareholding ratio, whether it has been reviewed, and whether the guaranteed party has the ability to repay." Step P2: The requirement analysis and agent generation module uses a large model to analyze the user's business requirement description and extract business elements: the business objective is the identification and verification of illegal guarantees, the task type is risk prevention and control, the input requirement is contract text, the output requirement is a verification result report, and the recommendation capabilities are text parsing, rule verification, and financial data analysis. Computational capability matching degree. Generate configuration information for the intelligent agent that provides illegal guarantees, including: data interfaces for connecting to the contract law system, financial system, and reform planning system; configuration rules for identifying substantive guarantee clauses (including core keywords such as "guarantee", "guarantee", "joint and several liability", "mortgage", and "pledge"); and configuration of four violation verification rules.

[0046] Step P3: The dynamic workflow orchestration module calculates the node selection probability based on the characteristics of the non-compliance guarantee task. Select workflow nodes. The generated workflow instance includes: data acquisition nodes. Text parsing node Identification Nodes of Guarantee Terms Non-compliance verification nodes (parallel execution of verification of external guarantees, guarantees exceeding shareholding ratio, guarantees without approval, and lack of repayment ability) Results summary node Output node. During workflow execution, if the guarantee clause identification node times out, the dynamic adjustment unit automatically adds a parallel text parsing node to distribute the workload.

[0047] Step P4: The central scheduling module decomposes workflow node tasks into atomic tasks: data acquisition task, text parsing task, guarantee clause recognition task, and four verification tasks. The scheduling decision engine calculates the comprehensive scheduling score. Assign an adaptive agent to each atomic task, where the capability matching degree =0.85, Current Load Factor =0.2, Historical Performance Score =0.9, Task Priority Weight =0.8, the calculated integrated scheduling score is 0.8. =0.84, select the agent with the highest comprehensive scheduling score as the adaptive agent.

[0048] Step P5: The feedback optimization module collects execution logs and finds that the execution time of the over-equity ratio guarantee verification node is relatively long. Large-scale model analysis suggests optimizing the verification algorithm. Automatic updates to rule weights and learning rate are also implemented. The current performance score is 0.1. =0.7, historical average performance score =0.85, original rule weight =0.5, calculate the updated weights. =0.485.

[0049] III. Three Major Issues Pre-review Intelligent Agent Taking the pre-review agent for the "Three Major Issues" as an example, the specific implementation process of this method is explained as follows: Step T1: User inputs business requirements: "We need an intelligent agent that can pre-review the agenda for the general manager's office meeting, automatically match the list of items, verify the permissions of the initiating department, and review the compliance of the decision-making process." Step T2: The requirements analysis and agent generation module generates the configuration for the pre-review agent of the "Three Major Issues" (a set of responsibilities and key tasks). This includes configuring the agent's capabilities for: issue information collection (interfacing with the OA system), item list matching (referring to the "List of 33 Decision-Making Items for the General Manager's Office Meeting"), compliance review by the initiating department, and compliance review of the decision-making process.

[0050] Step T3: The dynamic workflow orchestration module generates a workflow instance: topic data collection node. Issue Analysis Node Item list matching node Initiating department verification node Decision-making process verification node Results summary node Output node. The item list matching node calculates the similarity with 33 items. The item with the highest similarity is selected as the matching result.

[0051] Step T4: After execution, the feedback optimization module found that the accuracy of the item list matching was low, and the large model analysis suggested optimizing the matching algorithm. The system automatically updated the matching rules, upgrading the keyword matching algorithm to a semantic similarity matching algorithm.

[0052] IV. Intelligent Contract Review Agent Taking the intelligent contract review agent as an example, the specific implementation process of this method is explained as follows: Step F1: User inputs business requirements: "We need an intelligent agent that can intelligently review contracts, examine the integrity of the contract form and the risks of core clauses, and classify the review according to contract type." Step F2: The requirements analysis and agent generation module generates the intelligent agent configuration for contract review. This includes configuring capabilities for acquiring and parsing contract attachments (interfacing with the contract legal system), text compliance review (formal integrity review, core clause risk review), and business compliance review (categorized by procurement, sales, finance, and general categories).

[0053] Step F3: The dynamic workflow orchestration module generates a workflow instance. For procurement contracts, the workflow includes: a contract attachment retrieval node. Text parsing node Formal integrity review nodes (verification of name, number, signing date, full name of entity / unified social credit code) Key risk review points for core terms (subject matter, payment terms, performance period, acceptance, and liability for breach of contract) Compliance review points for procurement-related business (verified in conjunction with bidding management system) Risk outcome summary node.

[0054] Step F4: Calculate the contract risk score For example, if a risk is found in the terms of the subject matter, the risk factor is... =0.6, score for severity of violation =0.8, the risk score for this clause =0.48. Summarize the risk scores for all clauses to generate a contract risk review report.

[0055] V. Intelligent System for Financial Leasing Step A1: User inputs business requirements: "We need an intelligent agent that can screen risks in financial leasing business, including risks of collateral, leased assets, lessees, and guarantors, and check the fulfillment of loan disbursement conditions." Step A2: The requirements analysis and agent generation module generates the configuration of the financial leasing agent. This includes configuring high-risk screening capabilities for collateral (identifying uncollateralizable assets), high-risk screening capabilities for leased assets (identifying suitable, restricted, and prohibited leased assets), high-risk screening capabilities for lessees / guarantors (distinguishing between rules for state-owned enterprises and private enterprises), and risk screening capabilities for loan disbursement condition fulfillment.

[0056] Step A3: The dynamic workflow orchestration module generates a workflow example: Attachment collection node. Attachment parsing node Collateral screening nodes Rental item screening nodes Tenant screening nodes Guarantor screening process Loan disbursement screening process Risk summary node Output node. Each screening node executes in parallel.

[0057] Step A4: During the execution process, if the collateral screening node discovers unmortgageable properties such as schools and kindergartens, they will be automatically marked as high-risk, triggering an alert. The lessee screening node will verify according to the rules for private enterprises: if the lessee has been operating for less than 3 full years, or has committed serious illegal acts, or is included in the list of dishonest persons subject to enforcement, or has abnormal financial indicators (such as negative net profit attributable to the parent company for 2 consecutive years, negative net operating cash flow for 2 consecutive years, sales cash collection ratio for 2 consecutive years <0.6, and asset-liability ratio >80%), it will be automatically marked as high-risk and a risk report will be generated.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention also includes such modifications and variations.

Claims

1. A method for generating and scheduling management agents based on large models and dynamic workflows, characterized in that, include: Based on the large model, semantic parsing of user business requirements is performed to extract business elements. Through capability matching degree calculation, intelligent agent configuration information is automatically generated and instantiated to create a special management intelligent agent. Pre-build a workflow rule base; Receive tasks to be executed, adopt a dynamic workflow orchestration method based on a rule engine, select suitable workflow nodes by calculating the node selection probability, dynamically generate workflow instances that are suitable for business types and task attributes, and adjust the workflow structure according to the real-time execution status and dynamic adjustment rules during the execution process; A centralized multi-agent collaborative scheduling algorithm is adopted to decompose the workflow node tasks into atomic tasks. Through comprehensive scheduling and computation, an appropriate agent is assigned to each atomic task to achieve multi-agent collaborative scheduling and load balancing. Build an execution feedback and optimization iteration mechanism, collect log data during the execution process, analyze the log data using a large model, and update the rule weights and scheduling parameters in the workflow rule base.

2. The method according to claim 1, characterized in that, The process of automatically generating agent configuration information and completing instantiation includes: Input the user's business requirements into the large model and extract four business elements: business objectives, task types, input and output requirements, and constraints. The required capabilities are matched from the intelligent agent capability model based on business elements. The intelligent agent capability model includes document processing capability, data acquisition capability, natural language processing capability, business process processing capability, and system integration capability. Computational capability matching degree, automatically generate intelligent agent configuration information; The management agent is instantiated based on the agent configuration information. The management agent includes a perception module, a decision-making module, an execution module, and a feedback module.

3. The method according to claim 1, characterized in that, Creating a specialized management agent includes: Configure multi-system data access capabilities, substantive guarantee clause identification capabilities, and multi-dimensional verification capabilities for illegal guarantees to generate an intelligent agent for illegal guarantees; Configure the ability to collect and analyze topic information, intelligently match the list of items, conduct compliance review of the initiating department, and conduct compliance review of the decision-making process to generate a pre-review intelligent agent for the three major issues; Configure automatic document recognition and intelligent compliance verification capabilities to generate an intelligent entity for pre-auditing entertainment expenses; Configure the ability to acquire and parse contract attachments, the ability to review text compliance, and the ability to review business compliance, and generate a smart contract review agent; Configure high-risk screening capabilities for collateral, leased assets, lessees / guarantors, and loan disbursement conditions to generate a financial leasing intelligent agent.

4. The method according to claim 3, characterized in that, The compliance score for the pre-audit agent of entertainment expenses is calculated according to the following formula: ， In the above formula, For compliance scoring, To indicate whether the first condition is satisfied Indicator functions for compliance rules, This represents the total number of compliance rules.

5. The method according to claim 3, characterized in that, The contract risk score of the intelligent contract review agent is calculated according to the following formula: ， In the above formula, Score the contract risk. For the first Risk factor of such clauses For the first The severity score of the violation of the clauses, This represents the total number of clause categories.

6. The method according to claim 1, characterized in that, The ability matching degree is calculated according to the following formula: ， In the above formula, For capability matching, For the first The required score for each ability For the first The weighting coefficient of each ability, This represents the total number of capabilities.

7. The method according to claim 1, characterized in that, The node selection probability is calculated using the following formula: ， In the above formula, To select a node The probability, For tasks and nodes similarity, For temperature coefficient, This represents the total number of candidate nodes.

8. The method according to claim 1, characterized in that, The integrated scheduling score is calculated according to the following formula: ， In the above formula, For integrated scheduling, For capability matching, The current load factor, Based on historical performance, As task priority weight, This is the weighting coefficient for ability matching. The current load factor weighting coefficient. Weighting coefficients for historical performance. This is the task priority weight coefficient, and .

9. The method according to claim 1, characterized in that, The specific steps for updating rule weights include: Collect log data during workflow execution, including task execution time, resource consumption, and execution results; Analyze log data using large models to identify execution bottlenecks and update rule weights in the workflow rule base.

10. The method according to claim 1, characterized in that, Rule weights are updated according to the following formula: ， In the above formula, For the updated rule weights, The original rule weights, For learning rate, Rate the current execution performance. The historical average performance score is used.

Images