Business processing method and system based on multi-agent cooperation

Abstract

The application discloses a kind of based on multi-agent cooperation's service processing method and system, it is related to data processing technical field, it can be applied to finance, insurance field.The main purpose is to solve the problem of low efficiency of existing service processing.Mainly include response to service processing request, multi-dimensional task analysis is carried out to the business to be processed, and structured task tree is obtained;According to the task type of subtask and the execution logic relationship, based on hybrid triggering and dynamic routing mechanism, subtask is distributed to the predefined multiple agents, and the agent task execution sequence is obtained, wherein, any agent processes the task across business fields by calling and querying the corresponding dynamic knowledge base;Each agent cooperates to process subtask according to the agent task execution sequence, obtains the processing result of each different agent, and integrates the processing result of each agent through decision coordination agent, generates service processing result.It is mainly used for automated business processing.

Description

Technical Field

[0001] This invention relates to the field of data processing technology and can be applied to the fields of insurance and finance. In particular, it relates to a business processing method and system based on multi-agent collaboration. Background Technology

[0002] In highly complex and heavily regulated business sectors such as insurance and finance, business processes typically involve cross-step, multi-rule judgments and the processing of massive amounts of heterogeneous information. Traditional automated systems often employ rigid processes and rule engines, exhibiting shortcomings such as insufficient flexibility, lagging updates, and high maintenance costs when facing non-standardized cases, dynamically changing regulatory policies, and the need for cross-domain knowledge integration. In recent years, artificial intelligence system architectures centered on intelligent agents, with their modularity, autonomy, and collaboration, have provided a new technological paradigm for building next-generation business processing engines.

[0003] Existing agent-based business processing systems still face a critical bottleneck of rigid collaboration mechanisms when dealing with complex business scenarios such as insurance and finance. Specifically, most existing systems adopt predefined static workflows or simple message triggering mechanisms based on fixed rules, lacking the ability to adaptively perceive the inherent structure of tasks and real-time context. This rigid architecture prevents the system from dynamically decomposing tasks and scheduling resources based on the specific characteristics of the tasks (such as complexity, urgency, and domain overlap) and the current state of the system (such as the load of each agent and the update status of the knowledge base). As a result, when faced with non-standardized processes, sudden load fluctuations, or cross-domain complex tasks, the system exhibits slow response and uneven resource allocation, leading to low business processing efficiency. Summary of the Invention

[0004] In view of this, the present invention provides a business processing method and system based on multi-agent collaboration, the main purpose of which is to solve the problem of low efficiency in existing business processing.

[0005] According to one aspect of the present invention, a business processing method based on multi-agent collaboration is provided, comprising: responding to a business processing request, performing multi-dimensional task parsing on the business to be processed to obtain a structured task tree, wherein the structured task tree includes multiple sub-tasks and execution logic relationships between the sub-tasks; Based on the task type of the subtask and the execution logic relationship, the subtask is assigned to multiple predefined intelligent agents based on a hybrid triggering and dynamic routing mechanism, and an intelligent agent task execution sequence is obtained. In this sequence, any intelligent agent can process cross-business domain tasks by calling and querying the corresponding dynamic knowledge base. Each of the aforementioned intelligent agents performs collaborative processing on sub-tasks according to the task execution sequence of the intelligent agent, obtains the processing results of different intelligent agents, and integrates the processing results of each intelligent agent through a decision-coordinating intelligent agent to generate a business processing result.

[0006] According to another aspect of the present invention, a business processing system based on multi-agent collaboration is provided, comprising: The task parsing module is used to respond to business processing requests, perform multi-dimensional task parsing on the business to be processed, and obtain a structured task tree, wherein the structured task tree includes multiple sub-tasks and the execution logic relationship between each sub-task; The routing module is used to allocate the subtasks to multiple predefined agents based on the task type and execution logic relationship of the subtasks and a hybrid triggering and dynamic routing mechanism, and obtain the task execution sequence of the agents. Each agent can process cross-business domain tasks by calling and querying the corresponding dynamic knowledge base. The multi-agent collaboration module is used to collaboratively process sub-tasks according to the task execution sequence of each agent, obtain the processing results of each agent, and integrate the processing results of each agent through the decision coordination agent to generate business processing results.

[0007] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the above-described multi-agent cooperative business processing method.

[0008] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operations corresponding to the above-described multi-agent collaborative business processing method.

[0009] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a business processing method and system based on multi-agent collaboration. In this embodiment, in response to a business processing request, the system performs multi-dimensional task analysis to obtain a structured task tree. This structured task tree includes multiple sub-tasks and the execution logic relationships between them. Based on the task type and execution logic relationships of the sub-tasks, and using a hybrid triggering and dynamic routing mechanism, the sub-tasks are assigned to multiple predefined agents, resulting in an agent task execution sequence. Each agent processes cross-business domain tasks by calling and querying a corresponding dynamic knowledge base. Each agent collaboratively processes the sub-tasks according to the agent task execution sequence, obtaining processing results for each agent. A decision-coordinating agent integrates the processing results of each agent to generate the business processing result. This significantly reduces manual intervention and serial delays in traditional business processing, lowering business risks and compliance costs caused by human error or delayed knowledge updates. Simultaneously, its collaborative mechanism based on a dynamic knowledge base and hybrid routing ensures the system's adaptability and decision consistency in complex and changing business scenarios, thereby greatly improving the overall efficiency, accuracy, and intelligence level of business processing.

[0010] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. Furthermore, in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0011] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This invention provides a flowchart of a business processing method based on multi-agent collaboration. Figure 2 This invention provides a flowchart of another business processing method based on multi-agent collaboration. Figure 3 This diagram illustrates a block diagram of a multi-agent collaborative business processing system provided by an embodiment of the present invention. Figure 4 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] To address the problem of low business processing efficiency, this invention provides a business processing method based on multi-agent collaboration. This method can be based on a system implementation that integrates a task parser (task parsing module), a routing module (routing decision-maker), and a multi-agent coordination module. The multi-agent coordination module comprises a system consisting of multiple callable specialized agents. Figure 1 As shown, the method includes: 101. In response to a business processing request, perform multi-dimensional task analysis on the business to be processed to obtain a structured task tree.

[0014] In this embodiment of the invention, a business processing request can be a claim application form containing free text, images, and tables, etc. To transform the unstructured business processing request into a structured representation that can be understood and recognized by the computer system, the system receives the business processing request and performs task parsing through a task parser. This task parser integrates multiple technologies such as natural language understanding, image recognition, and key information extraction to perform multi-dimensional analysis of the request. Dimensions may include: business type identification (e.g., determining whether it is a car insurance claim, health insurance underwriting, or property insurance loss assessment), entity and relationship extraction (e.g., identifying "policyholder," "accident time," "loss object," "related responsible parties," and their relationships from the description), task complexity assessment (e.g., complexity assessment based on the number of responsible parties involved, the range of loss amounts, and the completeness of materials), and initial compliance screening (e.g., preliminary check whether the application is within the insurance period and whether it complies with the basic terms of the policy). Based on the above analysis results, the parser constructs a structured task tree. Each node in this tree structure represents an atomic or composite subtask (e.g., "verify driver's qualifications", "assess vehicle damage", "retrieve traffic police records from the accident scene"). The connections between nodes define the execution logic between subtasks, including but not limited to sequential execution, parallel execution, conditional branching (e.g., "if the loss exceeds 10,000 yuan, a fraud investigation must be initiated"), and circular dependencies. This structured task tree provides a precise execution basis for subsequent intelligent and dynamic scheduling.

[0015] 102. Based on the task type of the subtask and the execution logic relationship, and using a hybrid triggering and dynamic routing mechanism, the subtask is assigned to multiple predefined agents, and an agent task execution sequence is obtained.

[0016] In this embodiment of the invention, a hybrid triggering and dynamic routing mechanism is employed. Two driving factors jointly determine task allocation decisions, and dynamic routing identifies the agent responsible for handling sub-tasks. These two driving factors include event-driven mechanisms, where the completion of the previous task triggers the start of the next task; and data- or state-driven mechanisms, where specific data (e.g., "loss amount exceeds a preset threshold") or system status (e.g., "the current queue of the loss assessment agent is full") dynamically triggers or adjusts the task execution path. The dynamic routing process includes: the routing decision-maker receives each sub-task to be allocated and its metadata (task type, complexity, priority, data dependencies, etc.), and queries the state profile of each predefined agent in real time (e.g., current load, historical success rate for handling specific tasks, average response time, and the latest version number of its associated dynamic knowledge base). Based on preset optimization objectives (e.g., shortest global processing time, most balanced resources), the routing decision-maker uses a matching algorithm or multi-objective optimization algorithm to allocate sub-tasks to the most suitable agent in real time. For example, a subtask involving medical fraud detection might be routed to a specialized risk control agent that integrates the latest healthcare policy knowledge base and medical atlas, rather than a general rule engine. This ensures that even when faced with complex cross-domain tasks, each agent can perform its professional judgment supported by the latest and most relevant knowledge.

[0017] 103. Each of the aforementioned intelligent agents performs collaborative processing on sub-tasks according to the task execution sequence of the intelligent agent, obtains the processing results of different intelligent agents, and integrates the processing results of each intelligent agent through a decision-coordinating intelligent agent to generate a business processing result.

[0018] In this embodiment of the invention, various specialized intelligent agents work collaboratively under unified scheduling, ultimately forming a consistent business processing result. Each intelligent agent is activated in an ordered or parallel manner according to the agent task execution sequence defined in the task tree. After receiving the sub-task assigned to it and related context data, each intelligent agent independently invokes its internal specialized processing logic (such as model inference, rule matching, and knowledge retrieval) to process the data and generate its processing result. For example, the information collection intelligent agent outputs a "structured loss list," and the risk assessment intelligent agent outputs a "fraud risk score: 85 points." These results are accompanied by interpretable evidence, such as key influencing factors, inference logic, data sources used, and model confidence levels. Finally, the processing results of all intelligent agents are input into a decision coordination intelligent agent for integration of the various processing results. Specific integration methods can include conflict detection and resolution, such as arbitration when the loss assessment agent evaluates the loss at 50,000 yuan, while the anti-fraud agent believes the case has an extremely high fraud risk; information fusion and completion, such as piecing together fragmented evidence provided by different agents into a complete event chain; and global optimization decision-making, such as generating a final recommendation under multiple constraints including underwriting risk, customer satisfaction, operating costs, and compliance requirements, such as "pay compensation, but the amount is reduced to 40,000 yuan and recovery procedures are initiated" or "reject compensation, with specific terms and conditions as follows...". Finally, the decision-coordinating agent outputs a unified and interpretable business processing result.

[0019] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, the steps, based on the task type of the subtask and the execution logic relationship, and using a hybrid triggering and dynamic routing mechanism, assign the subtask to multiple predefined agents, and obtain the agent task execution sequence, including: 201. Determine the subtask triggering order based on the execution logic relationship and dynamic data.

[0020] 202. For each subtask, the subtask is assigned to multiple predefined agents according to the complexity of the subtask.

[0021] 203. Generate the subtask triggering order based on the subtask triggering order and the corresponding agents for each subtask.

[0022] In this embodiment of the invention, dynamic data includes dynamic data streams and system status events. After obtaining the inherent execution logic relationships (such as serial, parallel, and conditional dependencies) between subtasks in the task tree, the system not only relies on static logical relationships but also continuously monitors dynamic data streams and system status events. It dynamically calculates the triggering order of all subtasks to be processed at the current moment. This order is not completely static but can be adjusted in real time according to the execution context. For example, for logically parallel subtasks B and C, if only the data required to process subtask B is currently ready, the system will trigger subtask B first. As another example, even if the task tree presupposes that subtask G should be executed after subtask H, if the system detects in advance through the data stream that the key external data required to execute H (such as a third-party credit report) is ready, it can dynamically decide to trigger H earlier. Therefore, the final subtask triggering order is an actual execution sequence optimized by dynamic data streams and system status events within a static logical framework.

[0023] After determining the triggering time for each subtask, a suitable agent is assigned to it through a dynamic routing mechanism. The core dimension of the decision-making basis is the complexity of the subtask, which can be quantified and evaluated through one or more of the following factors: data dimension complexity, such as the number of data sources to be processed and the heterogeneity of data formats, such as whether it involves text, image, and speech fusion analysis; rule / model complexity, such as the number of rules to be invoked to solve the task, the computational intensity of the model algorithm, or the degree of overlap of the required domain knowledge; and decision risk level, such as the degree of impact of the task's conclusion on the final business outcome, for example, if the underwriting conclusion directly determines whether or not insurance is granted, it is considered a high-risk task. The system maintains an agent capability profile library, which records the task types that each predefined agent is good at handling, historical performance indicators (such as accuracy and average processing time) for handling tasks of different complexities, and its current load status. The dynamic routing algorithm matches the complexity characteristics of the subtasks with the agent's capability profile in real time.

[0024] The subtask triggering order obtained from the above steps is bound to the executing agent determined for each subtask to generate a complete and executable agent task execution sequence. This sequence is a list of instructions containing timing information. Each instruction clearly describes which data conditions must be met, which agent must execute which subtask, and the context of the required input data.

[0025] In one embodiment of the present invention, for further explanation and limitation, for each subtask, the subtask is assigned to a predefined plurality of intelligent agents according to the complexity of the subtask, including: Extract the task features of the subtask. If the task features match any complete processing path in the preset task library, treat the subtask as a regular task and assign an agent to the subtask through a pre-configured rule engine. If the task features do not match any complete processing path in the preset task library, the task features are classified and predicted using a lightweight natural language model. If the probability of a regular task is less than a preset threshold and the probability of a processable task is greater than a preset threshold, the subtask is determined as a semi-structured subtask, and an agent is assigned to the subtask using a lightweight machine learning model. If the task characteristics meet the conditions for a complex task, the subtask is identified as a complex task, and an agent is assigned to the subtask through an arbitrator based on reinforcement learning.

[0026] In this embodiment of the invention, for each subtask to be assigned, the system first extracts its task features from the task parsing results. This feature is a structured vector, typically including but not limited to: subtask type identifier, modality and scale of the input data (e.g., number of text characters, number of images), business domain labels, historical average processing time, structuredness score, and associated risk level labels. After obtaining the task features, a hierarchical, progressive decision-making logic is used to route the subtask to different levels of allocators based on the complexity reflected by the task features. Specifically, the system maintains a preset task library, which stores well-understood, deterministic standard task patterns and their complete standardized processing paths. The extracted task features are matched against the patterns in this library. If a match is successful, it indicates that the subtask belongs to a highly structured, process-fixed routine task, and this task is then handed over to a pre-configured rule engine for processing. This rule engine has built-in mapping rules between task types and default agents, enabling efficient and unambiguous assignment of specified agents to the subtask. For example, all "ID card OCR verification" subtasks are directly assigned to the information collection agent—image processing instance.

[0027] If a match fails, it indicates that the subtask does not belong to the standard pattern and may have some novelty or ambiguity. In this case, a lightweight natural language model (or classification model) is first invoked to analyze the task features. This model outputs the probability distribution of the task belonging to various known processing paths, paying particular attention to two key probability values: the probability of being classified as a regular task and the probability of being classified as a derable task, i.e., a non-standard task that the system is capable of handling. If the probability of a regular task is less than a low preset threshold (e.g., 0.3), and the probability of a derable task is greater than a high preset threshold (e.g., 0.7), then the subtask is classified as a semi-structured subtask. It is then handed over to a lightweight machine learning model (such as a gradient boosting decision tree or a small neural network) for assignment decisions. This model takes the task features as input, is trained to predict the classification probabilities of different agents as its output, and assigns an agent to the subtask accordingly.

[0028] Complex task conditions include cluster isolation, high-risk labeling, and low-confidence conditions. The cluster isolation condition involves calculating the distance between the task's features and the cluster centers formed after cluster analysis of all historical task features. If this distance exceeds a preset threshold, the task is considered a significant outlier or a completely new type. The high-risk labeling condition occurs when the subtask is marked as high-risk by business rules or upstream processes during its generation; for example, it may involve high insurance premiums, suspected fraud, or major disputed cases. The low-confidence condition occurs when the lightweight natural language model's prediction probability for any known processing path is below a very low preset threshold (e.g., all probabilities < 0.1), indicating extreme uncertainty on the part of the model. When any of these complex task conditions is met, the system submits the subtask to a reinforcement learning-based arbitrator. This arbitrator, acting as a global optimizer, considers not only the static matching degree between the task and the agent but also dynamically evaluates the current system load, the real-time performance status of each agent, task priority, and long-term learning objectives (such as exploring new strategies). By weighing immediate rewards against long-term value, it makes the final agent allocation decision. Through the aforementioned multi-layered judgment thresholds and progressive decision-making processes, the system achieves precise quantification and classification of task complexity. It can efficiently process a large number of routine tasks at high speed using a rule engine, ensuring basic efficiency; it can also flexibly handle common semi-structured tasks using machine learning models; and finally, it uses a reinforcement learning arbitrator to prudently handle complex, novel, or high-risk edge cases. This results in an overall optimized balance between processing efficiency and decision quality, significantly improving the robustness and adaptability of the intelligent agent collaborative system in real-world, complex business environments.

[0029] In one embodiment of the present invention, for further explanation and limitation, each of the intelligent agents performs collaborative processing of sub-tasks according to the task execution sequence of the intelligent agent to obtain the processing results of different intelligent agents, including: Based on the agent's task execution sequence, when a parent task node completes, a task completion event is published to the message queue using a publish-subscribe pattern; The triggered agent executes the processing of the corresponding subtask according to the received message, and generates the processing result of the agent; The subtask processing results, along with the task context, processing confidence, and new triggering events, are encapsulated into a standard format result message. The result message is published to the message queue to update the execution status of the corresponding node in the task tree and drive the triggering of subsequent logic.

[0030] In this embodiment of the invention, an asynchronous event-driven architecture based on a publish-subscribe pattern is used to advance the process according to a pre-generated sequence of agent task execution. Specifically, when a preceding (parent) task is completed, the executing agent publishes a task completion event to the message queue; subsequent agents subscribing to this event are then triggered, executing their corresponding subtasks and generating processing results. For subtasks that can be executed in parallel, the system achieves synchronous triggering by publishing broadcast events or a set of independent events, thereby improving the overall throughput.

[0031] After each agent completes a task, it encapsulates the processing result, task context, confidence level, and new triggering events into a standard-format result message and publishes it back to the message queue. This message serves two purposes: firstly, it updates the state of the corresponding node in the global task tree; secondly, its embedded triggering events automatically drive the execution of the next stage of the task, forming a closed-loop workflow. This loosely coupled design achieves decoupling and efficient collaboration between agents.

[0032] Throughout the message communication process, the system integrates key optimization strategies to ensure performance: compressed and decompressed messages to reduce network bandwidth and storage pressure; and intelligent caching of reusable intermediate data, prioritizing reading from the cache while ensuring data consistency, thus avoiding unnecessary duplicate calculations and data accesses, thereby significantly improving the system's processing efficiency and response speed when dealing with massive and multimodal data in insurance business.

[0033] In one embodiment of the present invention, for further explanation and limitation, the intelligent agent further includes an information collection intelligent agent, a risk assessment intelligent agent, a compliance and ethics intelligent agent, and a learning feedback intelligent agent; The triggered agent performs the processing of the corresponding sub-task according to the received message, including: If the trigger is an information collection agent, the processing includes: using a large model and a deep anti-counterfeiting algorithm to perform anti-counterfeiting verification on the data to be verified in the multi-source heterogeneous data carried by the sub-task, so that if the verification result passes, the multi-source heterogeneous data is subjected to multi-modal feature extraction and fusion to obtain a set of structured key information. If the risk assessment agent is triggered, the processing includes: performing risk prediction on the risk profile through a dynamic risk assessment model to obtain a dynamic risk score and risk profile; If the triggering agent is a compliance and ethics agent, the processing procedure includes: performing rule matching and compliance verification on the intermediate operations or processing results of any agent to obtain the compliance verification result; If the trigger is a learning feedback agent, the processing includes: constructing a fine-tuned sample set based on online prediction performance indicators and human feedback data within a preset historical period, and retraining parameters and calibrating rules for the specific model that has experienced performance degradation based on the fine-tuned sample set.

[0034] In this embodiment of the invention, at different stages of the entire business processing flow, each intelligent agent responds to a trigger and executes a corresponding sub-task. The execution relationship of each intelligent agent can be serial or parallel. The task processing process of each intelligent agent specifically includes: If the trigger is an information collection intelligent agent, this intelligent agent will use a large model and deep anti-counterfeiting algorithm to verify the authenticity and consistency of the multi-source heterogeneous data (such as text reports, on-site images, and voice recordings) carried by the sub-task. After the anti-counterfeiting verification is passed, the data will be subjected to multimodal feature extraction and fusion, and finally output a structured set of key information to provide standardized input for downstream tasks. If the trigger is a risk assessment intelligent agent, this intelligent agent will call a dynamic risk assessment model, and based on the risk profile and real-time data stream provided by the upstream, perform risk prediction and quantitative analysis, generate a dynamic risk score, and update the risk profile. If the trigger is a compliance and ethics intelligent agent, this intelligent agent will conduct real-time or post-event review of the intermediate operations or processing results of any intelligent agent, that is, by performing rule matching and logical verification with the embedded regulatory rule knowledge base, output compliance verification results to ensure the legality and ethics of the entire process.

[0035] Furthermore, the learning feedback agent, acting as the system's self-optimization engine, operates independently of specific business flows, focusing on the continuous evolution of the model. It continuously monitors the online predictive performance metrics of the models in the risk assessment agent and decision coordination agent, and simultaneously collects manual feedback on the system's final output. When performance metrics deviate from the benchmark or specific types of feedback occur, it proactively constructs a fine-tuning sample set and retrains and calibrates the parameters and rules of specific models identified as experiencing performance degradation, such as a classifier or risk assessment sub-model. After verifying that the optimized model's performance meets the standards on the validation set, it generates a model update package and deploys it during periods of low system load using hot updates or canary releases, replacing the old model version in the online service. This ensures the long-term stability and accuracy of the entire agent system's output.

[0036] In one embodiment of the present invention, for further explanation and limitation, the dynamic knowledge base is updated based on a dual-track update mechanism, and the update process specifically includes: Historical change data of insurance terms, regulatory policy texts, and risk model parameters are collected according to a preset cycle, and the structured features of the historical change data are extracted. An initial knowledge base is constructed based on the business type dimension features, clause validity dimension features, and time validity dimension features of the structured features; By adopting a bottom-up business-driven approach, dynamic information from business interaction processes is collected and processed in real time to generate the first type of update proposal; Through a top-down compliance-driven approach, normative documents from authoritative sources are collected and analyzed in real time to generate a second type of update proposal; The first type of update proposal and the second type of update proposal are compared with the benchmark knowledge base for consistency and conflict resolution. Based on the preset priority rules, the update content without conflict or after resolution is merged into the initial knowledge base to obtain the updated knowledge base. The updated knowledge base is distributed incrementally and synchronized to dynamic knowledge bases under different business types.

[0037] In this embodiment of the invention, the dynamic knowledge base achieves continuous evolution based on a dual-track update and fusion mechanism. The update process specifically includes: First, the system automatically collects historical change data of insurance clauses, regulatory policy texts, and risk model parameters according to a preset cycle, and extracts key structured features from them. Based on the business type dimension, clause validity dimension, and time validity dimension of these features, the system constructs and maintains a baseline initial knowledge base.

[0038] The update process is driven by two parallel paths: a bottom-up business-driven path, which collects and processes dynamic information and "bad cases" from actual business interactions (such as claims disputes and underwriting exceptions) in real time. The error understanding agent automatically determines the error type, outputs optimization solutions, and generates a first-type update proposal for specific rule or model optimization after expert confirmation. The top-down compliance-driven path monitors and analyzes normative documents from regulatory agencies, industry associations, and other relevant departments in real time, automatically extracts key change points, and generates a second-type update proposal to ensure compliance.

[0039] Subsequently, the two types of update proposals are automatically compared and conflict-resolved with the current initial knowledge base. Based on preset priority rules (preferably prioritizing compliance changes over business optimizations), conflict-free proposals or valid updates after resolution are securely integrated into the knowledge base, forming the updated knowledge base. Finally, to ensure system-wide knowledge consistency, the updated knowledge base is efficiently distributed and synchronized incrementally to dynamic knowledge base instances relied upon by different business types (such as auto insurance and health insurance) or different functional modules, ultimately completing the closed-loop update and global synchronization of the entire knowledge system.

[0040] In one embodiment of the present invention, for further explanation and limitation, the step of integrating the processing results of various intelligent agents through a decision-coordinating intelligent agent to generate business processing results includes: The conflict resolution rule base is invoked to perform consistency verification on the processing results of multiple agents, and a decision is made according to a preset priority in the event of a conflict during the consistency verification. The decision results and the processing results that pass the consistency check are merged to obtain the comprehensive processing result; Key decision-making factors and reasoning logic in the decision-making process are extracted through a pre-set decision interpretation mechanism; The business processing result is generated based on the key decision factors, reasoning logic, and comprehensive processing results.

[0041] In this embodiment of the invention, the decision-coordinating intelligent agent acts as the "command center" of the system, responsible for integrating and adjudicating the outputs of various specialized intelligent agents, ultimately forming a business conclusion. This process includes: First, the decision-coordinating intelligent agent invokes a built-in conflict resolution rule base to perform consistency verification on the received processing results from multiple intelligent agents (such as risk assessment results and compliance verification conclusions). When the verification finds logical or conclusion conflicts, the system will automatically adjudicate based on preset priority rules (such as compliance conclusions taking precedence over business risk assessments).

[0042] Subsequently, the decision-making coordination agent will integrate the final decision result after the adjudication with the processing result that has passed the consistency check, forming a comprehensive processing result. To improve the transparency and credibility of the decision, the system simultaneously uses a preset decision interpretation mechanism to trace and extract key decision factors that affect the final conclusion from the entire coordination and integration process, such as specific high-risk indicators, specific compliance clauses triggered, and their reasoning logic chains.

[0043] Ultimately, the decision-coordination agent encapsulates the aforementioned key decision factors, reasoning logic, and comprehensive processing results in a structured manner, generating a complete business processing result containing clear conclusions and interpretable evidence. This not only outputs intelligent judgments but also meets the requirements for transparency in the decision-making process during business review, regulatory audit, and customer communication.

[0044] This invention provides a business processing method based on multi-agent collaboration. In this embodiment, in response to a business processing request, the method performs multi-dimensional task analysis to obtain a structured task tree. This structured task tree includes multiple sub-tasks and the execution logic relationships between them. Based on the task type and execution logic relationships of the sub-tasks, and using a hybrid triggering and dynamic routing mechanism, the sub-tasks are assigned to multiple predefined agents, resulting in an agent task execution sequence. Each agent processes cross-business domain tasks by calling and querying a corresponding dynamic knowledge base. Each agent collaboratively processes the sub-tasks according to the agent task execution sequence, obtaining processing results for each agent. A decision-coordinating agent integrates the processing results of all agents to generate the business processing result. This significantly reduces manual intervention and serial delays in traditional business processing, lowering business risks and compliance costs caused by human error or delayed knowledge updates. Simultaneously, its collaborative mechanism based on a dynamic knowledge base and hybrid routing ensures the system's adaptability and decision consistency in complex and changing business scenarios, thereby greatly improving the overall efficiency, accuracy, and intelligence level of business processing. Furthermore, as a response to the above Figure 1 The implementation of the method shown in this invention provides a business processing system based on multi-agent collaboration, such as... Figure 3 As shown, the system includes: The task parsing module 31 is used to respond to a business processing request, perform multi-dimensional task parsing on the business to be processed, and obtain a structured task tree, wherein the structured task tree includes multiple sub-tasks and the execution logic relationship between each sub-task. The routing module 32 is used to allocate the subtask to multiple predefined intelligent agents based on the task type of the subtask and the execution logic relationship, and to obtain the task execution sequence of the intelligent agents, wherein any intelligent agent can process cross-business domain tasks by calling and querying the corresponding dynamic knowledge base; The multi-agent collaboration module 33 is used to collaboratively process sub-tasks according to the task execution sequence of each agent to obtain the processing results of each agent, and to integrate the processing results of each agent through the decision coordination agent to generate business processing results.

[0045] Furthermore, the routing module 32 includes: The determining unit is used to determine the subtask triggering order based on the execution logic relationship and dynamic data; An allocation unit is used to allocate each subtask to multiple predefined agents based on the complexity of the subtask. The generation unit is used to generate the subtask triggering order based on the subtask triggering order and the corresponding intelligent agents for each subtask.

[0046] Furthermore, in specific application scenarios, the allocation unit is specifically used to extract the task features of the subtask. If the task features match any complete processing path in the preset task library, the subtask is treated as a regular task, and an intelligent agent is allocated to the subtask through a pre-configured rule engine. If the task features do not match any complete processing path in the preset task library, the task features are classified and predicted using a lightweight natural language model. If the probability of a regular task is less than a preset threshold and the probability of a processable task is greater than a preset threshold, the subtask is determined as a semi-structured subtask, and an agent is assigned to the subtask using a lightweight machine learning model. If the task characteristics satisfy the conditions for a complex task, the subtask is identified as a complex task, and an agent is assigned to the subtask through an arbitrator based on reinforcement learning. The conditions for a complex task include cluster isolation conditions, high-risk labeling conditions, and low-confidence conditions.

[0047] Furthermore, the multi-agent collaborative module 33 includes: The publishing unit is used to publish a task completion event to the message queue in a publish-subscribe pattern when a parent task node completes, based on the task execution sequence of the intelligent agent. The processing unit is used to enable the triggered agent to perform the processing of the corresponding sub-task according to the received message and generate the processing result of the agent. In the case of multiple sub-tasks with parallel dependencies, multiple corresponding agents are synchronously triggered to execute in parallel by issuing a single event or broadcasting an event. The encapsulation unit is used to encapsulate the subtask processing result, task context, processing confidence, and new triggering events into a standard format result message. The update unit is used to publish the result message to the message queue to update the execution status of the corresponding node in the task tree and drive the triggering of subsequent logic. In the process of publishing the result message to the message queue, the transmitted message is compressed and decompressed, and reusable intermediate data is cached so that it can be read from the cache first under the condition of satisfying consistency.

[0048] Furthermore, in specific application scenarios, the intelligent agent also includes an information collection intelligent agent, a risk assessment intelligent agent, a compliance and ethics intelligent agent, and a learning feedback intelligent agent; the processing unit is specifically used to, if the information collection intelligent agent is triggered, perform anti-counterfeiting verification on the data to be verified in the multi-source heterogeneous data carried by the sub-task using a large model and a deep anti-counterfeiting algorithm, so that if the verification result passes, the multi-source heterogeneous data is subjected to multi-modal feature extraction and fusion to obtain a structured key information set; If the risk assessment agent is triggered, the processing includes: performing risk prediction on the risk profile through a dynamic risk assessment model to obtain a dynamic risk score and risk profile; If the triggering agent is a compliance and ethics agent, the processing procedure includes: performing rule matching and compliance verification on the intermediate operations or processing results of any agent to obtain the compliance verification result; If the trigger is a learning feedback agent, the processing includes: constructing a fine-tuned sample set based on online prediction performance indicators and human feedback data within a preset historical period, and retraining parameters and calibrating rules for the specific model that has experienced performance degradation based on the fine-tuned sample set.

[0049] Furthermore, the system also includes: The first extraction module is used to collect historical change data of insurance terms, regulatory policy texts and risk model parameters according to a preset period, and extract the structured features of the historical change data. The construction module is used to build an initial knowledge base based on the business type dimension features, clause validity dimension features, and time validity dimension features of the structured features; The first generation module is used to collect and process dynamic information from the business interaction process in real time through a bottom-up business-driven path, and generate the first type of update proposal; The second generation module is used to collect and parse normative documents from authoritative sources in real time through a top-down compliance-driven path to generate a second type of update proposal; The adjudication module is used to compare the first type of update proposal and the second type of update proposal with the benchmark knowledge base for consistency and conflict adjudication, and to merge the update content without conflict or after adjudication into the initial knowledge base based on preset priority rules to obtain the updated knowledge base. The distribution module is used to distribute and synchronize the updated knowledge base to the dynamic knowledge bases under different business types in an incremental manner.

[0050] Furthermore, the multi-agent collaborative module 33 includes: The consistency verification unit is used to call the conflict resolution rule base to perform consistency verification on the processing results of multiple agents, and to make decisions according to preset priorities in the event of a conflict during the consistency verification. The fusion unit is used to fuse the decision results and the processing results that have passed the consistency check to obtain a comprehensive processing result; The second extraction module is used to extract key decision factors and reasoning logic in the decision-making process through a preset decision interpretation mechanism. The third generation module is used to generate business processing results based on the key decision factors, reasoning logic, and comprehensive processing results.

[0051] This invention provides a business processing system based on multi-agent collaboration. In its embodiments, in response to a business processing request, the system performs multi-dimensional task analysis to obtain a structured task tree. This structured task tree includes multiple sub-tasks and their execution logic relationships. Based on the task type and execution logic relationships of the sub-tasks, and using a hybrid triggering and dynamic routing mechanism, the sub-tasks are assigned to multiple predefined agents, resulting in an agent task execution sequence. Each agent processes cross-business domain tasks by calling and querying a corresponding dynamic knowledge base. Each agent collaboratively processes the sub-tasks according to its task execution sequence, obtaining processing results for each agent. A decision-coordinating agent integrates these results to generate the business processing result. This significantly reduces manual intervention and serial delays in traditional business processing, lowering business risks and compliance costs caused by human error or delayed knowledge updates. Furthermore, its collaborative mechanism based on a dynamic knowledge base and hybrid routing ensures the system's adaptability and decision consistency in complex and changing business scenarios, thereby greatly improving the overall efficiency, accuracy, and intelligence level of business processing.

[0052] According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, which can execute the business processing method based on multi-agent cooperation in any of the above method embodiments.

[0053] Figure 4 The diagram illustrates the structure of a computer device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0054] like Figure 4 As shown, the computer device may include: a processor 402, a communication interface 404, a memory 406, and a communication bus 408.

[0055] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0056] Communication interface 404 is used for network communication with other devices such as clients or other servers.

[0057] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above-described embodiment of the business processing method based on multi-agent collaboration.

[0058] Specifically, program 410 may include program code that includes computer operation instructions.

[0059] Processor 402 may be a central processing unit (CPU), a specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. A computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0060] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0061] Specifically, program 410 can be used to cause processor 402 to perform the following operations: In response to a business processing request, the business to be processed is analyzed in multiple dimensions to obtain a structured task tree, wherein the structured task tree includes multiple sub-tasks and the execution logic relationship between each sub-task; Based on the task type of the subtask and the execution logic relationship, the subtask is assigned to multiple predefined intelligent agents based on a hybrid triggering and dynamic routing mechanism, and an intelligent agent task execution sequence is obtained. In this sequence, any intelligent agent can process cross-business domain tasks by calling and querying the corresponding dynamic knowledge base. Each of the aforementioned intelligent agents performs collaborative processing on sub-tasks according to the task execution sequence of the intelligent agent, obtains the processing results of different intelligent agents, and integrates the processing results of each intelligent agent through a decision-coordinating intelligent agent to generate a business processing result.

[0062] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A business processing method based on multi-agent collaboration, characterized in that, include: In response to a business processing request, the business to be processed is analyzed in multiple dimensions to obtain a structured task tree, wherein the structured task tree includes multiple sub-tasks and the execution logic relationship between each sub-task; Based on the task type of the subtask and the execution logic relationship, the subtask is assigned to multiple predefined intelligent agents based on a hybrid triggering and dynamic routing mechanism, and an intelligent agent task execution sequence is obtained. In this sequence, any intelligent agent can process cross-business domain tasks by calling and querying the corresponding dynamic knowledge base. Each of the aforementioned intelligent agents performs collaborative processing on sub-tasks according to the task execution sequence of the intelligent agent, obtains the processing results of different intelligent agents, and integrates the processing results of each intelligent agent through a decision-coordinating intelligent agent to generate a business processing result.

2. The method according to claim 1, characterized in that, Based on the task type of the subtask and the execution logic relationship, and using a hybrid triggering and dynamic routing mechanism, the subtask is assigned to multiple predefined agents, resulting in an agent task execution sequence, including: The subtask triggering order is determined based on the aforementioned execution logic relationship and dynamic data; For each subtask, the subtask is assigned to multiple predefined agents based on its complexity. The subtask triggering order is generated based on the subtask triggering order and the corresponding agents for each subtask.

3. The method according to claim 2, characterized in that, For each subtask, the subtask is assigned to multiple predefined agents based on its complexity, including: Extract the task features of the subtask. If the task features match any complete processing path in the preset task library, treat the subtask as a regular task and assign an agent to the subtask through a pre-configured rule engine. If the task features do not match any complete processing path in the preset task library, the task features are classified and predicted using a lightweight natural language model. If the probability of a regular task is less than a preset threshold and the probability of a processable task is greater than a preset threshold, the subtask is determined as a semi-structured subtask, and an agent is assigned to the subtask using a lightweight machine learning model. If the task characteristics satisfy the conditions for a complex task, the subtask is identified as a complex task, and an agent is assigned to the subtask through an arbitrator based on reinforcement learning. The conditions for a complex task include cluster isolation conditions, high-risk labeling conditions, and low-confidence conditions.

4. The method according to claim 1, characterized in that, Each of the aforementioned intelligent agents performs collaborative processing on subtasks according to its task execution sequence to obtain processing results for each agent, including: Based on the agent's task execution sequence, when a parent task node completes, a task completion event is published to the message queue using a publish-subscribe pattern; The triggered agent executes the processing of the corresponding sub-task according to the received message and generates the processing result of the agent. In the case of multiple sub-tasks with parallel dependencies, multiple corresponding agents are synchronously triggered to execute in parallel by issuing a single event or broadcasting an event. The subtask processing results, along with the task context, processing confidence, and new triggering events, are encapsulated into a standard format result message. The result message is published to the message queue to update the execution status of the corresponding node in the task tree and drive the triggering of subsequent logic. During the process of publishing the result message to the message queue, the transmitted message is compressed and decompressed, and reusable intermediate data is cached so that it can be read from the cache first under the condition of satisfying consistency.

5. The method according to claim 1, characterized in that, The intelligent agent also includes an information collection intelligent agent, a risk assessment intelligent agent, a compliance and ethics intelligent agent, and a learning feedback intelligent agent; The triggered agent performs the processing of the corresponding sub-task according to the received message, including: If the trigger is an information collection agent, the processing includes: using a large model and a deep anti-counterfeiting algorithm to perform anti-counterfeiting verification on the data to be verified in the multi-source heterogeneous data carried by the sub-task, so that if the verification result passes, the multi-source heterogeneous data is subjected to multi-modal feature extraction and fusion to obtain a set of structured key information. If the risk assessment agent is triggered, the processing includes: performing risk prediction on the risk profile through a dynamic risk assessment model to obtain a dynamic risk score and risk profile; If the triggering agent is a compliance and ethics agent, the processing procedure includes: performing rule matching and compliance verification on the intermediate operations or processing results of any agent to obtain the compliance verification result; If the trigger is a learning feedback agent, the processing includes: constructing a fine-tuned sample set based on online prediction performance indicators and human feedback data within a preset historical period, and retraining parameters and calibrating rules for the specific model that has experienced performance degradation based on the fine-tuned sample set.

6. The method according to claim 1, characterized in that, The dynamic knowledge base is updated based on a dual-track update mechanism. The update process specifically includes: Historical change data of insurance terms, regulatory policy texts, and risk model parameters are collected according to a preset cycle, and the structured features of the historical change data are extracted. An initial knowledge base is constructed based on the business type dimension features, clause validity dimension features, and time validity dimension features of the structured features; By adopting a bottom-up business-driven approach, dynamic information from business interaction processes is collected and processed in real time to generate the first type of update proposal; Through a top-down compliance-driven approach, normative documents from authoritative sources are collected and analyzed in real time to generate a second type of update proposal; The first type of update proposal and the second type of update proposal are compared with the benchmark knowledge base for consistency and conflict resolution. Based on the preset priority rules, the update content without conflict or after resolution is merged into the initial knowledge base to obtain the updated knowledge base. The updated knowledge base is distributed incrementally and synchronized to dynamic knowledge bases under different business types.

7. The method according to claim 1, characterized in that, The process of integrating the processing results of various intelligent agents through a decision-coordinating intelligent agent to generate business processing results includes: The conflict resolution rule base is invoked to perform consistency verification on the processing results of multiple agents, and a decision is made according to a preset priority in the event of a conflict during the consistency verification. The decision results and the processing results that pass the consistency check are merged to obtain the comprehensive processing result; Key decision-making factors and reasoning logic in the decision-making process are extracted through a pre-set decision interpretation mechanism; The business processing result is generated based on the key decision factors, reasoning logic, and comprehensive processing results.

8. A business processing system based on multi-agent collaboration, characterized in that, include: The task parsing module is used to respond to business processing requests, perform multi-dimensional task parsing on the business to be processed, and obtain a structured task tree, wherein the structured task tree includes multiple sub-tasks and the execution logic relationship between each sub-task; The routing module is used to allocate the subtasks to multiple predefined agents based on the task type and execution logic relationship of the subtasks and a hybrid triggering and dynamic routing mechanism, and obtain the task execution sequence of the agents. Each agent can process cross-business domain tasks by calling and querying the corresponding dynamic knowledge base. The multi-agent collaboration module is used to collaboratively process sub-tasks according to the task execution sequence of each agent, obtain the processing results of each agent, and integrate the processing results of each agent through the decision coordination agent to generate business processing results.

9. A storage medium, characterized in that, The storage medium stores at least one executable instruction, which causes the processor to perform the operation corresponding to the multi-agent collaborative business processing method as described in any one of claims 1-7.

10. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the business processing method based on multi-agent cooperation as described in any one of claims 1-7.

Images