Task processing method and device, equipment and storage medium

Abstract

The application discloses a task processing method and device, equipment and storage medium, and specifically discloses the following: receiving a to-be-processed task; performing multi-dimensional feature analysis on the to-be-processed task to obtain an analysis result; the multi-dimensional feature analysis comprises complexity analysis, dependency relationship analysis and resource requirement analysis; according to the analysis result, decomposing the to-be-processed task into a plurality of subtasks; according to the ability profile of a plurality of preset agents, matching corresponding target agents for the subtasks from the preset agents, and distributing the subtasks to the corresponding target agents; the ability profile is used for describing the task types and task processing capabilities that can be processed by the corresponding agents; receiving an execution result of the target agent for the distributed subtask, and performing aggregation on the execution result of the subtask based on the logical dependency relationship determined by the dependency relationship analysis, to generate an execution result of the to-be-processed task.

Description

Technical Field

[0001] This application relates to the field of insurance technology, and in particular to a task processing method, apparatus, device and storage medium. Background Technology

[0002] Currently, centralized control methods or rule-based control methods can be used to achieve task scheduling and collaboration among multiple agents. The centralized control method involves a single central node or master control unit performing calculations and decisions based on global state information, and sending unified action commands to all controlled agents to drive them to complete the overall task. The rule-based control method, on the other hand, involves pre-setting fixed behavioral rules, triggering conditions, or simple scripts for each agent, enabling them to react autonomously based on local information in specific scenarios, thereby achieving group collaboration.

[0003] While both of the above methods can achieve basic multi-agent collaboration, they both have significant technical problems: centralized control methods rely heavily on the reliability and computing power of the central node, making it difficult to cope with complex scenarios such as a surge in the number of agents or changes in network topology in dynamic and open environments, resulting in poor scalability, robustness, and real-time performance; on the other hand, control methods based on preset rules do not consider the dynamic uncertainty of the task environment and the heterogeneity of agents, and their static and fixed collaborative logic cannot be adaptively and optimized according to real-time task requirements and resource status, making it difficult to achieve efficient, flexible, and resource-efficient agent collaboration. Summary of the Invention

[0004] The main objective of this invention is to provide a task processing method, apparatus, device, and storage medium, which aims to solve the problems of poor scalability, low robustness, slow real-time response, insufficient adaptive optimization capability, and low resource utilization caused by existing multi-agent collaborative methods that rely on centralized control and control based on preset rules.

[0005] In a first aspect, embodiments of this disclosure provide a task processing method, including: Receive pending tasks; A multidimensional feature analysis is performed on the task to be processed to obtain the analysis results; the multidimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; the dependency analysis is used to identify data dependency, temporal dependency and logical dependency between operation units within the task to be processed; the resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed. Based on the analysis results, the task to be processed is decomposed into multiple sub-tasks; Based on the capability profiles of multiple preset intelligent agents, a corresponding target intelligent agent is matched from the preset intelligent agents for the subtask, and the subtask is assigned to the corresponding target intelligent agent; the capability profile is used to describe the types of tasks that the corresponding intelligent agent can handle and the task processing capabilities. The system receives the execution results of the target agent for the assigned sub-tasks, and aggregates the execution results of the sub-tasks based on the logical dependencies determined by the dependency analysis to generate the execution results of the task to be processed.

[0006] Secondly, embodiments of this disclosure provide a task processing apparatus, the apparatus comprising: The receiving module is used to receive tasks to be processed. The analysis module is used to perform multi-dimensional feature analysis on the task to be processed and obtain analysis results; the multi-dimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; the dependency analysis is used to identify data dependency, temporal dependency and logical dependency between operation units within the task to be processed; the resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed. The decomposition module is used to decompose the task to be processed into multiple sub-tasks based on the analysis results. The matching module is used to match a corresponding target intelligent agent for the subtask from the preset intelligent agents based on the capability profiles of multiple preset intelligent agents, and to assign the subtask to the corresponding target intelligent agent; the capability profile is used to describe the types of tasks that the corresponding intelligent agent can handle and the task processing capabilities. The aggregation module is used to receive the execution results of the target agent for the assigned sub-tasks, and aggregate the execution results of the sub-tasks based on the logical dependencies determined by the dependency analysis to generate the execution results of the task to be processed.

[0007] Thirdly, embodiments of this disclosure provide an electronic device, including: a processor; and a memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the method described in the first aspect above.

[0008] Fourthly, embodiments of this disclosure provide a computer-readable storage medium for storing computer-executable instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0009] Fifthly, embodiments of this disclosure provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect above.

[0010] The at least one technical solution provided by the embodiments of the present invention can achieve the following technical effects: In this embodiment of the invention, after receiving the task to be processed, a multi-dimensional feature analysis can be performed on the task to obtain the analysis results. This multi-dimensional feature analysis can include complexity analysis, dependency analysis, and resource requirement analysis. Dependency analysis can be used to identify data dependencies, temporal dependencies, and logical dependencies between operation units within the task to be processed. Resource requirement analysis can be used to assess the computational and data resource requirements when executing the task to be processed. Then, the task to be processed can be decomposed into multiple sub-tasks based on the analysis results. Based on the capability profiles of multiple preset agents, a corresponding target agent is matched for each sub-task from the preset agents. The capability profile describes the types of tasks and task processing capabilities that the agent can handle. After determining the target agent corresponding to the sub-task, the sub-task is assigned to the corresponding target agent, and the execution results of the assigned sub-task from the target agent are received. Based on the logical dependencies determined by the dependency analysis, the execution results of the sub-tasks are aggregated to generate the execution result of the task to be processed.

[0011] This invention achieves a decentralized, adaptive agent collaboration mechanism by performing refined multi-dimensional feature analysis on the task itself and dynamically decomposing and matching agents based on the analysis results. This overcomes the technical shortcomings of centralized control, which relies on a single central node for global decision-making, resulting in poor scalability and weak robustness. Furthermore, since it no longer relies on preset static rules but instead uses precise matching between task features and agent capability profiles, it can flexibly adapt to dynamically changing task environments and heterogeneous agent resources. This solves the technical problems of insufficient flexibility and inability to optimize orchestration in methods based on preset rules, thus achieving efficient, flexible, and resource-efficient agent collaboration. Attached Figure Description

[0012] Figure 1 This is one of the flowcharts illustrating a task processing method provided in an embodiment of the present invention; Figure 2 A second schematic flowchart of a task processing method provided in an embodiment of the present invention; Figure 3 A third schematic flowchart of a task processing method provided in an embodiment of the present invention; Figure 4 The fourth flowchart illustrates a task processing method provided in an embodiment of the present invention. Figure 5 The fifth flowchart illustrates a task processing method provided in an embodiment of the present invention. Figure 6 A flowchart illustrating a task processing method provided in an embodiment of the present invention (Figure 6); Figure 7 The seventh flowchart illustrating a task processing method provided in an embodiment of the present invention; Figure 8 This is the eighth flowchart illustrating a task processing method provided in an embodiment of the present invention. Figure 9 A flowchart illustrating a task processing method provided in an embodiment of the present invention is shown in Figure 9. Figure 10 A flowchart illustrating a task processing method provided in an embodiment of the present invention is shown in Figure 10. Figure 11 This is eleventh of a flowchart illustrating a task processing method provided in an embodiment of the present invention; Figure 12 The twelfth flowchart of a task processing method provided in an embodiment of the present invention; Figure 13 A flowchart of a task processing method provided in an embodiment of the present invention is shown as thirteenth. Figure 14 The fourteenth flowchart illustrates a task processing method provided in an embodiment of the present invention; Figure 15 A flowchart of a task processing method provided in an embodiment of the present invention is shown as Fifteen. Figure 16 A flowchart illustrating a task processing method provided in an embodiment of the present invention is shown in Figure 16. Figure 17 The seventeenth flowchart of a task processing method provided in an embodiment of the present invention; Figure 18 A schematic diagram of the module composition of a task processing device 1800 provided in one embodiment of the present invention; Figure 19 This is a schematic diagram of the hardware structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] Please see Figure 1 , Figure 1 This is a flowchart illustrating a task processing method provided in one embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps: Step 102: Receive the tasks to be processed.

[0016] Step 104: Perform multidimensional feature analysis on the task to be processed to obtain the analysis results; multidimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; dependency analysis is used to identify data dependencies, temporal dependencies and logical dependencies between operation units within the task to be processed; resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed.

[0017] Step 106: Based on the analysis results, the task to be processed is broken down into multiple sub-tasks.

[0018] Step 108: Based on the capability profiles of multiple preset agents, match the corresponding target agent for the subtask from the preset agents, and assign the subtask to the corresponding target agent; the capability profile is used to describe the types of tasks that the corresponding agent can handle and the task handling capabilities.

[0019] Step 110: Receive the execution results of the target agent for the assigned sub-tasks, and aggregate the execution results of the sub-tasks based on the logical dependencies determined by dependency analysis to generate the execution results of the task to be processed.

[0020] In this embodiment of the invention, various business requests from the application service layer can be received. Specifically, ITDEA (Intelligent Task Dispatcher & Execution Agent) can be used to receive these requests. These business requests can be uniformly abstracted as pending tasks. The application service layer may include multiple interfaces such as a customer portal, an agent portal, a management backend, and a mobile application. When a customer submits a new insurance application through the portal website, or an agent submits a customer service request through the backend system, or a mobile application triggers a policy expiration reminder event, a specific pending task can be generated. Each pending task can carry necessary contextual information, such as customer identifier, task type, priority identifier, and timestamp.

[0021] In one example, such as Figure 2The diagram illustrates a layered technical architecture. From top to bottom, the architecture is divided into a business access layer, an intelligent agent service layer, and a basic support layer. The business access layer can receive business requests through various channels such as customer portals, agent portals, management backends, and mobile applications. The core intelligent agent service layer can contain intelligent agents that perform domain-specific tasks, such as DPA (Data Perception Agent), PCA (Portrait Construction Agent), and VAA (Value Assessment Agent), as well as ITDEA, which serves as the coordination center. ITDEA is responsible for coordinating the work of all other intelligent agents and interacting with the underlying basic service layer. The basic support layer provides general-purpose technical components such as distributed computing, message queues, caching systems, and monitoring systems, as well as data and model assets such as customer databases, product databases, behavioral data lakes, and model warehouses.

[0022] In one example, such as Figure 3 As shown, the three-layer architecture is clearly presented from another perspective: the business application layer, the intelligent agent service layer, and the data support layer. The business application layer (such as customer management platforms and intelligent investment advisory systems) is the starting point for demand; the intelligent agent service layer is the core processing engine, where DPA and PCA are upstream, while VAA, RIA (Risk Identification Agent), PRA (Product Recommendation Agent), and other intelligent agents are downstream in parallel business decision-making positions, and ITDEA acts as the central control unit, coordinating the workflow of all intelligent agents; the data support layer (feature engineering platform, model service platform, data middle platform) provides unified algorithm features, model services, and data access capabilities for the upper-layer intelligent agents.

[0023] After receiving a task to be processed, a multidimensional feature analysis can be performed on the task to obtain the analysis results. This multidimensional feature analysis can include complexity analysis, dependency analysis, and resource requirement analysis. Specifically, dependency analysis can be used to identify data dependencies, temporal dependencies, and logical dependencies between operational units within the task; resource requirement analysis can be used to assess the computational and data resource requirements for executing the task. This multidimensional feature analysis process can be handled by the task analysis layer within ITDEA.

[0024] In one example, such as Figure 4As shown, the architecture of the underlying data platform is illustrated in detail, illustrating the complete data processing pipeline from multi-source data collection to high-quality data storage. The architecture can be divided into four layers: the data source layer (transaction data, behavioral data, external data), the data collection layer (collected through real-time streaming, batch ETL, API interfaces, etc.), the data processing layer (performing data cleaning, transformation, verification, and other quality control), and the data storage layer (including raw data storage, feature data storage, and a caching layer for acceleration). Figure 4 It demonstrates in detail how data is collected, processed, and stored in a standardized manner, thereby providing a reliable and consistent data supply for various intelligent agents at the upper level.

[0025] In one example, when performing multi-dimensional feature analysis on a task, the complexity of the task can first be graded based on a pre-defined complexity assessment model to obtain a complexity level. Complexity assessment can be based on multiple dimensions, including the number of operation units involved in the task, the number of data types involved, and the complexity of the conditional branches of the business rules. These assessment metrics are input into the pre-defined complexity assessment model, which is trained using machine learning methods based on the number of operation units in the sample task, the number of data types processed, and the number of conditional branches in the business rules. Specifically, the complexity assessment model can employ a gradient boosting decision tree algorithm, such as the XGBoost model. Its training features can include the number of operation units, the number of data types, and the number of conditional branches, while the target variable can be the complexity level labeled by experts. After the model is trained, it can quickly grade the complexity of new tasks. The complexity assessment model can output a quantified complexity level, for example, from 1 to 5, where 1 represents a simple task and 5 represents a highly complex task.

[0026] Then, the dependencies between the operational units within the task to be processed can be identified. There are three key types of dependencies among the operational units within the task: data dependencies, temporal dependencies, and logical dependencies. Data dependencies refer to the situation where the output data of one operational unit is the input data of another; for example, the profile building operation needs the cleaned data produced by the data collection operation. Temporal dependencies refer to the situation where operational units must follow a specific time sequence; for example, the customer access assessment must be completed before the policy activation operation can proceed. Logical dependencies refer to the business logic constraints between operational units; for example, the product recommendation operation needs to rely on both the value assessment result and the risk identification result. Based on these dependencies, an internal dependency graph can be constructed. This graph can use nodes to represent operational units and directed edges to represent dependencies. Simultaneously, historical task execution data can be referenced to analyze the correlation between the inputs and outputs of different tasks, and based on the analyzed correlations, a task dependency graph can be constructed to understand cross-task dependency patterns.

[0027] After obtaining the aforementioned complexity levels, internal dependency graphs, and task dependency graphs, these data can be combined to accurately estimate the computational and data resources required to execute the task. The required computational resources may include the number of CPU cores, memory size, and network bandwidth. The required data resources may include the types of databases to be accessed, the size of the data tables, and cache hit rate requirements. For example, an insurance customer lifecycle management task, after analysis, can be determined to have a complexity level of 4. Its internal dependency graph shows 5 key operational units and 7 dependency edges. The resource requirement assessment shows that it requires 8 CPU cores, 16GB of memory, and access to three main data sources: a customer database, a product knowledge base, and a behavioral data lake.

[0028] In this embodiment of the invention, the task to be processed can be decomposed into multiple subtasks. The decomposition process can be specifically executed by the task decomposition layer of ITDEA. During decomposition, based on the results of the multi-dimensional feature analysis, especially the internal dependency graph and task dependency graph generated by dependency analysis, the complex task to be processed can be broken down into a set of subtasks with finer granularity and simpler responsibilities. The logical structure of the task itself and the current available intelligent agent resources can be considered during decomposition. Parallel processing units in the task can be identified and decomposed into independent subtasks, while ensuring that operation units with dependencies are decomposed into a sequence of subtasks with sequential constraints. Each subtask can be given clear input and output definitions, execution constraints, and quality standards. The decomposed set of subtasks needs to maintain the overall business semantics of the original task to be processed, while improving parallel processing efficiency and resource utilization. For example, an insurance customer lifecycle management task can be decomposed into multiple subtasks such as data collection, profile building, value assessment, risk identification, and product recommendation. These subtasks have both temporal and data dependencies.

[0029] In one example, the task to be processed could be an insurance customer lifecycle management task. This task can be broken down into three sub-tasks: data collection, customer profile building, and business decision-making. The data collection sub-task can be used to collect customer transaction data, behavioral data, and credit data; the customer profile building sub-task can be used to build customer profiles based on these data; and the business decision-making sub-task can be used to perform at least one of the following based on the customer profile: customer value prediction, customer risk assessment, and insurance product recommendation analysis.

[0030] Among them, such as Figure 5The diagram illustrates the complete lifecycle of an insurance customer from potential client to eventual churn or value enhancement, highlighting the management actions and decision-making paths at each key juncture. The process begins with potential customer discovery and acquisition, followed by customer eligibility assessment—a crucial decision point determining whether a customer can proceed to the next stage. Customers who pass the assessment enter the activation, product sales, and service stages. In the later stages of the service cycle, renewal management becomes central, involving renewal decisions. This decision point leads to three paths: if the customer renews, the customer retention and preservation stage begins; if they do not renew, the customer recovery process is initiated; if the customer chooses to upgrade, cross-selling is pursued. Regardless of the path, the goal is to enhance customer value. Then, differentiated and refined management strategies are implemented based on the customer's lifecycle stage (e.g., maturity, growth, decline). For example, value maintenance is implemented during the maturity stage, while retention strategies are adopted during the decline stage. Failure to retain customers results in churn. All task processing and agent collaboration described in this embodiment of the invention ultimately aim to efficiently and automatically manage and optimize... Figure 5 Every step described aims to maximize customer value.

[0031] In one example, such as Figure 6 The diagram illustrates in detail the end-to-end processing flow in an insurance customer lifecycle management scenario, from customer event triggering to final result feedback. A customer event can trigger the start of the process, initiating ITDEA and overseeing the overall flow. The core of the process showcases three key stages: data collection, data preprocessing and integration, and profile building. The data collection stage is handled by DPA, which collects information in parallel from multiple sources, including transaction data, behavioral data, and external data. After preprocessing and integration, this data is further processed by PCA to build multi-dimensional customer profiles, such as behavioral profiles, risk profiles, and value profiles. Then, ITDEA makes judgments based on business needs and coordinates business decision-making sub-tasks such as value assessment, risk identification, and product recommendation. The results of these sub-tasks are ultimately integrated to form a complete task execution result, which is then fed back to the customer. This flowchart vividly illustrates the decomposition and execution process of an insurance customer lifecycle management task in this embodiment of the invention, namely, decomposing it into sub-tasks such as data collection, profile building, and business decision-making, which are completed collaboratively by corresponding intelligent agents.

[0032] In one example, such as Figure 7As shown, this document details the end-to-end workflow of PCA, from data input to high-quality profile output, including quality monitoring. The process begins with real-time data collection. PCA acquires transaction data, behavioral data, and external data collected by DPA and enters the data quality monitoring stage. This stage continuously performs anomaly detection, integrity verification, and timeliness assurance to ensure that the data input to PCA is reliable and complete—a prerequisite for building accurate profiles. The core profile building process is completed by PCA, which processes and integrates high-quality data to output a multi-dimensional customer profile including behavioral, preference, risk, and value profiles. After the profile is generated, the profile quality assessment module performs integrity checks and accuracy verification, and drives real-time updates to the profile based on business feedback and data changes. Figure 7 This clearly demonstrates that PCA is not a simple data aggregation process, but a rigorous process that includes data quality control, intelligent construction, continuous evaluation, and iterative optimization.

[0033] In one example, such as Figure 8 As shown, the internal architecture and core workflow of DPA and PCA working together are illustrated, clearly demonstrating the complete processing link from raw data to high-quality customer profiles and its inherent quality control closed loop. Figure 8 The upper layer focuses on DPA (Data Processing Automation), whose core responsibilities are broken down into two parallel functional pillars: real-time data collection and data quality monitoring. This reveals that the starting point of the data processing flow is not merely data acquisition, but also prioritizing data quality control, reflecting the design philosophy that quality originates from data collection. The real-time data collection function is responsible for continuously aggregating raw data from multiple sources, such as transaction logs, user behavior logs, and external credit scoring interfaces. Simultaneously, the data quality monitoring function operates as an independent assurance system, using preset rules and algorithms to perform anomaly detection, integrity verification, and consistency checks on the input data stream, ensuring that any flawed data is identified and processed before entering the core processing stage. The outputs of these two functional channels together provide clean and reliable raw materials for downstream profiling.

[0034] Figure 8The lower layer details how the profile-building agent utilizes high-quality data supplied from upstream to execute its core mission. It concretizes the customer profile building process into four measurable and manageable dimensions: behavioral profile, risk profile, value profile, and preference profile. This demonstrates that a customer profile is not a vague concept, but a comprehensive model organically composed of these concrete sub-profiles. More importantly, the diagram emphasizes that profile building is not a one-time action, but a continuous iterative optimization process. The profile quality assessment module constitutes the feedback node in this iterative closed loop. It performs quality checks on the generated profiles through mechanisms such as completeness checks and accuracy verification. Based on the assessment results and business feedback, the profiles are driven to be updated and dynamically optimized in real time, enabling customer profiles to maintain high accuracy over time and as customer status changes.

[0035] In one example, such as Figure 9 As shown, this demonstrates a real-time data processing and profiling management workflow centered on data quality. It visualizes the transformation of data from its raw state to become a trusted business asset. The workflow begins with the parallel collection of diverse data sources, including external data, user behavior data, and transaction data, reflecting the richness and comprehensiveness of the data supply. The collected raw data undergoes data cleaning to remove noise, correct errors, and fill in gaps, transforming the chaotic raw information into well-organized, consistent, and analyzable structured data, laying a solid foundation for subsequent in-depth analysis.

[0036] After cleaning, the data then enters the user profile calculation stage, transforming the basic data into business-meaning customer characteristics and tags to build or update user profiles that represent a comprehensive view of the customer. A crucial quality checkpoint is set up here—quality assessment. This decision node rigorously verifies the calculated profile data to determine if it meets preset standards for completeness, accuracy, and timeliness. If the assessment fails, the data flow is redirected upstream to re-enter the cleaning or calculation stage, forming a self-correcting feedback loop. This design ensures that any substandard intermediate products do not flow downstream and contaminate decision-making, forcing quality issues to be resolved in the current iteration. Only data that successfully passes the quality assessment can be output and participate in subsequent processing, such as business decision analysis or personalized recommendation engines. Figure 9 It showcases a data governance pipeline that integrates data cleansing, intelligent computing, rigorous quality inspection, and closed-loop feedback, ensuring the reliability and high confidence of data assets that ultimately drive business decisions.

[0037] In one example, such as Figure 10As shown, the core functional framework of three core business decision-making intelligent agents is presented side-by-side: Product Recommendation Agent (PRA), Risk Identification Agent (RIA), and Value Assessment Agent (VAA). For PRA, its functions are clearly divided into two main sections: intelligent product recommendation and recommendation effectiveness evaluation. The intelligent product recommendation section reveals three key characteristics of recommendations: personalized recommendation, indicating its ability to customize based on unique customer profiles; scenario-aware recommendation, demonstrating its consideration of the customer's current specific context (such as policy renewal period, new product launch period); and real-time updates, meaning the recommendation model can dynamically adjust according to the latest customer behavior and data. Together, these constitute a flexible and adaptive recommendation engine. The recommendation effectiveness evaluation section constructs a complete metric system for measuring the value of the recommendation system, including accuracy evaluation, conversion rate analysis, and satisfaction evaluation. This reflects a focus not only on the output of the recommendation algorithm but also on the final effectiveness in real-world business scenarios.

[0038] For RIA, its functional framework is also divided into two levels: multidimensional risk assessment and risk transmission analysis. Multidimensional risk assessment clarifies the three main perspectives of risk analysis: credit risk, behavioral risk, and associated risk, indicating that risk assessment is three-dimensional and comprehensive. Risk transmission analysis goes a step further, focusing on how risks spread and diffuse within customer relationship networks or business processes, and ultimately achieving risk early warning, demonstrating an advanced capability from static risk assessment to dynamic risk management.

[0039] For VAA, its framework focuses on customer value prediction and value contribution analysis. Customer value prediction emphasizes multi-time-domain forecasting capabilities, enabling the estimation of value for different future time periods, supplemented by interpretability analysis to ensure the prediction results are not a black box, while also segmenting value for differentiated operations. Value contribution analysis focuses on attribution, identifying key factors affecting customer value through feature contribution analysis, grasping value change trends through trend analysis, and quantifying the credibility of the prediction results through prediction confidence level.

[0040] In one example, such as Figure 11As shown, a complete end-to-end business decision-making process driven by three intelligent agents is dynamically presented. First, work is carried out collaboratively or in parallel along three paths: value assessment, risk identification, and product recommendation. The left side of the process shows the VAA (Value Assessment Agent) path, starting with customer value prediction, proceeding through specific analytical steps such as multi-time-domain prediction and interpretability analysis, and delving into feature contribution analysis and prediction confidence calculation, producing a quantified customer value score and a credible explanation. The middle of the process shows the RIA (Risk Analysis Agent) path, centered on multi-dimensional risk assessment, covering multiple assessment dimensions such as credit risk, behavioral risk, and association risk, and focusing on the dynamic process of risk propagation and diffusion. The right side of the process shows the PRA (Professional Product Recommendation) path, revolving around intelligent product recommendation, emphasizing personalization and context awareness, and also including evaluation stages for the accuracy and conversion rate of recommendation effectiveness. Figure 11 The middle arrows clearly indicate the logical sequence and possible convergence points between analytical steps. For example, the results of value assessment and risk identification are likely to serve as input together, influencing product recommendation strategies. Figure 11 This demonstrates that intelligent agents do not work in isolation, but rather within a unified task flow framework, each performing its own duties while collaborating with each other to jointly complete the transformation from customer data to business decisions.

[0041] In this embodiment of the invention, after decomposing the task to be processed into multiple sub-tasks, a corresponding target intelligent agent can be matched to the sub-tasks based on the capability profiles of multiple preset intelligent agents, and the sub-tasks can be assigned to the corresponding target intelligent agents. The capability profiles can be used to describe the types of tasks that the corresponding intelligent agent can handle and its task processing capabilities. Specifically, multiple types of intelligent agents can be preset, each with a detailed capability profile describing its area of ​​expertise and processing capabilities. For example, the capability profile of DPA describes its expertise in data collection, data cleaning, and data quality monitoring, and its task processing capability indicators include data throughput, cleaning accuracy, and real-time performance, thus characterizing its task processing capabilities. The capability profile of PCA describes its expertise in customer profile building, profile quality assessment, and real-time updates, and its task processing capability indicators include profile dimensional completeness, feature accuracy, and update timeliness. The capability profile of VAA describes its expertise in customer value prediction, multi-time-domain prediction, and interpretability analysis, and its task processing capability indicators include prediction accuracy, prediction time span, and interpretability credibility. The capability profile of RIA describes its expertise in multi-dimensional risk assessment, risk transmission analysis, and risk early warning, and its task processing capability indicators include risk identification accuracy, transmission analysis depth, and early warning timeliness. The capability profile of PRA describes its expertise in intelligent product recommendation, scenario-aware recommendation, and recommendation effect evaluation, and its task processing capability indicators include recommendation accuracy, scenario adaptability, and conversion rate. Based on the task type requirements of the sub-tasks, intelligent agents matching the capability profiles can be found for each sub-task.

[0042] In one example, the task to be processed could be a lifecycle management task for insurance customers, which can be broken down into data collection sub-tasks, profile building sub-tasks, and business decision-making sub-tasks. Based on the pre-defined intelligent agent's capability profile, the sub-tasks can be matched with corresponding target intelligent agents, and the sub-tasks can be assigned to the appropriate target intelligent agents. The intelligent agent matching process can be executed by the ITDEA's intelligent agent matching layer. For example, the data collection sub-task can be executed by the DPA to collect customer-related information from multiple data sources. These data sources can include internal transaction databases (which can record customer purchase records, payment history, and claim records), behavioral data platforms (which can record customer access behavior on websites and mobile applications, and interactions with customer service), and external data sources (such as credit data from credit reporting agencies, public information from social media, etc.). The data collection sub-task can include various methods such as real-time data stream collection, batch data collection, and API interface calls; the profile building sub-task can be executed by the PCA to build a comprehensive customer profile based on the collected raw data. Customer profile building can include multiple dimensions such as behavioral profiles (profiles built by analyzing customer purchasing habits and interaction preferences), risk profiles (profiles built by identifying potential risk factors in customers), value profiles (profiles built by assessing customer contribution potential), and preference profiles (profiles built based on customer product and service preferences). Business decision-making sub-tasks can be executed by corresponding specialized intelligent agents. Specifically, customer value prediction can be handled by the VAA (Value Agent), customer risk assessment by the RIA (Risk Agent), and insurance product recommendation analysis by the PRA (Product Recommendation Agent). These business decision-making sub-tasks can be executed in parallel based on customer profiles, but they all depend on the output results of the profile building sub-tasks. For example, in managing customer Zhang San's insurance policy, the decomposed data collection subtask will collect Zhang San's recent transaction records, website browsing behavior, and the latest credit report; the profile building subtask will update Zhang San's behavioral profile (showing his preference for online services), risk profile (low credit risk), value profile (high-value customer), and preference profile (focus on health insurance products) based on this data; the value assessment in the business decision-making subtask will predict Zhang San's value contribution for the next year, the risk assessment will analyze Zhang San's policy risk level, and the product recommendation will generate personalized insurance product suggestions for Zhang San.

[0043] In one example, when matching a target agent for a subtask based on the capability profiles of multiple preset agents, the runtime status values ​​of each preset agent can be obtained in real time. These runtime status values ​​can include CPU load, memory usage, task queue length, and historical task success rate. Then, based on the subtask's task type, runtime status value, and the task processing capabilities required to handle the subtask, a matching score between each preset agent and the subtask can be calculated. The preset agent with the highest matching score is then identified as the target agent for the subtask. The matching score can be calculated using a weighted summation model, with the matching score corresponding to task type having the highest weight, followed by the matching score corresponding to runtime status, and the matching score corresponding to historical performance having the lowest weight. The task type matching score can be calculated by comparing the subtask's task type with the task type support list declared in the agent's capability profile. A high score is awarded for a perfect match to the task type in the supported list, a medium score for a partial match, and a zero score for a no-match. The running status match score is calculated based on the agent's current load; agents with lower CPU load, lower memory usage, and shorter task queue lengths score higher. The historical task success rate match score can be obtained by directly using the agent's average success rate for recently executing similar tasks. The total match score is obtained by multiplying each sub-score by its weight and then summing the results. For example, a data collection subtask needs to match a DPA, and three preset DPAs are available. DPA 1 has a perfect task type match, a CPU load of 30%, a memory usage of 40%, a task queue length of 3, and a historical success rate of 98%, resulting in a score of 95. DPA 2 has a perfect task type match, but a CPU load of 80%, a memory usage of 70%, a task queue length of 8, and a historical success rate of 95%, resulting in a score of 75. DPA 3 has a perfect task type match, a CPU load of 20%, a memory usage of 30%, a task queue length of 1, and a historical success rate of 99%, resulting in a score of 98. Therefore, DPA3 was chosen as the target agent. The weights corresponding to the matching degree calculation can be continuously optimized based on actual operating data to maximize overall efficiency.

[0044] In this embodiment of the invention, after assigning a subtask to the corresponding target agent, the execution status of the target agent on the assigned subtask can be monitored. If the monitored execution status indicates that the target agent has timed out or failed, a reassignment mechanism can be triggered. The reassignment mechanism can be to assign the subtask to a preset agent with the second-highest matching score, or to recalculate the matching scores between each preset agent and the subtask, and then reassign the subtask to the preset agent with the highest recalculated matching score. Execution status monitoring can be handled by the execution coordination layer of ITDEA. When monitoring the execution status, monitoring indicators can include subtask execution time, resource consumption progress, and abnormal log output. Each subtask has a preset timeout threshold; if the execution time exceeds the threshold, it is determined to have timed out. If an agent reports an execution abnormality or an agent failure is detected, it is determined to have failed. Once the reassignment mechanism is triggered, the availability of the agent with the second-highest initial matching score can be checked first. If it is available and the load is within an acceptable range, the subtask can be directly reassigned to that agent. If the second-highest performing agent is also unavailable, the current running status values ​​of all available agents can be recollected, the matching score recalculated, and a new optimal agent selected for assignment. The reassignment process can be logged for subsequent analysis and optimization. For example, a profile building subtask was initially assigned to PCA2, but monitoring revealed that PCA2 failed to complete profile building within the specified 2-second timeout, thus triggering reassignment. PCA1, initially ranked second in matching score, was found to have a current CPU load of 50%, within acceptable limits, so the subtask was reassigned to PCA1. If PCA1 is also unavailable, the matching scores of all PCAs are recalculated, and a new target agent is selected.

[0045] In embodiments of the present invention, the execution results of the target agent for the assigned sub-tasks can be received, and the execution results of the sub-tasks can be aggregated based on the logical dependencies determined by dependency analysis to generate the execution results of the task to be processed. Result reception and aggregation can be performed by the ITDEA result aggregator. After each target agent completes a sub-task, it can send its execution results to the result aggregator. The results may include task output data, execution metadata (such as start time, end time, resource consumption), and quality metrics (such as confidence level, accuracy). The result aggregator can integrate the sub-task results based on the logical dependencies determined by the preceding dependency analysis. For sub-tasks with sequential dependencies, the result aggregator can process the results one by one according to the dependency order, ensuring that the results of subsequent sub-tasks correctly depend on the results of preceding sub-tasks. For sub-tasks that can be executed in parallel, the aggregator can receive and process their results in parallel, improving aggregation efficiency. The aggregation logic of the aggregator may include data format conversion, result consistency verification, conflict resolution, and semantic fusion. Based on the aggregation results, the overall execution result of the task to be processed can be generated. For example, in the task of insurance customer lifecycle management, the results aggregator can sequentially receive the results of the data collection subtask (cleaned customer data), the results of the profile building subtask (multi-dimensional customer profiles), and the results of the business decision-making subtask (value assessment scores, risk levels, and product recommendation lists). Based on the logical dependencies between these subtask results, the results aggregator generates a final comprehensive report, which may include customer value analysis, risk assessment conclusions, product recommendation suggestions, as well as corresponding decision-making basis and actionable actions.

[0046] In one example, such as Figure 12 The diagram illustrates the internal workings of the intelligent task decomposition and execution algorithm, breaking down the entire process from task reception to coordination into three logically rigorous and sequentially connected core stages. The process begins with the task analysis and decomposition stage. At this stage, instead of simply and crudely splitting the task, a series of professional evaluators are activated to conduct in-depth analysis of the task. The task complexity evaluator quantifies and classifies the difficulty and scale of the task; the task dependency analyzer carefully examines the intricate data, temporal, and logical dependencies between various operational units within the task; the resource requirement evaluator is responsible for calculating the computing, storage, and other resources required to execute the task; and the priority sorter determines the processing order based on business rules. The collaborative work of these analyzers provides a comprehensive basis for the reasonable decomposition of tasks.

[0047] Based on the results of in-depth analysis, the process enters the agent selection and allocation phase. The core of this phase is to optimally match the decomposed subtasks with various specialized agents. The capability matching algorithm matches subtasks based on their requirements and the capability profiles of each agent; the load balancing algorithm takes a global perspective to ensure that no agent is overloaded and that overall throughput is optimal; the performance monitor and fault detector continuously monitor the agent status, providing real-time data for dynamically adjusting the allocation strategy.

[0048] The final stage is task execution and monitoring. The concurrent execution controller ensures that subtasks that can be executed in parallel are efficiently scheduled; the result aggregator is responsible for systematically integrating the scattered subtask results into a final unified result according to the dependencies analyzed earlier. The quality assurance mechanism and performance optimizer work continuously in the background to ensure output quality and seek efficiency improvements. Figure 12 It fully reveals how ITDEA manages tasks through a standardized process of analysis, planning, allocation, monitoring, and summarization.

[0049] In embodiments of the present invention, after generating the execution results of the tasks to be processed, the actual execution performance data of each subtask can be recorded. Execution performance data may include execution time. Based on the difference between the actual execution performance data and the expected performance data, the actual processing efficiency of the target agent executing each subtask can be evaluated. According to the actual processing efficiency, the processing capability rating for similar subtasks in the target agent's capability profile can be adjusted. The processing capability rating can be used to determine the number of subtasks to be decomposed into similar subtasks. The task type of similar subtasks is the same as the task type of the subtasks executed by the target agent. Execution performance data can be obtained by monitoring the execution status of the target agent on the assigned subtasks. Execution performance data may specifically include the start time, end time, resource usage details, etc., of each subtask. When determining the actual processing efficiency, it can be determined by comparing the actual execution time with the expected execution time, combined with task complexity and data scale, etc. The evaluation results can be used to dynamically adjust the processing capability rating in the agent's capability profile. If the agent consistently demonstrates processing efficiency higher than expected, its processing capability rating is increased; if it consistently falls below expectations, the rating is decreased. The processing capability rating can directly affect the future decomposition strategy for similar tasks. For agents with high processing power ratings, the number of subtasks can be reduced when decomposing similar tasks, and larger-granularity tasks can be assigned to them to reduce coordination overhead. For agents with low processing power ratings, tasks can be decomposed into finer-grained subtasks to facilitate parallel processing and load balancing. For example, VAA's initial processing power rating is A, indicating that it can efficiently handle complex value prediction tasks. In the obtained execution performance data, VAA1's actual processing time for a high-complexity value prediction subtask is 3 seconds, which is shorter than the expected 5 seconds, indicating that its actual processing efficiency is higher than expected. Therefore, VAA1's processing power rating is upgraded to A+. Subsequently, when decomposing similar value prediction tasks, the work that originally needed to be decomposed into two subtasks can be merged into one subtask and assigned to VAA1 to improve efficiency. This dynamic adjustment mechanism based on actual performance can achieve continuous self-optimization.

[0050] In embodiments of this invention, a distributed agent architecture can be employed to ensure scalability and reliability. Agents are deployed across multiple nodes and coordinated uniformly by a cluster management node. A service registry is responsible for agent registration and discovery, a load balancer distributes load, and a monitoring center monitors health status in real time. The communication layer enables asynchronous communication between agents through message queues (such as Kafka or RabbitMQ), a service mesh (such as Istio) provides secure inter-service communication, and an API gateway (such as Kong or Nginx) uniformly manages external interfaces. This architecture supports horizontal scaling of agents; agent instances can be dynamically added as load increases, and backup instances can be automatically switched when an agent fails, ensuring high availability.

[0051] In one example, such as Figure 13 As shown, the underlying distributed system architecture that supports the efficient, reliable, and collaborative operation of all the aforementioned intelligent agents and services is demonstrated, namely, the microservices and service governance system. Figure 13 The document showcases the service-oriented architecture components: a service registry records the addresses and states of all agent service instances, enabling services to discover each other; a load balancer intelligently distributes external requests or internal calls to multiple identical agent instances, preventing overload on a single instance and ensuring overall high availability and performance; and an API gateway, acting as a unified portal and security barrier, handles all requests from the application service layer (customer portal, mobile applications, etc.), performing routing, authentication, rate limiting, and monitoring. The specific business logic is handled by individual microservices (i.e.,... Figure 13 This is achieved through various intelligent agents (such as DPA, VAA, etc.). Figure 13 It demonstrates how to deploy components such as ITDEA, DPA, and VAA as independently developable, deployable, and scalable microservices using a cloud-native architecture, and how to achieve secure and efficient communication and collaboration between them through governance tools such as service mesh and API gateway, so as to make the entire system elastic, scalable, and highly reliable.

[0052] This invention also proposes a real-time processing optimization mechanism. This includes optimizing data stream processing through a stream processor; optimizing caching strategies through a cache manager; compressing transmitted data through a compression engine; and optimizing database queries through a query optimizer. These optimizations enable efficient handling of real-time business needs. For example, when a large number of customers simultaneously initiate insurance inquiries, tasks can be quickly assigned to available agent instances, and data processing latency can be reduced through caching and compression techniques.

[0053] At the technical implementation level, this invention also includes detailed algorithm design. For example, the EVPA (Enhanced Value Prediction Algorithm) used in VAA combines multiple machine learning models, including logistic regression, random forest, gradient boosting, and neural networks; the model fusion layer employs a weighted fusion strategy, dynamically adjusting the weights based on the performance of each model in different scenarios. The interpretability layer uses the SHAP (SHapley Additive Explanations) method to interpret the prediction results. The core formula of SHAP is:

[0054] in, It is the original input feature vector, representing various feature data of the customer; It is the perturbed feature vector, obtained by applying a small perturbation to the original features; It is a distance metric function used to measure prediction results. Compared with the benchmark Differences; These are disturbance constraint parameters that ensure and The difference does not exceed the infinite norm. This formula quantifies the contribution of each feature to the prediction result by calculating the impact of different feature perturbations on the prediction result. For example, when assessing the value of customer Zhang San, VAA uses SHAP analysis to analyze the contribution of features such as Zhang San's purchase frequency, average premium, and credit score to the final value prediction result, thereby providing an interpretable basis for decision-making.

[0055] In one example, such as Figure 14 The diagram illustrates the advanced machine learning architecture employed by VAA for customer value prediction. This is a layered, modular modeling process. The input layer defines the three core data foundations of the model: customer profile data, historical behavior data, and external environment data. The feature engineering layer, crucial for model success, comprises multiple modules including basic feature extraction, temporal feature generation, interaction feature construction, and domain feature engineering, aiming to extract meaningful signals for prediction from the raw data. The prediction model layer employs a fusion strategy, deploying multiple algorithms in parallel, such as logistic regression, random forest, XGBoost, and neural networks, leveraging the complementary strengths of different models. The model fusion layer optimizes and integrates the prediction results of multiple base models through techniques such as weighted fusion, stacking ensemble, voting mechanisms, and dynamic weight adjustment to improve the accuracy and stability of the final prediction. The interpretability layer integrates tools such as SHAP global interpretation, LIME (Local Interpretable Model-agnostic Explanations) local interpretation, feature importance analysis, and decision path tracing to ensure that the model's prediction results are not only accurate but also understandable and reliable, meeting interpretability requirements. Ultimately, the output layer transforms the prediction results into business-usable value scores, value grades, trend predictions, and recommended actions.

[0056] RIA uses CRAA (Comprehensive Risk Assessment Algorithm) to assess risk from multiple dimensions, including credit risk, behavioral risk, and association risk. The algorithm employs a risk transmission analysis model to simulate the propagation path and impact range of risk within the customer network. PRA uses EIRA (Enhanced Insurance Recommendation Algorithm) which combines collaborative filtering, content recommendation, and deep learning technologies, and incorporates a context-aware mechanism to adjust recommendation strategies based on the customer's current situation.

[0057] In one example, such as Figure 15 As shown, this paper elucidates the core algorithm framework for PRA to generate personalized insurance product recommendations. The framework is divided into an input layer, an algorithm layer, a fusion optimization layer, and an output layer. The input layer clarifies the two main categories of information upon which the recommendations rely: user profile input (including basic attributes, behavioral characteristics, and preference information) and product information input (including product feature library, historical sales data, and product association rules). The algorithm layer is the core engine. It does not rely on a single algorithm but integrates three mainstream recommendation technologies: collaborative filtering (based on user or item similarity), content recommendation (based on feature or semantic matching), and deep learning (such as neural collaborative filtering) to cover different scenarios and data types. The fusion optimization layer intelligently integrates the results of multiple algorithms, adopting a multi-algorithm fusion strategy (such as dynamic weight allocation) and introducing a scene-aware mechanism. This allows the recommendations to be dynamically adjusted according to the specific time and scenario of the customer, while also supporting real-time model updates to adapt to data changes. The output layer ultimately produces a list of recommended products, along with recommendation confidence and explainable reasons for the recommendations. It also includes a built-in effect feedback mechanism for continuous algorithm optimization.

[0058] The core of ITDEA is its intelligent task decomposition and execution algorithm. This algorithm employs a graph theory-based task decomposition method, breaking down complex tasks into a set of subtasks represented by a DAG (Directed Acyclic Graph). Task allocation uses a market-based bidding model, where agents bid for subtasks based on their capabilities and states, and a coordinator selects the optimal agent to execute the subtask. This distributed decision-making mechanism improves flexibility and efficiency.

[0059] In one example, such as Figure 16As shown, the collaborative workflow of various functional components within ITDEA is described in detail, revealing the foundation for its intelligent scheduling and coordination capabilities. The workflow is top-down. The task receiving layer is responsible for receiving and initially parsing externally input business tasks; the task classification layer identifies task types; the task analysis layer is the core pre-processing stage, containing a complexity analyzer, dependency analyzer, resource requirement evaluator, and priority ranking module, which together complete the aforementioned multi-dimensional feature analysis process; based on this analysis, the task decomposition layer uses task decomposition algorithms, subtask generators, and constructs task dependency graphs and execution order planning to scientifically break down complex tasks into executable subtasks; the agent matching layer uses capability matching algorithms, combined with information from load balancers and performance predictors, to find the most suitable target agent for each subtask, while a fault detector ensures the robustness of the matching; the execution coordination layer monitors the execution of subtasks, handles faults through anomaly handlers, and integrates the final results through a result aggregator; the performance evaluator and learning optimizer in the feedback optimization layer continuously evaluate the execution effect, adjust parameters, and update knowledge, enabling the entire ITDEA to continuously self-optimize.

[0060] In one example, such as Figure 17 As shown, the process of systematically decomposing and processing complex tasks is clearly illustrated. The process begins with classifying and identifying the input task, determining its domain and basic attributes. Then, task analysis and decomposition are performed to deconstruct the complex task, identifying its components and internal logic. After decomposition, the resulting sub-tasks are systematically summarized into three specialized functional modules. Module one focuses on information collection and processing, corresponding to the responsibilities of DPA (Data Acquisition, Processing, and Application), which involves acquiring raw information from the internal and external environment and performing preliminary processing. Module two focuses on decision-making and planning, corresponding to the core functions of business decision-making agents such as VAA (Vehicle Application Analyzer) and RIA (Resource Application Analyzer), which involves analyzing, judging, and generating strategies based on information. Module three is responsible for execution and control, corresponding to translating decisions into specific action instructions and monitoring the execution process; in customer management scenarios, this can manifest as triggering specific marketing activities or service processes. Figure 17 This demonstrates a universally applicable intelligent task processing flow.

[0061] In this embodiment of the invention, after receiving the task to be processed, a multi-dimensional feature analysis can be performed on the task to obtain the analysis results. This multi-dimensional feature analysis can include complexity analysis, dependency analysis, and resource requirement analysis. Dependency analysis can be used to identify data dependencies, temporal dependencies, and logical dependencies between operation units within the task to be processed. Resource requirement analysis can be used to assess the computational and data resource requirements when executing the task to be processed. Then, the task to be processed can be decomposed into multiple sub-tasks based on the analysis results. Based on the capability profiles of multiple preset agents, a corresponding target agent is matched for each sub-task from the preset agents. The capability profile describes the types of tasks and task processing capabilities that the agent can handle. After determining the target agent corresponding to the sub-task, the sub-task is assigned to the corresponding target agent, and the execution results of the assigned sub-task from the target agent are received. Based on the logical dependencies determined by the dependency analysis, the execution results of the sub-tasks are aggregated to generate the execution result of the task to be processed.

[0062] This invention achieves a decentralized, adaptive agent collaboration mechanism by performing refined multi-dimensional feature analysis on the task itself and dynamically decomposing and matching agents based on the analysis results. This overcomes the technical shortcomings of centralized control, which relies on a single central node for global decision-making, resulting in poor scalability and weak robustness. Furthermore, since it no longer relies on preset static rules but instead uses precise matching between task features and agent capability profiles, it can flexibly adapt to dynamically changing task environments and heterogeneous agent resources. This solves the technical problems of insufficient flexibility and inability to optimize orchestration in methods based on preset rules, thus achieving efficient, flexible, and resource-efficient agent collaboration.

[0063] Figure 18 The task processing device 1800 shown is capable of achieving Figure 1 The method described in the embodiment achieves the same technical effect, and can be specifically referred to in the above description. Figure 1 The task processing method of the illustrated embodiment will not be described in detail here. The task processing apparatus 1800 includes: Receiver module 1801 is used to receive tasks to be processed; The analysis module 1802 is used to perform multi-dimensional feature analysis on the task to be processed and obtain analysis results; the multi-dimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; the dependency analysis is used to identify data dependency, temporal dependency and logical dependency between operation units within the task to be processed; the resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed. The decomposition module 1803 is used to decompose the task to be processed into multiple sub-tasks based on the analysis results. The matching module 1804 is used to match a corresponding target intelligent agent for the subtask from the preset intelligent agents according to the capability profiles of multiple preset intelligent agents, and to assign the subtask to the corresponding target intelligent agent; the capability profile is used to describe the task types and task processing capabilities that the corresponding intelligent agent can handle; The aggregation module 1805 is used to receive the execution results of the target intelligent agent for the assigned sub-tasks, and aggregate the execution results of the sub-tasks based on the logical dependencies determined by the dependency analysis to generate the execution results of the task to be processed.

[0064] Optionally, the task to be processed is a lifecycle management task for insurance customers; the decomposition module 1803 is used for: The lifecycle management task for the insurance customer is broken down into a data collection subtask, a profile building subtask, and a business decision-making subtask. The data collection subtask collects customer transaction data, behavioral data, and credit data. The profile building subtask constructs a customer profile based on the transaction data, behavioral data, and credit data. The business decision-making subtask performs at least one of the following based on the customer profile: customer value prediction, customer risk assessment, and insurance product recommendation analysis.

[0065] Optionally, the analysis module 1802 is used for: Based on a preset complexity assessment model, the task to be processed is classified into complexity levels to obtain a complexity level; the complexity assessment model is trained based on the number of operation units of the sample task, the number of data types processed, and the number of conditional branches of the business rules. Based on the data dependencies, temporal dependencies, and logical dependencies between the operation units within the task to be processed, an internal dependency graph is constructed. Furthermore, based on historical task execution data, the input and output relationships between different tasks are analyzed, and a task dependency graph is constructed based on the analyzed relationships. Based on the complexity level, the internal dependency graph, and the task dependency graph, the computing resources required for each processing stage and the data resources that need to be accessed when executing the task to be processed are determined.

[0066] Optionally, the matching module 1804 is used for: The running status values ​​of each preset intelligent agent are acquired in real time; the running status values ​​include CPU load, memory usage, task queue length, and historical task success rate. Based on the task type of the subtask, the running status value, and the task processing capability required to process the subtask, calculate the matching score between each preset agent and the subtask; The preset agent with the highest matching score is determined as the target agent corresponding to the subtask.

[0067] Optionally, the device further includes ( Figure 18 (not shown in the image) The monitoring module 1806 is used to monitor the execution status of the target intelligent agent on the assigned sub-task after the sub-task is assigned to the corresponding target intelligent agent; The reassignment module 1807 is used to trigger a reassignment mechanism if the monitored execution status indicates that the target agent has timed out or failed to execute. The reassignment mechanism is to assign the subtask to a preset agent with the second highest matching score, or to recalculate the matching score between each preset agent and the subtask, and reassign the subtask to the preset agent with the highest matching score after recalculation.

[0068] Optionally, the device further includes ( Figure 18 (not shown in the image) The recording module 1808 is used to record the actual execution performance data of each subtask after the execution result of the task to be processed is generated; the execution performance data includes the execution time. Evaluation module 1809 is used to evaluate the actual processing efficiency of the target agent executing each of the sub-tasks based on the difference between the actual performance data and the expected performance data. The adjustment module 1810 is used to adjust the processing capability rating of the same type of sub-task in the capability profile of the target intelligent agent according to the actual processing efficiency; the processing capability rating is used to determine the number of sub-tasks of the same type; the task type of the sub-tasks of the same type is the same as the task type of the sub-tasks executed by the target intelligent agent.

[0069] In this embodiment of the invention, after receiving the task to be processed, a multi-dimensional feature analysis can be performed on the task to obtain the analysis results. This multi-dimensional feature analysis can include complexity analysis, dependency analysis, and resource requirement analysis. Dependency analysis can be used to identify data dependencies, temporal dependencies, and logical dependencies between operation units within the task to be processed. Resource requirement analysis can be used to assess the computational and data resource requirements when executing the task to be processed. Then, the task to be processed can be decomposed into multiple sub-tasks based on the analysis results. Based on the capability profiles of multiple preset agents, a corresponding target agent is matched for each sub-task from the preset agents. The capability profile describes the types of tasks and task processing capabilities that the agent can handle. After determining the target agent corresponding to the sub-task, the sub-task is assigned to the corresponding target agent, and the execution results of the assigned sub-task from the target agent are received. Based on the logical dependencies determined by the dependency analysis, the execution results of the sub-tasks are aggregated to generate the execution result of the task to be processed.

[0070] This invention achieves a decentralized, adaptive agent collaboration mechanism by performing refined multi-dimensional feature analysis on the task itself and dynamically decomposing and matching agents based on the analysis results. This overcomes the technical shortcomings of centralized control, which relies on a single central node for global decision-making, resulting in poor scalability and weak robustness. Furthermore, since it no longer relies on preset static rules but instead uses precise matching between task features and agent capability profiles, it can flexibly adapt to dynamically changing task environments and heterogeneous agent resources. This solves the technical problems of insufficient flexibility and inability to optimize orchestration in methods based on preset rules, thus achieving efficient, flexible, and resource-efficient agent collaboration.

[0071] Figure 19 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Please refer to it. Figure 19 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0072] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 19 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0073] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0074] The processor reads the corresponding computer program from non-volatile memory into main memory and then executes it, forming a non-contiguous transfer configuration at the logical level. The processor executes the program stored in memory and specifically performs the following operations: Receive pending tasks; A multidimensional feature analysis is performed on the task to be processed to obtain the analysis results; the multidimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; the dependency analysis is used to identify data dependency, temporal dependency and logical dependency between operation units within the task to be processed; the resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed. Based on the analysis results, the task to be processed is decomposed into multiple sub-tasks; Based on the capability profiles of multiple preset intelligent agents, a corresponding target intelligent agent is matched from the preset intelligent agents for the subtask, and the subtask is assigned to the corresponding target intelligent agent; the capability profile is used to describe the types of tasks that the corresponding intelligent agent can handle and the task processing capabilities. The system receives the execution results of the target agent for the assigned sub-tasks, and aggregates the execution results of the sub-tasks based on the logical dependencies determined by the dependency analysis to generate the execution results of the task to be processed.

[0075] The above is as stated in this application. Figure 1The task processing method disclosed in the embodiments described above can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor mentioned above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in one or more embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in one or more embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0076] The electronic device can also perform Figure 1 The task processing method described herein will not be elaborated further here.

[0077] This application also proposes a computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by a portable electronic device including multiple applications, enable the portable electronic device to perform... Figure 1 The methods of the embodiments shown are not described in detail here.

[0078] This application also proposes a computer program product, which is stored in a storage medium and executed by at least one processor to implement... Figure 1 The methods of the embodiments shown are not described in detail here.

[0079] Of course, in addition to software implementation, the electronic device of this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0080] In summary, the above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this application should be included within the scope of protection of one or more embodiments of this application.

[0081] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0082] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined in the embodiments of this application, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0083] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0084] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A task processing method, characterized in that, The method includes: Receive pending tasks; A multidimensional feature analysis is performed on the task to be processed to obtain the analysis results; the multidimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; the dependency analysis is used to identify data dependency, temporal dependency and logical dependency between operation units within the task to be processed; the resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed. Based on the analysis results, the task to be processed is decomposed into multiple sub-tasks; Based on the capability profiles of multiple preset intelligent agents, a corresponding target intelligent agent is matched from the preset intelligent agents for the subtask, and the subtask is assigned to the corresponding target intelligent agent; the capability profile is used to describe the types of tasks that the corresponding intelligent agent can handle and the task processing capabilities. The system receives the execution results of the target agent for the assigned sub-tasks, and aggregates the execution results of the sub-tasks based on the logical dependencies determined by the dependency analysis to generate the execution results of the task to be processed.

2. The method according to claim 1, characterized in that, The task to be processed is a lifecycle management task for insurance customers; the task to be processed is decomposed into multiple sub-tasks, including: The lifecycle management task for the insurance customer is broken down into a data collection subtask, a profile building subtask, and a business decision-making subtask. The data collection subtask collects customer transaction data, behavioral data, and credit data. The profile building subtask constructs a customer profile based on the transaction data, behavioral data, and credit data. The business decision-making subtask performs at least one of the following based on the customer profile: customer value prediction, customer risk assessment, and insurance product recommendation analysis.

3. The method according to claim 1, characterized in that, The multidimensional feature analysis of the task to be processed includes: Based on a preset complexity assessment model, the task to be processed is classified into complexity levels to obtain a complexity level; the complexity assessment model is trained based on the number of operation units of the sample task, the number of data types processed, and the number of conditional branches of the business rules. Based on the data dependencies, temporal dependencies, and logical dependencies between the operation units within the task to be processed, an internal dependency graph is constructed. Furthermore, based on historical task execution data, the input and output relationships between different tasks are analyzed, and a task dependency graph is constructed based on the analyzed relationships. Based on the complexity level, the internal dependency graph, and the task dependency graph, the computing resources required for each processing stage and the data resources that need to be accessed when executing the task to be processed are determined.

4. The method according to claim 1, characterized in that, The step of matching a target intelligent agent for the subtask from the preset intelligent agents' capability profiles includes: The running status values ​​of each preset intelligent agent are acquired in real time; the running status values ​​include CPU load, memory usage, task queue length, and historical task success rate. Based on the task type of the subtask, the running status value, and the task processing capability required to process the subtask, calculate the matching score between each preset agent and the subtask; The preset agent with the highest matching score is determined as the target agent corresponding to the subtask.

5. The method according to claim 4, characterized in that, After assigning the subtask to the corresponding target agent, the method further includes: Monitor the execution status of the target intelligent agent on the assigned sub-tasks; If the monitored execution status indicates that the target agent has timed out or failed to execute, a reassignment mechanism is triggered. The reassignment mechanism is to assign the subtask to a preset agent with the second highest matching score, or to recalculate the matching score between each preset agent and the subtask, and reassign the subtask to the preset agent with the highest matching score after recalculation.

6. The method according to claim 1, characterized in that, After generating the execution result of the task to be processed, the method further includes: Record the actual execution performance data of each subtask; the execution performance data includes the execution time. Based on the difference between the actual performance data and the expected performance data, the actual processing efficiency of the target agent executing each of the sub-tasks is evaluated. Based on the actual processing efficiency, the processing capability rating of the same type of sub-task in the capability profile of the target intelligent agent is adjusted; the processing capability rating is used to determine the number of sub-tasks of the same type; the task type of the sub-tasks of the same type is the same as the task type of the sub-tasks executed by the target intelligent agent.

7. A task processing device, characterized in that, The device includes: The receiving module is used to receive tasks to be processed. The analysis module is used to perform multi-dimensional feature analysis on the task to be processed and obtain analysis results; the multi-dimensional feature analysis includes complexity analysis, dependency analysis and resource requirement analysis; the dependency analysis is used to identify data dependency, temporal dependency and logical dependency between operation units within the task to be processed; the resource requirement analysis is used to evaluate the computing resource requirements and data resource requirements when executing the task to be processed. The decomposition module is used to decompose the task to be processed into multiple sub-tasks based on the analysis results. The matching module is used to match a corresponding target intelligent agent for the subtask from the preset intelligent agents based on the capability profiles of multiple preset intelligent agents, and to assign the subtask to the corresponding target intelligent agent; the capability profile is used to describe the types of tasks that the corresponding intelligent agent can handle and the task processing capabilities. The aggregation module is used to receive the execution results of the target agent for the assigned sub-tasks, and aggregate the execution results of the sub-tasks based on the logical dependencies determined by the dependency analysis to generate the execution results of the task to be processed.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store computer-executable instructions that, when executed by a processor, implement the steps of the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the method described in any one of claims 1 to 6.

Images