Task cooperation processing method and device using distributed intelligent agent network and medium

Abstract

This invention provides a task collaboration processing method, apparatus, and medium using a distributed intelligent agent network. It receives task processing requests from user terminals, performs hierarchical semantic parsing and task dependency analysis to generate a structured set of atomic tasks, inputs these tasks into a distributed intelligent agent network, performs collaborative decision-making and execution intention confirmation on the atomic task units, constructs a dynamic task graph based on the decision results and dependency descriptors, and executes the corresponding atomic task units in parallel according to the task allocation relationships in the dynamic task graph. Task execution status information and intermediate data output are exchanged in real time through a preset publish-subscribe communication method. A distributed execution trajectory is generated based on the exchange results, and the execution results of each atomic task unit are correlated and fused to calculate a global consistency weight. Based on this weighted aggregation, a global task processing result is generated and sent to the user terminal. This invention can improve the accuracy of task collaboration processing.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method, apparatus, and medium for task collaboration processing using a distributed intelligent agent network. Background Technology

[0002] With the development of distributed computing technology, task collaboration processing technology based on intelligent agent networks has been widely applied in complex task execution scenarios, such as financial services and AI+investment advisory scenarios. Distributed task processing through the collaborative work of multiple intelligent agent nodes can improve task execution efficiency and resource utilization. Currently, common task collaboration processing typically involves breaking down a task into multiple sub-tasks, assigning these sub-tasks to different intelligent agent nodes according to preset allocation rules, with each node executing independently and then summarizing the results. During execution, node status is periodically polled and then aggregated. However, the task decomposition stage struggles to accurately identify implicit dependencies between sub-tasks, leading to execution order conflicts or resource contention; sub-task allocation cannot dynamically adapt to real-time load changes of intelligent agent nodes, easily resulting in some nodes being overloaded while others are idle; and there is a lag in the transmission of status information during execution, making it difficult to trace the root cause of task anomalies, thus affecting the overall accuracy and reliability of task processing. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide at least one method, apparatus and medium for task collaboration processing using distributed intelligent agent networks.

[0004] According to one aspect of the present invention, a task collaboration processing method using a distributed agent network is provided. The method includes: receiving a task processing request sent by a user terminal; performing hierarchical semantic parsing and task dependency analysis on the task processing request to generate a structured atomic task set, wherein the atomic task set includes multiple atomic task units with execution order constraints and dependency descriptors between the atomic task units; inputting the atomic task set into a preset distributed agent network; having each agent in the network perform collaborative decision-making and execution intention confirmation on the atomic task units based on its own capability description vector and current load state; and constructing a task execution path and agent allocation based on the decision results and the dependency descriptors. A dynamic task graph of relationships is constructed. Each agent executes corresponding atomic task units in parallel according to the task allocation relationships in the dynamic task graph. During execution, task execution status information and intermediate data outputs are exchanged in real time through a preset publish-subscribe communication method. Based on the exchange results, a distributed execution trajectory containing timestamps is generated. The execution results of each atomic task unit in the distributed execution trajectory are correlated and fused. A result dependency network is constructed by combining the dependency relationship descriptor, and the global consistency weight of each result node is calculated. Based on the global consistency weight, the execution results in the result dependency network are weighted and aggregated to generate a global task processing result that satisfies the task processing request. The global task processing result is then sent to the user terminal.

[0005] According to another aspect of the present invention, a task collaboration processing apparatus is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable code, which, when executed by the processor, causes the processor to perform the method described above.

[0006] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method.

[0007] This invention receives task processing requests from user terminals and generates a structured atomic task set containing atomic task units and dependency descriptors through hierarchical semantic parsing and task dependency analysis. This allows for precise identification of the inherent logic and execution order constraints of tasks, avoiding execution conflicts caused by ambiguous dependencies in traditional task decomposition. The atomic task set is then input into a distributed agent network, where each agent makes collaborative decisions and confirms execution intentions based on its capability description vector and current load status. A dynamic task graph is constructed using dependency descriptors, enabling dynamic adaptation between tasks and agent resources and resolving issues of low resource utilization or load imbalance in static allocation. Execution status information and intermediate data outputs are exchanged in real time through a preset publish-subscribe communication method, generating a structured atomic task set containing... The timestamped distributed execution trajectory ensures the traceability of the task execution process and the real-time and efficient data transmission, overcoming the limitations of information lag or fragmentation in traditional execution monitoring. The execution results in the distributed execution trajectory are correlated and fused, and a result dependency network is constructed by combining dependency descriptors and calculating the global consistency weight of each result node. This makes the correlation between task unit execution results structured and the contribution quantifiable, avoiding the averaging of result importance in simple aggregation. Based on the global consistency weight, the execution results in the result dependency network are weighted and aggregated to generate a global task processing result that satisfies the task processing request. This ensures that the final result accurately reflects the contribution logic of each task unit and the overall reliability, improving the accuracy and efficiency of complex task collaborative processing. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the architecture of an application environment provided by the present invention; Figure 2 This is a flowchart illustrating a task collaboration processing method using a distributed intelligent agent network provided by the present invention. Figure 3 This is a schematic diagram of the structure of a task collaboration processing device provided in an embodiment of the present invention. Detailed Implementation

[0009] To facilitate a clearer understanding of this invention, a schematic diagram of the application environment in which this invention is implemented is first presented, such as... Figure 1 As shown, this includes a task collaboration processing unit 10 and a user terminal cluster. The user terminal cluster may include one or more user terminals; the number of user terminals is not limited here. Figure 1 As shown, the user terminal cluster may specifically include user terminal 1, user terminal 2, ..., user terminal n; it can be understood that user terminal 1, user terminal 2, user terminal 3, ..., user terminal n can all be connected to the task collaboration processing device 10 via network so that each user terminal can interact with the task collaboration processing device 10 via network connection.

[0010] It is understood that the task collaboration processing device 10 can refer to a device that executes the method of the present invention, such as a server. The server can be a single physical server, or a server cluster or distributed system consisting of at least two physical servers. The user terminal can specifically refer to a smartphone, tablet computer, laptop computer, etc., but is not limited thereto. The various user terminals and servers can be directly or indirectly connected through wired or wireless communication. At the same time, the number of user terminals and servers can be one or at least two, and the present invention does not limit this.

[0011] Further, please see Figure 2 This is a flowchart illustrating a task collaboration processing method using a distributed intelligent agent network provided in an embodiment of the present invention. Figure 2 As shown, this method can be derived from... Figure 1 The task collaboration processing device 10 in the middle performs the following steps: Step S100: Receive a task processing request sent by a user terminal, perform hierarchical semantic parsing and task dependency analysis on the task processing request, and generate a structured atomic task set. The atomic task set contains multiple atomic task units with execution order constraints and dependency descriptors between each atomic task unit.

[0012] This invention uses financial services and AI+investment advisory scenarios as examples for illustration. For instance, a task processing request sent by a user terminal is to create a personalized investment portfolio plan for a client. A task processing request is an instruction from the user that they want the system to complete a corresponding task; for example, a client might request that a suitable investment portfolio be generated based on their risk tolerance, investment goals, and asset situation.

[0013] To generate a structured set of atomic tasks, natural language processing techniques can be used to perform hierarchical semantic parsing of task processing requests. For example, deep learning-based semantic analysis models, such as BERT (Bidirectional Encoder Representations from Transformers), can be used to encode and classify the request text, identifying semantic information at different levels. Regarding task dependency analysis, a task dependency graph can be constructed, and graph algorithms can be used to determine the dependencies between atomic task units. For example, topological sorting algorithms can be used to process the task dependency graph to clarify the execution order of tasks.

[0014] In one embodiment, step S100 may specifically include the following steps S110 to S160: Step S110: Perform task feature modeling on the task processing request, extract the task objective description, constraint set and context association information from the task processing request, and construct a task requirement feature model containing multi-dimensional features, including functional objective features, resource requirement features and time constraint features.

[0015] In the scenario described in this invention, for a task processing request to formulate an investment portfolio plan, the task objective description is a detailed explanation of the final goal to be achieved, such as generating an investment portfolio for the client with an expected return of a certain level over the next year and a risk level within an acceptable range. The constraint set refers to the restrictions on the task execution process, such as the client stipulating that investment funds cannot exceed a certain amount, or that investment targets must come from specific industries. Contextual information is background information related to the task, such as the client's previous investment history and current market conditions.

[0016] Functional objective features describe the functional aspects of the task to be achieved, such as the portfolio plan aiming to increase asset value or diversify risk. Resource requirement features involve the resources needed to complete the task, such as computing resources (for complex investment calculations) and data resources (e.g., financial market data). Time constraint features are the time limits for completing the task, such as requiring the portfolio plan to be developed within a week. To build a task requirement feature model, feature engineering can be used. First, text mining is performed on the task processing request to extract the task objective description, the set of constraints, and contextual information. For example, regular expression matching techniques can be used to extract key constraints from the request text. Then, this information is quantified and encoded, converting it into a multi-dimensional feature vector. For example, functional objective features can be represented using binary encoding to indicate whether different functions need to be implemented, and resource requirement features can be represented using specific resource quantities or specifications. Finally, machine learning algorithms, such as Support Vector Machines (SVM) or Random Forests, are used to model these features, resulting in a task requirement feature model.

[0017] Step S120: Construct the task execution logic chain based on the task requirement feature model, decompose the high-level task objective into multiple interrelated mid-level task logic units, and each mid-level task logic unit contains the operation sequence and input / output interface definition for implementing the sub-objective.

[0018] In one embodiment, step S120, constructing a task execution logic chain based on the task requirement feature model, may specifically include the following steps S121~S126: Step S121: Based on the functional target features in the task requirement feature model, query the preset domain causal relationship network to obtain the causal relationship path to achieve the functional target. The causal relationship path includes multiple event nodes and causal triggering relationships between events.

[0019] In this embodiment of the invention, the pre-defined domain causal relationship network is a network that stores the causal relationships between various events in the financial field. For example, this network records the causal relationship between a customer's risk tolerance and their choice of investment targets; that is, a customer with a high risk tolerance is more likely to choose a higher-risk investment target. The event nodes in the causal relationship path can be specific financial events, such as customer risk assessment results or the price fluctuations of a stock. The causal triggering relationships between events describe the sequence and causal connection between these events. A graph database can be used to store the domain causal relationship network. For example, using the Neo4j graph database, event nodes are treated as nodes in the graph, and the causal triggering relationships between events are treated as edges. Then, a graph query language (such as Cypher) is used to query based on functional target features to determine the causal relationship paths that meet the conditions.

[0020] Step S122: Map the event nodes in the causal path to task execution steps, determine the logical dependencies between steps based on the causal triggering relationship between events, and obtain a preliminary sequence of task execution steps.

[0021] When mapping and determining logical dependencies, a mapping table can be established to map event nodes one-to-one with task execution steps. Simultaneously, the order of task execution steps is determined based on the edge information in the causal path. A dictionary data structure can be used to implement the mapping table, and a preliminary sequence of task execution steps can be generated by traversing the causal path.

[0022] Step S123: Combining the resource requirement features and time constraint features in the task requirement feature model, perform resource-time constraint matching on the preliminary task execution step sequence, and allocate resource occupation time periods and time windows for each step.

[0023] When performing resource-time constraint matching, resource allocation algorithms, such as the Earliest Start Time Scheduling algorithm, can be used. This algorithm allocates resources to steps that start earlier, based on the order of task execution and resource requirements. Simultaneously, it incorporates time constraint characteristics to ensure that each step can be completed within the specified time window.

[0024] Step S124: Decompose each task execution step into a mid-level task logic unit containing preconditions, execution actions, and postconditions.

[0025] Taking the task of "collecting basic customer information for risk assessment" as an example, the preconditions are that the customer agrees to provide basic information and the relevant data collection system is functioning normally. The execution action involves collecting basic customer information, such as age, income, and occupation, through online questionnaires or interviews, and conducting a risk assessment based on this information. The postcondition is obtaining the customer's risk tolerance level and storing the assessment results in the system. This breakdown can be achieved using text analytics and rule engine technologies. For example, rules can be defined using a rule engine (such as Drools) to automatically identify the preconditions, execution actions, and postconditions based on the description of the task execution steps.

[0026] Step S125: Perform time-series correlation analysis on the mid-level task logic units, identify logic unit groups that may be executed in parallel, determine the feasibility of parallel execution based on resource occupation periods, and generate parallel execution suggestions.

[0027] When performing temporal correlation analysis and assessing the feasibility of parallel execution, graph theory and resource allocation algorithms can be used. First, a task execution graph is constructed, with mid-level task logic units as nodes and temporal relationships between nodes as edges. Then, a resource allocation algorithm (such as a greedy algorithm) is used to analyze the resource usage periods of each node to determine the feasibility of parallel execution. Based on the analysis results, parallel execution suggestions are generated, such as suggesting that "collecting basic customer information for risk assessment" and "querying financial market data" be executed in parallel.

[0028] Step S126: Integrate logical dependencies, resource-time constraints, and parallel execution suggestions to construct a task execution logic chain that includes sequential execution units and parallel execution unit groups.

[0029] Step S130: Refine the operation granularity of the middle-level task logic unit, identify the smallest execution unit in combination with the input and output interface definition, and decompose each middle-level task logic unit into atomic task units containing specific operation instructions and parameter constraints to obtain a set of atomic task units.

[0030] The input / output interface definition clearly defines the input data and output results of the mid-level task logic unit, which is used to identify the smallest execution unit. For example, the smallest execution unit "obtain customer risk assessment results" takes the storage location of the customer risk assessment system as input and outputs the customer's risk assessment results as output. Each smallest execution unit is further broken down into atomic task units containing specific operation instructions and parameter constraints. For example, the smallest execution unit "obtain customer risk assessment results" can be broken down into atomic task units such as "connect to the customer risk assessment system database" and "query the risk assessment results of a specified customer." Each atomic task unit has specific operation instructions and parameter constraints, such as the database connection instruction using a specific database connection string, and the database username and password as parameter constraints.

[0031] By breaking down the intermediate-level task logic units layer by layer, a set of atomic task units is finally obtained. A recursive algorithm can be used to implement the decomposition process, starting from the top-level intermediate-level task logic unit and continuously decomposing downwards until the smallest atomic task unit is obtained.

[0032] Step S140: Traverse the set of atomic task units, analyze the source of input parameters and the destination of output parameters of different atomic task units, identify task unit pairs with parameter passing relationships, and generate task execution order constraints based on parameter passing direction.

[0033] In one embodiment, step S140 may specifically include the following steps S141 to S146: Step S141: Construct a parameter hierarchy table for each atomic task unit. The parameter hierarchy table includes an input parameter column and an output parameter column. The input parameter column records the names and data types of all parameters required for the execution of the task unit, and the output parameter column records the names and data types of the parameters generated after execution.

[0034] A parameter hierarchy table is a table used to record the input and output parameter information of atomic task units. By constructing a parameter hierarchy table, the parameter requirements and results of each atomic task unit can be clearly understood. To construct a parameter hierarchy table, a database table or a two-dimensional array can be used to store the parameter information.

[0035] Step S142: Traverse the output parameter columns of all atomic task units, match the output parameter names and data types with the input parameter columns of other atomic task units, and mark parameter pairs with the same parameter names and compatible data types.

[0036] For example, for the output parameter "Customer Risk Assessment Result" of the atomic task unit "Query the risk assessment results of a specified customer", its name and data type are matched against the input parameter columns of other atomic task units. If the input parameter name of the atomic task unit "Filter investment targets based on risk assessment results" is also "Customer Risk Assessment Result", and the data types are compatible (i.e., the data types are the same or can be converted), then these two parameters are marked as a parameter pair. Through this matching and marking operation, potential task unit pairs with parameter passing relationships can be identified. In implementing the matching process, a loop structure can be used to traverse the parameter hierarchy table of all atomic task units. For each output parameter, the input parameter columns of other atomic task units are then traversed for matching. String comparison and data type judgment methods can be used to determine whether the parameter names are the same and whether the data types are compatible.

[0037] Step S143: Perform source analysis on the parameter pairs to determine that the atomic task unit to which the output parameter belongs is the source task unit, and the atomic task unit to which the input parameter belongs is the target task unit, and establish a source-target task unit mapping relationship.

[0038] For marked parameter pairs, such as the "Customer Risk Assessment Result" output from "Query Risk Assessment Results for a Specified Customer" and the "Customer Risk Assessment Result" input from "Filter Investment Targets Based on Risk Assessment Result," a source analysis is performed. "Query Risk Assessment Results for a Specified Customer" is identified as the source task unit, and "Filter Investment Targets Based on Risk Assessment Result" as the target task unit. A source-target task unit mapping relationship is established, recording from which atomic task unit the parameter is passed to which atomic task unit. A dictionary data structure can be used to store the source-target task unit mapping relationship, where the dictionary keys are the identifiers of the source task units, and the values ​​are the identifiers of the target task units.

[0039] Step S144: Determine whether there is a time overlap between the output parameter generation time of the source task unit and the input parameter requirement time of the target task unit. If there is an overlap, confirm the time validity of parameter transmission; otherwise, mark it as a potential timing conflict.

[0040] To determine time overlap, the start and end times of each atomic task unit can be recorded in the task execution logic chain. Comparing these times allows for the determination of whether a time overlap exists. Timestamps can be used to represent time, and conditional statements can be used to determine time overlap.

[0041] Step S145: Based on the parameter transfer direction and time validity analysis, generate a parameter transfer path record containing the source task unit identifier, target task unit identifier, parameter name, and transfer time window.

[0042] In the scenario of this invention embodiment, a parameter transmission path record is generated based on the parameter transmission direction (from the source task unit to the target task unit) and the results of time validity analysis. For example, for the two atomic task units "querying the risk assessment results of a specified customer" and "screening investment targets based on the risk assessment results", the parameter transmission path record includes the source task unit identifier "querying the risk assessment results of a specified customer", the target task unit identifier "screening investment targets based on the risk assessment results", the parameter name "customer risk assessment results", and the transmission time window (i.e., the intersection of the time when the source task unit outputs the parameters and the time when the target task unit requires the input parameters).

[0043] Step S146: Summarize all parameter transfer path records, identify complex parameter transfer relationships with multiple input sources or multiple target outputs, and generate a list of task unit pairs containing direct and indirect transfer relationships.

[0044] In this embodiment of the invention, all generated parameter transfer path records are summarized. For example, in addition to the parameter transfer path record between "querying the risk assessment results of a specified customer" and "screening investment targets based on the risk assessment results," there are also parameter transfer path records between other atomic task units. Complex parameter transfer relationships with multiple source inputs or multiple target outputs are identified. For example, a target task unit may require the output parameters of multiple source task units as input, or the output parameters of a source task unit may be transferred to multiple target task units. These complex parameter transfer relationships are identified by analyzing the summarized parameter transfer path records. A list of task unit pairs containing direct and indirect transfer relationships is generated. A direct transfer relationship means that the parameter is directly transferred from the source task unit to the target task unit, while an indirect transfer relationship means that the parameter reaches the target task unit after being transferred through intermediate task units. A graph structure can be used to represent the task unit pair list, where nodes represent atomic task units, edges represent parameter transfer relationships, and edge attributes can represent information such as the name of the transferred parameter and the transfer time window.

[0045] Step S150: Construct a directed dependency graph based on the task execution order constraints. Nodes in the graph represent atomic task units, and directed edges represent the execution order dependencies between task units.

[0046] A directed dependency graph is constructed based on the execution order constraints of these tasks. Nodes in the graph represent atomic task units, such as "querying the risk assessment results of a specified customer" and "screening investment targets based on the risk assessment results". Directed edges represent the execution order dependencies between task units. For example, there is a directed edge from the node "querying the risk assessment results of a specified customer" to the node "screening investment targets based on the risk assessment results", indicating that the former must be executed before the latter.

[0047] Step S160: Integrate the node and edge information in the directed dependency graph in a structured manner to obtain a structured atomic task set containing atomic task unit sequences and dependency descriptors.

[0048] An atomic task unit sequence is a list of atomic task units arranged in the order of task execution. A dependency descriptor describes the dependencies between these atomic task units. For example, in a directed dependency graph, for the nodes "querying the risk assessment results of a specified customer" and "screening investment targets based on the risk assessment results," the dependency descriptor can record that the former is a prerequisite task for the latter. To achieve structured integration, a data structure can be used to store node and edge information. For example, using a JSON-formatted data structure, the atomic task unit sequence can be stored as an array, and the dependency descriptor as an object, where key-value pairs represent the dependencies between task units. This structured integration yields a structured set of atomic tasks containing the atomic task unit sequence and dependency descriptors, facilitating subsequent task processing and management.

[0049] Step S200: Input the atomic task set into the preset distributed agent network, and have each agent make collaborative decisions and confirm execution intentions for the atomic task units based on its own capability description vector and current load status. Construct a dynamic task graph containing task execution paths and agent allocation relationships based on the decision results and dependency descriptors.

[0050] In one embodiment, step S200 may specifically include the following steps S210 to S260: Step S210: Receive the atomic task set through the coordination node in the distributed intelligent agent network, prioritize the atomic task units, and generate a task priority sequence. The sorting criteria include the position of the task unit in the dependency descriptor and the urgency of resource requirements. The sorting interval of adjacent task units in the task priority sequence is dynamically adjusted through the resource requirement volatility.

[0051] The sorting interval between adjacent task units in the task priority sequence is dynamically adjusted based on resource demand volatility. Resource demand volatility refers to the rate of change in a task unit's resource demand over different time periods. If a task unit has high resource demand volatility, its sorting interval with adjacent task units may be appropriately increased to avoid resource conflicts. To implement priority sorting, sorting algorithms, such as quicksort, can be used. During the sorting process, each task unit is assigned a priority value based on its position in the dependency descriptor and the urgency of its resource demand, and then sorted according to these priority values. Simultaneously, the sorting interval between adjacent task units is dynamically adjusted based on resource demand volatility. Statistical analysis methods can be used to calculate resource demand volatility, such as calculating the standard deviation of a task unit's resource demand over a period of time.

[0052] Step S220: Broadcast a task capability requirement query to all agent nodes in the network through the coordinating node. The task capability requirement query includes the operation type, input and output data specifications and execution quality requirements of each atomic task unit, triggering the agent node's capability self-check and load assessment process.

[0053] In this embodiment of the invention, the coordinating node broadcasts a task capability requirement query to all agent nodes in the network. The task capability requirement query is a query containing detailed information about each atomic task unit. The operation type refers to the specific operation to be performed by the atomic task unit, such as a query operation or a calculation operation. The input / output data specifications refer to the format, type, and size requirements of the input and output data of the atomic task unit. The execution quality requirements refer to the quality requirements for the execution results of the atomic task unit, such as the accuracy of the calculation results and the completeness of the query results. Upon receiving the task capability requirement query, the agent node triggers a capability self-check and load assessment process. The capability self-check refers to the agent node checking whether it has the capability to execute each atomic task unit, for example, checking whether it has the corresponding software tools or algorithms to perform the query operation.

[0054] Step S230: When each intelligent agent node responds to a query, it parses the operation type and compares it with the operation sequence in its own historical execution record. It determines the capability adaptation basis through the operation sequence matching degree, and then calculates the data flow efficiency coefficient in combination with the input and output data specifications. The data flow efficiency coefficient is dynamically generated by the ratio of data throughput to format conversion time. It also analyzes the current processing queue length and resource usage ratio simultaneously to obtain a load margin assessment.

[0055] In one embodiment, step S230 may specifically include the following steps S231 to S236: Step S231: Extract the operation type identifier of the atomic task unit through the agent node, query the local operation skill graph, where the node represents the operation type, the edge weight represents the transfer proficiency between operations, and calculate the skill transfer cost between the current operation and the historically executed operation through path analysis.

[0056] A local operation skill graph is a graph that stores the relationships between types of operations historically executed by an agent. Path analysis is used to calculate the skill migration cost between the current operation and historical operations. Path analysis involves finding a path from the current operation type to a historical operation type within the local operation skill graph. The skill migration cost is calculated based on the edge weights along the path. For example, a larger edge weight indicates a higher skill migration cost. Path analysis and skill migration cost calculation can use graph algorithms, such as Dijkstra's algorithm or A* algorithm. Agent nodes find the shortest path from the current operation type node to the historical operation type node within the local operation skill graph and calculate the skill migration cost based on the edge weights along the path.

[0057] Step S232: Determine whether basic execution capability is available based on the skill transfer cost threshold. If the skill transfer cost is lower than the skill transfer cost threshold, analyze the quality requirement parameters corresponding to the operation type, decompose the quality requirement parameters into precision control indicators and stability indicators, and calculate the probability of quality achievement through the frequency of parameter compliance in historical execution data.

[0058] The quality requirement parameters corresponding to the operation type refer to the quality requirements for the execution result of that operation. For example, the quality requirements for a query operation may include the accuracy and completeness of the query results. These quality requirement parameters can be broken down into accuracy control indicators and stability indicators. Accuracy control indicators represent the precision of the operation results, while stability indicators represent the consistency of the operation results.

[0059] The probability of quality achievement is calculated by analyzing the frequency of parameter compliance in historical execution data. For example, in historical query operations, the proportion of query results meeting accuracy standards out of the total number of executions represents the probability of quality achievement. Conditional statements can be used to compare the skill migration cost with a skill migration cost threshold. When calculating the probability of quality achievement, the number of times parameters met standards and the total number of executions in historical execution data can be counted, and then the ratio can be calculated.

[0060] Step S233: For the input and output data specifications, the agent node simulates the data flow process, verifies the data format compatibility, determines whether format conversion is required through the mapping relationship in the format conversion rule base, calculates the ratio of conversion time to data throughput, and generates a data processing efficiency coefficient.

[0061] Data format validation tools can be used to verify data format compatibility and check whether the input data format meets the requirements. When determining whether format conversion is necessary, the mapping relationship between input and output data formats can be found in a format conversion rule base. Performance testing tools can be used to calculate conversion time and throughput; for example, recording the start and end times of format conversion to calculate conversion time, and calculating the amount of data processed per unit time to calculate data throughput.

[0062] Step S234: The skill transfer cost, the probability of achieving quality and the data processing efficiency coefficient are comprehensively calculated for the adaptability. The basic value of capability adaptability is generated by weighted calculation. The weight allocation is adjusted in real time according to the complexity of the operation. The complexity is determined by the number of operation steps and the correlation between parameters.

[0063] A weighted calculation is used to generate a baseline value for capability adaptation, with the weight allocation adjusted in real time based on the complexity of the operation. The complexity of the operation is determined by a combination of the number of operation steps and the correlation between parameters. For example, an operation with many steps and high parameter correlation is more complex and may require a higher weight to be assigned to the probability of achieving the desired quality.

[0064] Step S235: Normalize the capability adaptation base value. In the process, introduce the historical collaboration penalty factor. If there are records of failure of the same type of task, reduce the adaptation base value to generate a capability adaptation base that reflects the overall adaptation level, which serves as the core content of the response.

[0065] Normalization makes the baseline capability values ​​of different agents comparable. The historical collaboration penalty factor is determined based on the agent's historical collaboration records. If an agent has a history of failures in similar tasks, it indicates a potential problem in performing those tasks, thus lowering the baseline capability value. For example, if an agent has failed multiple times in historical query tasks, its baseline capability value will be lowered when responding to the task of "querying the risk assessment results of a specified customer." Normalization algorithms, such as the Min-Max normalization algorithm, can be used. When introducing the historical collaboration penalty factor, a penalty coefficient can be set for the agent based on historical collaboration records, and the adjusted baseline capability value is obtained by multiplying the penalty coefficient by the baseline capability value.

[0066] Step S236: Perform correlation analysis between the capability adaptation basis and the current load status. The load status is converted into load margin by the reciprocal of the resource utilization rate. Together, they form the preliminary response content of the agent node to the task unit and send it to the coordination node.

[0067] Load status is converted into load margin by the reciprocal of resource occupancy rate, which refers to the proportion of computing resources, storage resources, etc., currently occupied by the agent. Load margin indicates how many remaining resources the agent has to execute new tasks. The capability adaptation baseline and load margin together form the agent node's initial response to the task unit, which is sent to the coordination node. The coordination node can then make subsequent decisions and allocate tasks based on these initial responses.

[0068] Step S240: After collecting the capability adaptation basis and load margin assessment of the intelligent agent nodes through the coordination node, a decision evaluation system is constructed from three aspects: capability matching depth, resource margin and collaborative response speed. The adaptation index of each intelligent agent is determined, and multiple intelligent agents with the best index are selected to form a candidate execution group.

[0069] In one embodiment, step S240 may specifically include the following steps S241 to S246: Step S241: Initialize the decision evaluation system for each atomic task unit through the coordination node. The first dimension is the capability matching depth, which is quantified by the difference between the capability adaptation base value and the task requirements. The second dimension is the resource slack, which is calculated by the ratio of the available resources in the load slack assessment to the resource requirements of the task. The third dimension is the collaborative response speed, which is determined by the ratio of the time taken for the agent node to respond to the query to the preset response threshold.

[0070] Step S242: Map the capability adaptation basis, load margin assessment and response time data of each agent node to three assessment dimensions to obtain a multi-dimensional assessment matrix, where each element corresponds to the assessment value of an agent node in a specific dimension.

[0071] For example, for an agent node, its basic capability adaptation value is mapped to the capability matching depth dimension, the load margin assessment result is mapped to the resource margin dimension, and the response time data is mapped to the collaborative response speed dimension. This results in a multi-dimensional evaluation matrix, where rows represent agent nodes, columns represent evaluation dimensions, and each element corresponds to the evaluation value of an agent node in a specific dimension. For instance, an element in the matrix might represent the evaluation value of a certain agent node in the capability matching depth dimension.

[0072] Step S243: Calculate the degree of difference between the evaluation value of each agent node and the ideal fit state. The ideal fit state is when the evaluation value of each dimension reaches the optimal. If the value of a certain dimension exceeds the required threshold, the overall fit degree is determined to be unqualified.

[0073] The ideal fit is the state where all evaluation values ​​across dimensions are optimal, such as maximum capability matching depth, maximum resource sufficiency, and fastest collaborative response speed. The degree of difference between the evaluation value of each agent node and the ideal fit can be calculated using distance metrics, such as Euclidean distance. For example, calculating the Euclidean distance between the evaluation value of an agent node in the three dimensions of capability matching depth, resource sufficiency, and collaborative response speed and the corresponding dimension value in the ideal fit state indicates a smaller distance and a smaller degree of difference. If a value in a certain dimension exceeds the required threshold, it means that the agent cannot meet the task requirements in that dimension, and the overall fit is determined to be unsatisfactory. For example, if the resource sufficiency of an agent node is lower than the required threshold, it means that its remaining resources are insufficient to undertake the task.

[0074] Step S244: After sorting the degree of difference, select the multiple agent nodes with the smallest degree of difference as candidate execution groups. The number of candidates is dynamically adjusted according to the importance level of the task unit. The importance level is determined by the critical path length in the dependency descriptor.

[0075] The number of candidates is dynamically adjusted based on the importance level of the task unit, which is determined by the critical path length in the dependency descriptor. For example, if a task unit has a long critical path in the dependency descriptor, it indicates that it has a greater impact on the completion of the entire task and has a higher importance level, so it may be necessary to select more agent nodes as candidate execution groups.

[0076] Step S245: Analyze the distribution characteristics of the evaluation values ​​of agent nodes in the candidate execution group. If multiple nodes have similar differences, conflict detection is performed. The detection is achieved by calculating the overlap of resource requirements. Among the nodes with high overlap, the nodes with faster collaborative response speed are retained.

[0077] If multiple nodes exhibit similar levels of difference, it indicates that these nodes are relatively close in overall adaptability, potentially indicating resource contention and necessitating conflict detection. Conflict detection is achieved through resource demand overlap calculation, which refers to the degree to which multiple agent nodes require the same resource. For example, if two agent nodes both require significant computing resources, their resource demand overlap is high.

[0078] Step S246: Arrange the selected candidate execution groups in ascending order of their degree of difference, and generate a comprehensive fit index ranking. The ranking results include the three-dimensional evaluation value of each node and the reasons for the difference, which serve as the basis for task pre-allocation decisions.

[0079] Step S250: Conduct multiple rounds of task pre-allocation negotiation with the candidate execution group through the coordination node. The first round of negotiation determines the resource reservation baseline. The second round of negotiation adjusts the collaboration window with the preceding task unit according to the dependency descriptor. The timing feasibility of data transmission is verified through time axis overlap analysis. The final round of negotiation resolves resource competition conflicts.

[0080] In one embodiment, step S250 may specifically include the following steps S251 to S256: Step S251: In the first round of negotiation, the coordinating node sends a list of task resource requirements to the candidate execution group. The list of task resource requirements includes the amount of processing resources, storage resources, and communication bandwidth requirements. The candidate node provides a resource reservation plan based on the current load status, which specifies the reservation period and release conditions for each resource.

[0081] In the first round of negotiation, the coordinating node sends a list of task resource requirements to the candidate execution group. For the task unit "querying the risk assessment results of a specified customer," the task resource requirement list may include the amount of computing resources required to process the query task, the amount of storage resources required to store the query results, and the communication bandwidth requirements for communicating with the database. The candidate nodes provide a resource reservation plan based on their current load status, which represents the current workload of the agent nodes. The resource reservation plan specifies the reservation period and release conditions for each resource.

[0082] Step S252: By resolving the dependency descriptor through the coordinating node, extract the predecessor task identifier of the current task unit, query the candidate execution nodes of the predecessor task unit, establish the collaboration relationship pair between the current task and the predecessor task, and obtain the expected completion time window of the predecessor task.

[0083] In this embodiment of the invention, the coordinating node resolves dependency descriptors, which record the dependencies between atomic task units. For the task unit "querying the risk assessment results of a specified customer," its preceding task identifier is extracted; for example, the preceding task might be "obtaining basic customer information." Candidate execution nodes for the preceding task unit are queried to determine which agent nodes might execute the preceding task. A collaborative relationship pair between the current task and the preceding task is established to clarify the collaborative relationship between the two task units.

[0084] Step S253: Align the resource reservation period of the current task with the completion time window of the preceding task on the time axis, calculate the time overlap interval, and if there is an overlap, further analyze the time required for data transmission, determine whether the data handover can be completed within the overlap interval, and generate a time series feasibility assessment.

[0085] In this embodiment of the invention, the resource reservation period for the current task "querying the risk assessment results of a specified customer" is aligned with the completion time window of the preceding task "obtaining basic customer information" on the timeline. Assuming the resource reservation period for "querying the risk assessment results of a specified customer" is 10:00-10:30, and the completion time window for "obtaining basic customer information" is 9:30-10:00, the relative positions of the two time periods can be clearly seen through timeline alignment.

[0086] The time overlap interval is calculated; in this example, it's 10:00 AM. If overlap exists, the time required for data transfer is further analyzed. The time required for data transfer depends on factors such as data size and transmission rate. For example, after the "obtain basic customer information" task is completed, the basic customer information needs to be transferred to the "query risk assessment results for a specified customer" task. Assume the data transfer time is 5 minutes. It is then determined whether the data handover can be completed within the overlap interval. In this example, since the overlap interval is only 10:00 AM, and data transfer takes 5 minutes, the data handover cannot be completed within the overlap interval. Therefore, the resulting time series feasibility assessment result is infeasible.

[0087] Step S254: The second round of negotiation adjusts the risk points in the timing feasibility assessment. If the time overlap is insufficient, the execution time of the preceding task node is compressed, or the start time of the current task node is delayed. The adjusted resource loss is calculated through time compensation. If the resource loss exceeds the threshold, the candidate node is replaced.

[0088] In the scenario of this invention embodiment, the second round of negotiation focuses on the risk points identified in the time-series feasibility assessment. Taking the two tasks of "querying the risk assessment results of a specified customer" and "obtaining basic customer information" as examples, if the time-series feasibility assessment shows insufficient time overlap and data handover cannot be completed, adjustments are required.

[0089] If there is insufficient time overlap, one adjustment method is to coordinate the execution time of the preceding task node "Acquiring Basic Customer Information" to compress its execution time. For example, if "Acquiring Basic Customer Information" was originally expected to be completed between 9:30 and 10:00, by optimizing the data collection process or improving processing efficiency, the completion time can be brought forward to 9:20-9:50. This increases the time overlap with the resource-reserved time period (10:00-10:30) for "Querying Risk Assessment Results for a Specific Customer". Another method is to delay the start time of the current task node "Querying Risk Assessment Results for a Specific Customer", for example, delaying the start time from 10:00 to 10:10, to increase the overlap with the completion time of the preceding task.

[0090] While adjusting the time, it is necessary to calculate the resource consumption after the adjustment through time compensation. The time compensation calculation takes into account the impact of changes in task execution time on resource usage. For example, compressing the execution time of a preceding task node may require additional computing resources, and delaying the start time of the current task node may cause resources to be idle for a period of time before being used, thus resulting in certain resource consumption.

[0091] Step S255: In the final round of negotiation, when multiple candidate nodes apply for the same physical resource at the same time, the resource usage rights are allocated according to the task priority sequence. The task nodes with lower priority adjust their resource requirements or change their resource types to generate the final resource allocation agreement.

[0092] In this embodiment of the invention, the final round of negotiation primarily addresses resource contention conflicts. Assume that when executing the task of "querying the risk assessment results of a specified customer," multiple candidate nodes simultaneously request computing resources from the same server. Lower-priority task nodes need to adjust their resource requirements or change their resource type. For example, a lower-priority task node that originally needed to use a large amount of the server's computing resources can adjust to using fewer resources or switch to using computing resources from another server. During the negotiation process, the coordinating node communicates and negotiates with the candidate nodes to determine the final resource allocation scheme. A final resource allocation agreement is generated, which specifies the type, quantity, and usage time of resources obtained by each task node.

[0093] Step S256: Once the negotiation is reached, the resource reservation scheme, timing coordination protocol and conflict resolution are integrated into the task execution benchmark through the coordination node. This benchmark includes a resource allocation table, a timing coordination matrix and an exception handling plan, which serve as the basic input for constructing a dynamic task graph.

[0094] Resource reservation schemes define the reservation periods and release conditions for various resources reported by candidate nodes in the first round of negotiation. For example, in the task of "querying the risk assessment results of a specified customer," this involves the usage periods and release rules for the computing resources, storage resources, and communication bandwidth reserved by the agent node. Timing coordination protocols define the time coordination relationships between tasks after adjustments based on timing feasibility assessments in the second round of negotiation. For example, the timing coordination rules between "querying the risk assessment results of a specified customer" and the preceding task "obtaining basic customer information." Conflict resolution solutions resolve resource contention conflicts in the final round of negotiation. For example, specific measures for lower-priority task nodes to adjust resource requirements or change resource types.

[0095] The coordinating node integrates this information into task execution benchmarks. The resource allocation table details the type, quantity, and usage time of resources allocated to each task node. The contingency plan outlines countermeasures for potential anomalies, such as resource failures and task timeouts. These task execution benchmarks serve as the foundational input for constructing a dynamic task graph. Nodes in the dynamic task graph represent atomic task units and agents, while edges represent the task execution order and the allocation relationships between agents and task units. Through these task execution benchmarks, task execution paths and agent allocation relationships can be accurately determined, thereby constructing a reasonable dynamic task graph.

[0096] Step S260: Based on the negotiated resource allocation scheme and timing coordination protocol, connect the atomic task units and agent nodes through directed edges. The edge attributes include the data transmission delay threshold and resource mutual exclusion identifier. Construct the task execution path topology to obtain the dynamic task graph.

[0097] Edge attributes include a data transfer latency threshold and a resource mutual exclusion flag. The data transfer latency threshold refers to the maximum allowed delay time when transferring data between tasks. For example, after completing the task "Query the risk assessment results of a specified customer," the maximum delay time required to transfer the customer's risk assessment results to the task "Select investment targets based on the risk assessment results." The resource mutual exclusion flag indicates whether there is a mutual exclusion relationship between tasks when using resources, such as whether two tasks can simultaneously use the same server's computing resources. By connecting all atomic task units and agent nodes in the above manner, a task execution path topology is constructed, ultimately resulting in a dynamic task graph.

[0098] Step S300: Each agent executes the corresponding atomic task unit in parallel according to the task allocation relationship in the dynamic task graph. During the execution, task execution status information and intermediate data output are exchanged in real time through a preset publish-subscribe communication method. Based on the exchange results, a distributed execution trajectory containing timestamps is generated.

[0099] In one embodiment, step S300 may specifically include the following steps S310 to S350: Step S310: Each agent node locks its own atomic task unit according to the allocation relationship in the dynamic task graph, obtains the operation instructions, parameter constraints and the identifier of the preceding task unit of the atomic task unit, starts the dynamic resource pre-allocation process, reserves processing resources, storage resources and communication bandwidth according to the task complexity, and synchronously negotiates the data receiving window with the preceding task node.

[0100] Each agent node first identifies its assigned atomic task unit based on the allocation relationships in the dynamic task graph. For example, agent node A determines it is responsible for the task "querying the risk assessment results of a specified customer" based on the dynamic task graph. It then obtains the operation instructions, parameter constraints, and identifiers of the preceding task units for that atomic task unit. The operation instructions specify the concrete execution steps of the task; for example, the operation instructions for the task "querying the risk assessment results of a specified customer" might include connecting to the database and executing the query statement. Parameter constraints specify the range and requirements of parameter values ​​during task execution; for example, the customer identifier in the query statement must be in a valid format. The identifier of the preceding task unit is used to identify the preceding task; for example, the preceding task for "querying the risk assessment results of a specified customer" might be "obtaining basic customer information." A dynamic resource pre-allocation process is then initiated, reserving processing resources, storage resources, and communication bandwidth based on the task complexity. Task complexity can be assessed based on factors such as the number of operation steps and the amount of data processed. For example, if the task "querying the risk assessment results of a specified customer" involves a large amount of data querying and complex calculations, it may require reserving more processing and storage resources. Simultaneously, corresponding communication bandwidth is reserved based on the task's data transmission requirements. The data receiving window refers to the time range within which the current task node can receive data from the preceding task node.

[0101] Step S320: Send a data subscription request to the agent node corresponding to the preceding task unit marked in the dynamic task graph. The data subscription request includes data type identifier, update frequency requirements and transmission format description. Establish a dynamic communication link based on task dependency relationship. The priority of the dynamic communication link is dynamically adjusted according to the importance of data in task execution.

[0102] Taking agent node B executing the task of "screening investment targets based on risk assessment results" as an example, it needs to send a data subscription request to agent node A, which corresponds to the preceding task unit "querying the risk assessment results of a specified customer" marked in the dynamic task diagram. The data subscription request includes a data type identifier, update frequency requirements, and transmission format specifications. The data type identifier clarifies the data type to be subscribed to, such as "customer risk assessment results." The update frequency requirements specify the time interval for data updates, such as requiring data to be updated every 5 minutes. Transmission formats include JSON and XML formats.

[0103] Dynamic communication links are established based on task dependencies. These links are dynamically created according to the dependencies between tasks. For example, since the task of "screening investment targets based on risk assessment results" depends on the result of the task of "querying the risk assessment results of a specified customer," a communication link is established from agent node A to agent node B. The priority of these dynamic communication links is dynamically adjusted based on the importance of the data in task execution. If the "customer risk assessment results" are important to the task of "screening investment targets based on risk assessment results," then the communication link has a higher priority to ensure that the data can be transmitted in a timely and accurate manner. For example, in the event of network congestion, high-priority communication links will receive bandwidth resources first.

[0104] Step S330: When the preceding task unit generates intermediate data, the data is received through the dynamic communication link. Feature comparison and context association verification are performed on the data. Feature comparison is achieved by matching the key attributes of the data with the expected feature template. Context association verification is combined with the execution environment parameters of the preceding task to determine the validity of the data and generate data quality assessment results.

[0105] In this embodiment of the invention, when agent node A generates intermediate data (such as customer risk assessment results) by executing the task of "querying the risk assessment results of a specified customer," agent node B receives this data through a dynamic communication link. The received data undergoes feature comparison and contextual verification. Feature comparison is achieved by matching key attributes of the data with expected feature templates. For example, the expected feature template specifies that the customer risk assessment results should include key attributes such as risk level and risk score. Agent node B compares the key attributes of the received data with the expected feature template to check for a match. If the risk level range is specified as 1-5 in the expected template, but the risk level in the received data is 6, it indicates that the data is abnormal.

[0106] Contextual verification, combined with parameters of the preceding task execution environment, determines data validity. These parameters include execution time and the data source used. For example, if the task "query the risk assessment results for a specified customer" uses a specific data source within a specific time frame, agent node B will use these environment parameters to determine the reasonableness of the received data. If the displayed risk assessment results differ significantly from the historical data of the data source, it may indicate a data problem. Based on the results of feature comparison and contextual verification, a data quality assessment result is generated. This result can be categorized into different levels, such as acceptable, unacceptable, and partially acceptable, providing a reference for subsequent task execution.

[0107] Step S340: Configure input parameters according to the parameter requirement list, execute the operation logic corresponding to the atomic task unit, adjust the number of execution threads and the size of data processing blocks according to the real-time resource load, continuously collect execution progress, resource consumption and environmental parameters during execution, and generate a time-series state sequence containing multi-dimensional state parameters.

[0108] Taking the task of "screening investment targets based on risk assessment results" performed by agent node B as an example, the input parameters are first configured according to the parameter requirement list. The parameter requirement list clarifies the parameters required for task execution; for example, the "screening investment targets based on risk assessment results" task may require parameters such as customer risk assessment results and an investment target database. Agent node B configures the received customer risk assessment results and other parameters to ensure the task can be executed correctly. Then, the operational logic corresponding to the atomic task unit is executed, that is, the specific operation of screening investment targets is performed according to the configured parameters. For example, a list of investment targets that meet the risk requirements is screened based on the customer's risk level.

[0109] The number of execution threads and the size of data processing blocks are adjusted based on real-time resource load. Real-time resource load refers to the current usage of computing and storage resources by the agent node. If the resource load is high, agent node B can reduce the number of execution threads to avoid excessive resource contention; simultaneously, the size of the data processing blocks is adjusted, dividing the data into smaller blocks for processing to improve efficiency. For example, if 100 investment targets are processed each time, the processing time can be adjusted to 50 targets per time when the resource load is high. During execution, execution progress, resource consumption, and environmental parameters are continuously collected. Execution progress can be measured by the number of completed operation steps or the amount of data processed, such as the percentage of selected investment targets out of the total number of investment targets. Resource consumption includes information such as computing resource usage time and storage resource occupancy. Environmental parameters include information such as system temperature and network bandwidth. Based on the collected information, a time-series state sequence containing multi-dimensional state parameters is generated. The time-series state sequence records the state information at different points in time during task execution.

[0110] Step S350: Push the time-series state sequence and intermediate data output to the distributed message bus. The coordinating node and the intelligent agent node that subscribes to the distributed message bus adopt a differentiated receiving strategy according to their own task priorities. High-priority tasks receive and process data in real time, while low-priority tasks receive and cache data in batches. After receiving, they return a receiving feedback containing data integrity and processing suggestions to the publishing node.

[0111] In one embodiment, step S350 may specifically include the following steps S351 to S356: Step S351: Generate dynamic topic identifiers based on atomic task unit identifiers, agent node identifiers, and data generation stage identifiers. The topic identifiers adopt a multi-level structure, which includes domain identifiers, task flow identifiers, task unit identifiers, and data type identifiers in sequence.

[0112] Taking the task of "screening investment targets based on risk assessment results" performed by agent node B as an example, dynamic topic identifiers are generated based on the atomic task unit identifier, agent node identifier, and data generation stage. The atomic task unit identifier is the unique identifier for the task "screening investment targets based on risk assessment results." The agent node identifier is the identifier for agent node B. The data generation stage indicates at which stage of task execution the data was generated, such as the data when some investment targets are screened. The topic identifier adopts a multi-level structure, successively including a domain identifier, task flow identifier, task unit identifier, and data type identifier. The domain identifier indicates the domain to which the data belongs, such as the financial investment domain. The task flow identifier indicates the identifier of the entire task flow, such as the task flow of formulating an investment portfolio plan. The task unit identifier is the identifier of a specific atomic task unit, such as "screening investment targets based on risk assessment results." The data type identifier indicates the type of data, such as a list of investment targets or a time-series state sequence.

[0113] Step S352: Perform semantic enhancement processing on the time-series state sequence, converting the execution progress parameters into a structured state description that includes execution stage descriptions, key milestone achievement status, and remaining step predictions, and converting the resource consumption parameters into a resource usage overview that includes resource type, current consumption, and trend predictions.

[0114] For example, if the execution progress parameter shows that the current task is 30% complete, semantic enhancement processing can transform it into a structured status description: "Currently in the preliminary screening stage of investment targets, risk level screening of some investment targets has been completed. The key milestone 'Complete preliminary screening' has not yet been achieved, and further detailed benefit analysis and comparison steps are expected." Similarly, resource consumption parameters can be transformed into a resource usage overview including resource type, current consumption, and trend prediction. For instance, if the resource consumption parameter shows CPU utilization at 30%, semantic enhancement processing can transform it into a resource usage overview: "The current resource type is CPU, and the current consumption is 30%. Based on task execution and historical data prediction, as the screening scope expands, CPU utilization may rise to 40% within the next 10 minutes." Step S353: Construct a feature chain for the intermediate data output. Extract key feature values ​​from the original data and associate them with the feature values ​​of the output data of the preceding task to form a feature transfer chain. Each feature node in the feature transfer chain includes the feature name, numerical range, and correlation strength with the preceding and following features.

[0115] Key features are extracted from the raw data. For example, for a list of investment targets, key features might include the rate of return, risk level, and investment period. Agent node B uses data mining and analysis techniques to extract these key features from the raw data. Feature values ​​from the output data of preceding tasks are then linked to form a feature transfer chain. For example, the feature values ​​output by the preceding task "Query the risk assessment results of a specified customer" might include the customer's risk level and risk score. These feature values ​​from the output data of preceding tasks are then linked to the feature values ​​from the intermediate data produced by the current task to form a feature transfer chain. For instance, a customer's risk level is correlated with the risk level of an investment target, and investment targets that meet the requirements are selected based on the customer's risk level.

[0116] Each feature node in the feature propagation chain includes a feature name, a numerical range, and the strength of its correlation with preceding and following features. For example, the feature name of the feature node "investment target return rate" is "return rate", the numerical range may be 0-20%, and the strength of its correlation with preceding and following features indicates the degree of correlation between this feature and other features (such as risk level, investment period, etc.).

[0117] Step S354: Add metadata to the head of the feature transmission chain. The metadata includes the data generation timestamp, the health status of the publishing node, the data transmission priority, and the expected processing time limit.

[0118] The data generation timestamp records the generation time of data in the feature transmission chain, such as the generation time of intermediate data generated by the task of "screening investment targets based on risk assessment results". The health status of the publishing node indicates the current operating status of agent node B, such as whether it is operating normally or whether there are any faults. Data transmission priority is determined based on the importance and timeliness requirements of the data; for example, data related to real-time investment decisions has a higher transmission priority. The expected processing time limit specifies the time limit for the receiving node to process the data; for example, it requires completing further analysis of the investment target list within 10 minutes.

[0119] Step S355: Push the structured state description, resource usage overview, feature transmission chain and metadata to the distributed message bus. Adjust the push order according to the network node load and data urgency. The adjustment of the push order is achieved through a priority queue algorithm.

[0120] In the scenario of this invention embodiment, agent node B pushes the processed structured state description, resource usage overview, feature transmission chain, and metadata to the distributed message bus.

[0121] The push order is adjusted based on network node load and data urgency. Network node load represents the current workload of each node in the distributed message bus; a high load on a node may affect data transmission and processing efficiency. Data urgency is determined by the importance and timeliness requirements of the data; for example, data related to real-time investment decisions has a high urgency. The push order adjustment is implemented using a priority queue algorithm. A priority queue is a data structure that sorts elements according to their priority. Agent node B assigns a priority to each data item based on network node load and data urgency, inserting the data item into the priority queue. The priority queue automatically sorts data items according to priority, and agent node B pushes the data to the distributed message bus in the order of the queue.

[0122] Step S356: Record the data reception status and processing time of each subscription node, dynamically adjust the subsequent data push frequency and data compression ratio according to the reception delay, and generate a communication quality report that includes transmission success rate, average delay and node processing capacity.

[0123] In this embodiment of the invention, agent node B records the data reception status and processing time of each subscribing node (such as the coordinating node and agent nodes that depend on subsequent tasks). The data reception status includes information such as whether data was successfully received and whether the received data is complete. The processing time represents the time spent by the subscribing node processing the received data.

[0124] The system dynamically adjusts the subsequent data push frequency and data compression ratio based on the reception delay. Reception delay reflects the transmission time of data from the sending node to the receiving node; a long reception delay may affect the timeliness and availability of the data. Agent node B dynamically adjusts the subsequent data push frequency based on the reception delay. For example, if the reception delay is long, the data push frequency is appropriately reduced to avoid data backlog. Simultaneously, the data compression ratio is adjusted to improve data transmission efficiency. If data transmission bandwidth is limited, the data compression ratio can be increased to reduce the amount of data transmitted. A communication quality report is generated, including transmission success rate, average delay, and node processing capacity. The transmission success rate represents the proportion of data successfully transmitted to the subscribed nodes, the average delay represents the average latency of data transmission, and the node processing capacity represents the subscribed node's ability to process data, such as the amount of data processed per second.

[0125] Step S400: Perform correlation and fusion processing on the execution results of each atomic task unit in the distributed execution trajectory, construct a result dependency network by combining the dependency descriptor, and calculate the global consistency weight of each result node.

[0126] In one embodiment, step S400 may specifically include the following steps S410 to S460: Step S410: Extract the execution results of each atomic task unit from the distributed execution trajectory, classify and organize the execution result set according to the task unit identifier, the execution result set includes the task completion status, output data body and process log, trace the source of the preceding data of each result through the dependency descriptor, and establish the result-previous data association index.

[0127] For example, extract the execution results of the tasks "query the risk assessment results of a specified customer" and "screen investment targets based on the risk assessment results" from the distributed execution trajectory.

[0128] The execution result set includes task completion status, output data body, and process log. Task completion status indicates whether the task was successfully completed, such as whether the task "Query the risk assessment results of a specified client" successfully retrieved the client's risk assessment results. The output data body contains the data generated after task execution, such as the client's risk assessment results and the list of selected investment targets. The process log records detailed information about the task execution process, such as execution time and resources used. The preceding data source for each result is traced using dependency descriptors. For example, the result of the task "Filter investment targets based on risk assessment results" depends on the output data of the task "Query the risk assessment results of a specified client," and this preceding data source can be traced using dependency descriptors. A result-preceding data relationship index is established, which records the correspondence between each result and its preceding data source.

[0129] Step S420: Traverse the results - previous association index, compare the feature association degree between the output data body of the current task unit and the output data body of the previous task unit. The feature association degree is calculated by comprehensively considering the number of matching data fields, the consistency of numerical fluctuation trends and the continuity of timestamps, and generates a description of the association strength.

[0130] In the scenario of this invention embodiment, the result-previous association index is traversed, and the feature association degree of each result is compared with the preceding data source. For example, the output data body (the list of selected investment targets) of the task "screening investment targets based on risk assessment results" is compared with the output data body (the risk assessment results of the customer) of the preceding task "querying the risk assessment results of a specified customer".

[0131] Feature correlation is calculated by comprehensively considering the number of matching data fields, the consistency of numerical fluctuation trends, and the continuity of timestamps. The number of matching data fields indicates the number of identical fields in two data volumes. For example, if a customer's risk level field exists in both data volumes, then that field matches. The consistency of numerical fluctuation trends indicates whether the changing trends of related values ​​in the two data volumes are consistent. For example, do the numerical fluctuation trends of a customer's risk score and the risk level of an investment target match? The continuity of timestamps indicates whether the generation times of the two data volumes are consecutive. If the task of "querying the risk assessment results of a specified customer" is completed immediately followed by the task of "filtering investment targets based on risk assessment results," and the time interval is reasonable, then the timestamps are continuous. The correlation strength can be described numerically or textually, such as high correlation strength, medium correlation strength, and low correlation strength.

[0132] Step S430: Construct an initial association network with atomic task units as nodes and association strength described as edge attributes. Node attributes include the reliability of execution results and the scope of data influence, while edge attributes include data transmission loss rate and time series deviation value. Isolated nodes are identified and indirect association paths are supplemented through network density analysis.

[0133] In the scenario of this invention embodiment, an initial association network is constructed using atomic task units as nodes and association strength descriptions as edge attributes. For example, "querying the risk assessment results of a specified customer" and "screening investment targets based on risk assessment results" are used as nodes, and the association strength description between them is used as an edge attribute.

[0134] Node attributes include the reliability of the execution result and the scope of data influence. The reliability of the execution result indicates its dependability; for example, the reliability of the execution result for the task "Query the risk assessment results of a specified customer" can be determined based on factors such as the data source and query method. The scope of data influence indicates the extent to which the execution result affects other tasks or results; for example, a customer's risk assessment results may influence the selection of investment targets and the formulation of investment portfolios.

[0135] The edge attribute includes the data transmission loss rate and the timing deviation value. The data transmission loss rate represents the information loss that may occur during the transmission of data from one task unit to another, such as data loss or errors during transmission. The timing deviation value represents the time deviation of data transmission, such as the deviation between the completion time of the task "Query the risk assessment results of a specified customer" and the start time of the task "Filter investment targets based on risk assessment results" receiving data.

[0136] Network density analysis identifies isolated nodes and adds indirect connection paths. Network density analysis calculates the connection density between nodes in a network. A node with few connections to other nodes is likely an isolated node. For isolated nodes, indirect connection paths are added to establish connections with other nodes. For example, if the result of an atomic task unit is not directly related to the results of other task units, but can be indirectly related through intermediate task units, then those indirect connection paths are added.

[0137] Step S440: Perform multiple rounds of iterative optimization on the initial association network. In the first round of optimization, merge duplicate nodes whose feature association degree exceeds the set threshold. In the second round of optimization, adjust the edge directions with excessive time deviation values. In the final round of optimization, supplement resource competition relationship identifiers to obtain a multi-level result dependency network containing direct and indirect associations.

[0138] In this embodiment of the invention, the initial correlation network undergoes multiple rounds of iterative optimization. The first round of optimization merges duplicate nodes whose feature correlation exceeds a set threshold. This threshold is a pre-determined standard; if the feature correlation of two nodes exceeds this threshold, it indicates that their results are very similar, and they can be merged into one node. The second round of optimization adjusts the direction of edges with excessively large temporal deviations. Excessive temporal deviations may cause problems with the time order of data transmission, affecting task execution efficiency and result accuracy. Adjusting the direction of edges makes the time order of data transmission more reasonable. The final round of optimization supplements resource competition relationship identifiers. Resource competition relationships indicate that different task units may compete for the same resources during execution, such as computing resources and storage resources. In the final round of optimization, resource competition relationships are identified and corresponding identifiers are added to the network to better manage and coordinate task execution. After multiple rounds of iterative optimization, a multi-level result dependency network containing direct and indirect correlations is obtained. This multi-level result dependency network can more accurately reflect the correlations and data transmission paths between task units, providing a more reliable foundation for subsequent node weight calculations and result fusion.

[0139] Step S450: Based on the result-dependent network topology, calculate the node propagation weight layer by layer starting from the network's starting node. The propagation weight is dynamically generated by the product of the association strength description, path length, and node influence range. The path length is the number of edges traversed from the current node to the starting node.

[0140] In one embodiment, step S450 may specifically include the following steps S451 to S456: Step S451: Determine the set of starting nodes in the result dependency network. The set of starting nodes consists of the result nodes corresponding to atomic task units without prior dependencies. Assign basic propagation weights to each starting node. The basic propagation weights are set comprehensively based on the task priority sequence and the reliability of the execution results.

[0141] In the scenario of this invention embodiment, the set of starting nodes in the result dependency network is determined. The set of starting nodes consists of the result nodes corresponding to atomic task units without prior dependencies. For example, the task "querying basic information of a specified customer" has no prior dependencies, and its result node is the starting node.

[0142] Each starting node is assigned a basic propagation weight. The basic propagation weight is set based on a combination of task priority sequence and execution result reliability. The task priority sequence is generated in step S210, which prioritizes tasks according to their position in the dependency descriptor and the urgency of resource requirements. Execution result reliability indicates the reliability of the execution result; for example, the reliability of the execution result for the task "query basic information of a specified customer" can be determined based on factors such as the data source and query method.

[0143] Step S452: Construct a node propagation path table, recording the shortest path from each node to all starting nodes. The path includes the identifiers of intermediate nodes and their corresponding edge attributes. Generate a path priority ranking using a breadth-first search algorithm, with shorter paths and stronger associations having higher priority.

[0144] In this embodiment of the invention, a node propagation path table is constructed. The node propagation path table records the shortest path from each node to all starting nodes, and the path includes the identifiers of intermediate nodes and their corresponding edge attributes. For example, for the node "screening investment targets based on risk assessment results," the shortest path from it to the starting node "querying basic information of a specified customer" is recorded. This path may pass through the intermediate node "querying the risk assessment results of a specified customer," and the corresponding edge attributes include information such as association strength description and data transmission loss rate.

[0145] A path priority ranking is generated using a breadth-first search (BFS) algorithm, which traverses graphs by visiting nodes in a hierarchical order. Paths are prioritized based on their length and association strength, with shorter paths and stronger associations having higher priority.

[0146] Step S453: Start the first round of weight propagation from the starting node. Each starting node distributes the basic propagation weight to the directly associated nodes according to the association strength ratio. The receiving node adds up the allocated weights of all directly associated starting nodes to generate the intermediate weight value after the first round of propagation.

[0147] Taking the starting node "Query basic information of a specified customer" as an example, the basic propagation weight is allocated to directly related nodes according to the association strength ratio. Assuming the basic propagation weight of the starting node "Query basic information of a specified customer" is 1, and the association strength with the node "Query risk assessment results of a specified customer" is 0.8, then a weight of 1 × 0.8 = 0.8 is allocated to the node "Query risk assessment results of a specified customer". The receiving node accumulates the allocated weights of all directly related starting nodes. For example, the node "Query risk assessment results of a specified customer" may be directly related to multiple starting nodes; it accumulates the weights allocated to all directly related starting nodes to generate the intermediate weight value after the first round of propagation.

[0148] Step S454: Perform subsequent rounds of weight propagation. Non-starting nodes distribute their intermediate weight values ​​to downstream associated nodes according to the association strength ratio. During the distribution process, a path attenuation coefficient is introduced. The weight value decreases by the attenuation coefficient for each intermediate node the path passes through. The attenuation coefficient is dynamically adjusted according to the average association strength of the network.

[0149] In the scenario of this invention embodiment, weight propagation is performed in subsequent rounds. Non-starting nodes distribute their intermediate weight values ​​to downstream related nodes according to the correlation strength ratio. For example, the node "Query the risk assessment results of a specified customer" obtains an intermediate weight value after the first round of propagation, and it distributes this intermediate weight value to the node "Select investment targets based on risk assessment results" according to the correlation strength ratio with the node "Select investment targets based on risk assessment results".

[0150] A path attenuation coefficient is introduced during the allocation process. The weight value decreases by the coefficient for each intermediate node the path passes through. The path attenuation coefficient represents the degree of weight loss when data passes through intermediate nodes during propagation. For example, if the path attenuation coefficient is 0.9, the weight value will be multiplied by 0.9 if the path passes through an intermediate node. The attenuation coefficient is dynamically adjusted based on the average network correlation strength. If the average network correlation strength is high, it indicates that the nodes are closely related, and the attenuation coefficient can be appropriately increased; conversely, if the average network correlation strength is low, the attenuation coefficient can be appropriately decreased.

[0151] Step S455: After each round of propagation, record the weight change of each node. When the weight change of all nodes is less than the set threshold for two consecutive rounds, stop propagation. The threshold is determined by multiplying the number of network nodes by the average association strength.

[0152] In this embodiment of the invention, the weight change of each node is recorded after each round of propagation. The weight change represents the difference between the weight value of a node after the current round of propagation and the weight value of the previous round. Propagation stops when the weight changes of all nodes for two consecutive rounds are less than a set threshold. The set threshold is determined by multiplying the number of network nodes by the average association strength. The number of network nodes represents the total number of nodes in the network to which the result depends, and the average association strength represents the average association strength of all edges in the network. By recording the weight changes and setting the threshold, the termination condition of weight propagation can be controlled, ensuring that the weight propagation process converges to a stable state.

[0153] Step S456: The final propagation weight is weighted and fused with the credibility of the node execution result. The credibility is calculated by multiplying the historical execution success rate with the current data integrity assessment. After fusion, the global consistency weight of each node is obtained. The weight value reflects the node's contribution to the network and its reliability level.

[0154] In this embodiment of the invention, the final propagation weight and the reliability of the node execution result are weighted and fused together. The final propagation weight is the node weight value obtained after multiple rounds of weight propagation, and the reliability of the node execution result represents the reliability of the node execution result.

[0155] Credibility is calculated by multiplying the historical execution success rate by the current data integrity assessment. The historical execution success rate represents the proportion of times a task corresponding to a node has been successfully completed in its historical execution history. For example, if the task "Query the risk assessment results of a specified customer" has been successfully completed 8 out of the past 10 executions, then the historical execution success rate is 0.8. The current data integrity assessment indicates whether the data in the current execution result is complete; for example, whether the customer's risk assessment results contain all the necessary information.

[0156] The final propagation weight and the reliability of the node execution result are weighted and fused. For example, a weighted average method can be used to combine the final propagation weight and reliability according to a certain weight ratio. After fusion, the global consistency weight of each node is obtained. The weight value reflects the node's contribution to the network and its reliability level. If a node has a high global consistency weight, it means that the node's result has a greater impact on the entire network and the result is relatively reliable.

[0157] Step S460: Accumulate the weights of all propagation paths to obtain the initial global weight of each node. Adjust the weight allocation ratio based on the credibility of the execution result, increasing the weight ratio of nodes with high credibility, and finally generating a globally consistent weight distribution that reflects the correlation and reliability of the results.

[0158] In this embodiment of the invention, the initial global weight of each node is obtained by accumulating the weights of all propagation paths. Each node may have multiple propagation paths; the weights of these paths are summed to obtain the initial global weight of that node. For example, the node "Query the risk assessment results of a specified customer" may obtain weight values ​​from the starting node through different paths; these weight values ​​are accumulated to obtain the initial global weight. Execution result credibility indicates the reliability of the node's execution result. A node with higher credibility indicates that its result is more reliable, and its weight percentage should be increased. For example, if the execution result credibility of the "Query the risk assessment results of a specified customer" task is high, the weight percentage of that node is appropriately increased based on the initial global weight. Finally, a globally consistent weight distribution reflecting the correlation and reliability of the results is generated. The globally consistent weight distribution shows the weight values ​​of each node in the result dependency network, reflecting the correlation and reliability level between nodes.

[0159] Step S500: Based on the global consistency weight, perform weighted aggregation on the execution results in the result-dependent network to generate a global task processing result that satisfies the task processing request, and send the global task processing result to the user terminal.

[0160] In one embodiment, step S500 may specifically include the following steps S510-S560: Step S510: Implement a bottom-up hierarchical aggregation strategy according to the hierarchical topology of the result-dependent network. Start from the leaf nodes of the network and merge the results layer by layer upwards. Each aggregation layer uses the global consistency weight of the nodes at the current level as the fusion benchmark. The results of nodes with higher weights retain more feature details during the fusion process.

[0161] The hierarchical topology of the result-dependent network illustrates the hierarchical relationships between nodes. Leaf nodes are nodes at the bottom of the network that have no child nodes. For example, in a result-dependent network for portfolio strategy development, the result node corresponding to an atomic task like obtaining basic data for a single investment target might be a leaf node. The network merges execution results layer by layer upwards from the leaf nodes, with each layer's aggregation using the global consistency weight of the current-level node as the fusion benchmark. The global consistency weight reflects the node's contribution and reliability level in the network's results; nodes with higher weights retain more feature details during the fusion process. For instance, when merging the results of two leaf nodes, "Query the historical return of a stock" and "Query the historical return of a bond," if the "Query the historical return of a stock" node has a higher global consistency weight, then more detailed return data features and fluctuation trends from that node's result will be preserved during the fusion process.

[0162] When performing hierarchical aggregation, methods such as weighted average or weighted summation can be used. Taking weighted average as an example, to fuse the results of multiple nodes at a certain level, each node result is multiplied by its corresponding global consistency weight, and then these products are added together and divided by the sum of the global consistency weights of all nodes to obtain the fusion result of that level.

[0163] Step S520: Perform a multi-dimensional consistency assessment during each aggregation process. By comparing the feature vector cosine similarity, numerical fluctuation range overlap, and time series trend consistency between the fusion result and the results of each input node, a fusion quality assessment report is generated. The report includes consistency achievement items, deviation items, and deviation magnitude.

[0164] In this embodiment of the invention, a multi-dimensional consistency assessment is performed during each aggregation process. The cosine similarity of feature vectors measures the similarity between the fusion result and the results of each input node in the feature space. For example, the various features of the investment target (such as rate of return, risk level, liquidity, etc.) are represented as feature vectors, and the cosine similarity between the feature vector of the fusion result and the feature vector of each input node is calculated. The closer the cosine similarity is to 1, the more similar the two are. The overlap of numerical fluctuation ranges compares the overlap between the numerical fluctuation ranges of the fusion result and the results of each input node. For example, regarding the rate of return fluctuation range of the investment target, if the rate of return fluctuation range of the fusion result has a large overlap with the rate of return fluctuation range of a certain input node result, it indicates good consistency in this aspect.

[0165] Time series trend consistency examines whether the trends of change in the fusion result and the results of each input node are consistent over time. For example, observing the price trend of an investment target over a period of time, if the price trend of the fusion result is similar to that of the input node results, it indicates a high degree of time series trend consistency. Based on these multi-dimensional evaluation indicators, a fusion quality assessment report is generated. The report includes consistency achievement items, i.e., indicators that the fusion result and the input node results have reached consistency in certain dimensions; deviation items, i.e., indicators that differ; and deviation magnitude, i.e., the specific degree of deviation. To achieve multi-dimensional consistency assessment and report generation, mathematical calculation libraries and data analysis tools can be used. The cosine similarity of eigenvectors can be calculated using vector calculation functions, the overlap of numerical fluctuation ranges can be determined by comparing numerical intervals, and time series trend consistency can be achieved using time series analysis methods. The evaluation results are compiled into a report to provide a basis for subsequent deviation adjustments.

[0166] Step S530: Initiate a dynamic deviation adjustment mechanism for deviation items in the evaluation report. If the deviation is due to weight allocation, fine-tune the global consistency weight ratio of the current level node. If the deviation is due to data feature conflict, trace the source of the conflict through the feature evolution path and make targeted corrections based on the result features of the source node.

[0167] In this embodiment of the invention, a dynamic deviation adjustment mechanism is initiated for deviation items in the fusion quality assessment report. If the deviation stems from weight allocation, for example, in a certain level of aggregation, an unreasonable global consistency weight setting for a certain node leads to a significant deviation between the fusion result and the result of that node in certain dimensions. In this case, the global consistency weight ratio of the current level node is fine-tuned. For example, the global consistency weight of that node is appropriately increased or decreased, and the level aggregation is re-performed to observe whether the deviation decreases. If the deviation stems from data feature conflicts, for example, there are contradictions between certain features (such as risk level and rate of return) of different investment targets, leading to deviations in the fusion result. The source of the conflict is traced through the feature evolution path, which records the transmission and change process of data features in the result dependency network. For example, starting from the fusion result, the source node that generated the conflicting features is traced along the edges of the result dependency network. Targeted correction is performed based on the result features of the source node. For example, if the source node is "querying the risk assessment data of a certain investment target", and it is found that the risk level in the result of this node conflicts with the results of other related nodes, the risk level is corrected based on the actual situation of the investment target and other reliable data. After correction, perform hierarchical aggregation and consistency assessment again until the deviation meets the requirements.

[0168] Step S540: After completing the aggregation of all levels, a preliminary global result is generated. Feature enhancement processing is performed on the preliminary result, key features retained during the fusion process of each level are integrated, and context information of the task processing request is associated to form a result-context association graph. The graph contains the mapping relationship between result features and request targets.

[0169] Key features are those retained during hierarchical aggregation due to their high global consistency weight among nodes, such as detailed return data of investment targets and trends in risk level changes. Integrating these key features enriches and improves the accuracy of the preliminary results. The contextual information of the task processing request includes the user's risk tolerance, investment objectives, and investment horizon. The features in the preliminary results are correlated with this contextual information to form a graph. The graph contains the mapping relationship between result features and request objectives; for example, the matching relationship between the risk level of recommended investment targets and the user's risk tolerance, and the correlation between the expected return of investment targets and the user's asset appreciation goals. Feature extraction and correlation algorithms can be used. Key features are extracted from the fusion results of each level and correlated with the contextual information of the task processing request to construct the graph.

[0170] Step S550: Based on the output format requirements of the user terminal, the enhanced global results are structurally adapted, and the result-context association graph is converted into a standard response format that includes core conclusions, key evidence, process summary and confidence assessment. The confidence assessment is calculated by averaging the global consistency weights of all participating aggregation nodes.

[0171] The output format requirements of user terminals may vary depending on factors such as device type and user preferences. For example, mobile devices may require a simple and clear format, while computer devices may be able to display more detailed content.

[0172] The core conclusion is a concise summary of the enhanced global results. For example, in a portfolio strategy, the core conclusion might be "Recommended investment in stock A, bond B, and fund C, with investment proportions of 30%, 50%, and 20%, respectively." Key evidence supports the core conclusion with important information such as risk assessment data and return forecasts for the investment targets. The process summary describes the main steps and processes from task processing request to generating the final result, such as how to decompose the task, execute atomic tasks, and fuse results. The confidence assessment is calculated by averaging the global consistency weights of all participating aggregation nodes. The global consistency weights reflect the reliability and contribution of the node results; calculating the average global consistency weights of all participating aggregation nodes yields a comprehensive confidence assessment value.

[0173] Step S560: Encapsulate the global task processing results after structured adaptation, add result generation timestamps, processing node signature chains and data traceability identifiers, and send them to the user terminal through an encrypted transmission channel. At the same time, back up the results and associated aggregation process logs in local distributed storage. The logs contain fusion parameters and deviation adjustment records at each level.

[0174] Please see Figure 3 This is a schematic diagram of the structure of a task collaboration processing device provided in an embodiment of the present invention. Figure 3 As shown, the aforementioned task collaboration processing device 10 may include a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the task collaboration processing device 10 may also include a user interface 1003 and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 3As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0175] exist Figure 3 In the task collaboration processing device 10 shown, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the methods provided in the above embodiments.

Claims

1. A task collaboration processing method using a distributed intelligent agent network, characterized in that, The method includes: The system receives a task processing request sent by a user terminal, performs hierarchical semantic parsing and task dependency analysis on the task processing request, and generates a structured set of atomic tasks. The set of atomic tasks contains multiple atomic task units with execution order constraints and dependency descriptors between each atomic task unit. The atomic task set is input into a preset distributed agent network. Each agent in the network makes collaborative decisions and confirms the execution intention of the atomic task unit based on its own capability description vector and current load status. A dynamic task graph containing task execution paths and agent allocation relationships is constructed based on the decision results and the dependency descriptor. Each intelligent agent executes the corresponding atomic task unit in parallel according to the task allocation relationship in the dynamic task graph. During the execution process, task execution status information and intermediate data output are exchanged in real time through a preset publish-subscribe communication method. Based on the exchange results, a distributed execution trajectory containing timestamps is generated. The execution results of each atomic task unit in the distributed execution trajectory are correlated and fused, and a result dependency network is constructed in combination with the dependency descriptor to calculate the global consistency weight of each result node. Based on the global consistency weight, the execution results in the result-dependent network are weighted and aggregated to generate a global task processing result that satisfies the task processing request, and the global task processing result is sent to the user terminal.

2. The method of claim 1, wherein, The process of receiving a task processing request from a user terminal involves performing hierarchical semantic parsing and task dependency analysis on the task processing request to generate a structured set of atomic tasks, including: Task feature modeling is performed on the task processing request, extracting the task objective description, constraint set and context association information from the task processing request, and constructing a task requirement feature model containing multi-dimensional features, including functional objective features, resource requirement features and time constraint features. Based on the task requirement feature model, a task execution logic chain is constructed, decomposing the high-level task objective into multiple interrelated mid-level task logic units. Each mid-level task logic unit contains the operation sequence for implementing the sub-objective and the definition of input and output interfaces. The operation granularity of the middle-level task logic unit is refined, and the smallest execution unit is identified in combination with the input and output interface definition. Each middle-level task logic unit is decomposed into atomic task units containing specific operation instructions and parameter constraints, resulting in a set of atomic task units. Traverse the set of atomic task units, analyze the source of input parameters and the destination of output parameters of different atomic task units, identify task unit pairs with parameter passing relationships, and generate task execution order constraints based on parameter passing direction; A directed dependency graph is constructed based on the task execution order constraints. In the graph, nodes represent atomic task units, and directed edges represent the execution order dependencies between task units. The node and edge information in the directed dependency graph are structurally integrated to obtain a structured atomic task set containing atomic task unit sequences and dependency descriptors.

3. The method of claim 2, wherein, The construction of the task execution logic chain based on the task requirement feature model includes: Based on the functional target features in the task requirement feature model, a preset domain causal relationship network is queried to obtain the causal association path to achieve the functional target. The causal association path includes multiple event nodes and causal triggering relationships between events. The event nodes in the causal path are mapped to task execution steps, and the logical dependencies between steps are determined based on the causal triggering relationship between events, thus obtaining a preliminary sequence of task execution steps. Combining the resource requirement features and time constraint features in the task requirement feature model, resource-time constraint matching is performed on the preliminary task execution step sequence, and resource occupation time period and time window are allocated to each step; Each task execution step is broken down into a mid-level task logic unit that includes preconditions, execution actions, and postconditions. A time-series correlation analysis is performed on the mid-level task logic units to identify logic unit groups that may be executed in parallel. The feasibility of parallel execution is judged based on the resource occupation period, and parallel execution suggestions are generated. By integrating logical dependencies, resource-time constraints, and parallel execution suggestions, a task execution logic chain containing sequential execution units and parallel execution unit groups is constructed.

4. The method of claim 2, wherein, The process of traversing the set of atomic task units, analyzing the source of input parameters and the destination of output parameters for different atomic task units, and identifying task unit pairs with parameter transfer relationships includes: A parameter hierarchy table is constructed for each atomic task unit. The parameter hierarchy table includes an input parameter column and an output parameter column. The input parameter column records the names and data types of all parameters required for the execution of the task unit, and the output parameter column records the names and data types of parameters generated after execution. Iterate through the output parameter columns of all atomic task units, match the output parameter names and data types with the input parameter columns of other atomic task units, and mark parameter pairs with the same parameter names and compatible data types; The parameter pairs are analyzed to determine the atomic task unit to which the output parameter belongs as the source task unit and the atomic task unit to which the input parameter belongs as the target task unit, and a source-target task unit mapping relationship is established. Determine if there is a time overlap between the generation time of the output parameters of the source task unit and the required time of the input parameters of the target task unit. If there is an overlap, confirm the time validity of the parameter transmission; otherwise, mark it as a potential timing conflict. Based on the analysis of parameter transfer direction and time validity, a parameter transfer path record is generated, which includes the source task unit identifier, the target task unit identifier, the parameter name, and the transfer time window. Summarize all parameter transfer path records, identify complex parameter transfer relationships with multiple input sources or multiple output targets, and generate a list of task unit pairs containing direct and indirect transfer relationships.

5. The method of claim 1, wherein, The process involves inputting the atomic task set into a preset distributed agent network, where each agent, based on its own capability description vector and current load state, makes collaborative decisions and confirms the execution intention of the atomic task units. A dynamic task graph, including task execution paths and agent allocation relationships, is constructed based on the decision results and the dependency descriptor. The atomic task set is received by the coordination node in the distributed intelligent agent network, the atomic task units are prioritized, and a task priority sequence is generated. The prioritization is based on the position of the task unit in the dependency descriptor and the urgency of resource requirements. The task capability requirement query is broadcast to all intelligent agent nodes in the network by the coordinating node. The task capability requirement query includes the operation type, input and output data specifications and execution quality requirements of each atomic task unit. When each intelligent agent node responds to a query, it parses the operation type and compares it with the operation sequence in its own historical execution record. It determines the capability adaptation basis by the operation sequence matching degree, and then calculates the data flow efficiency coefficient by combining the input and output data specifications. The data flow efficiency coefficient is dynamically generated by the ratio of data throughput to format conversion time. It also analyzes the current processing queue length and resource usage ratio to obtain a load margin assessment. After coordinating the collection of capability adaptation and load margin assessments of agent nodes by the coordinating nodes, a decision evaluation system is constructed to determine the adaptation index of each agent and select multiple agents with the best index to form a candidate execution group. The coordination node and candidate execution group conduct multiple rounds of task pre-allocation negotiation. The first round of negotiation determines the resource reservation baseline. The second round of negotiation adjusts the collaboration window with the preceding task unit according to the dependency descriptor. The time sequence feasibility of data transmission is verified by time axis overlap analysis. The final round of negotiation resolves resource competition conflicts. Based on the negotiated resource allocation scheme and timing coordination protocol, atomic task units and agent nodes are connected by directed edges. The edge attributes include data transmission delay thresholds and resource mutual exclusion identifiers. The task execution path topology is constructed to obtain a dynamic task graph.

6. The method of claim 5, wherein, When each intelligent agent node responds to a query, it parses the operation type and compares it with the operation sequence in its own historical execution record. The capability adaptation basis is determined through the operation sequence matching degree, including: The operation type identifier of the atomic task unit is extracted by the intelligent agent node, and the local operation skill graph is queried. The node represents the operation type, the edge weight represents the transfer proficiency between operations, and the skill transfer cost between the current operation and the historically executed operation is calculated through path analysis. The basic execution capability is determined based on the skill transfer cost threshold. If the skill transfer cost is lower than the skill transfer cost threshold, the quality requirement parameters corresponding to the operation type are analyzed. The quality requirement parameters are broken down into precision control indicators and stability indicators. The probability of quality achievement is calculated by the frequency of parameter compliance in historical execution data. For input and output data specifications, the agent node simulates the data flow process, verifies data format compatibility, determines whether format conversion is required through the mapping relationship in the format conversion rule base, calculates the ratio of conversion time to data throughput, and generates a data processing efficiency coefficient. The adaptation degree is comprehensively calculated by considering the skill transfer cost, the probability of achieving quality, and the data processing efficiency coefficient, and a basic value for capability adaptation is generated through weighted calculation. The capability adaptation base value is normalized, and a historical collaboration penalty factor is introduced during the process. If there are records of failure of similar tasks, the adaptation base value is reduced to generate a capability adaptation base that reflects the overall adaptation level, which serves as the core content of the response. The capability adaptation basis is correlated with the current load status. The load status is converted into load margin by the reciprocal of the resource utilization rate. Together, they form the initial response content of the intelligent agent node to the task unit and are sent to the coordination node.

7. The method of claim 5, wherein, After assessing the capability adaptation of agent nodes through coordination nodes and the load margin, a decision evaluation system is constructed, including: The decision evaluation system is initialized for each atomic task unit by coordinating nodes. The first dimension is the capability matching depth, which is quantified by the difference between the capability adaptation base value and the task requirements. The second dimension is the resource slack, which is calculated by the ratio of the available resources in the load slack assessment to the resource requirements of the task. The third dimension is the collaborative response speed, which is determined by the ratio of the time taken for the agent node to respond to the query to the preset response threshold. The capability adaptation basis, load margin assessment and response time data of each agent node are mapped to three assessment dimensions to obtain a multi-dimensional assessment matrix, where each element corresponds to the assessment value of an agent node in a specific dimension. Calculate the difference between the evaluation value of each agent node and the ideal fit state. The ideal fit state is when the evaluation value of each dimension is optimal. If the value of a certain dimension exceeds the required threshold, the overall fit is determined to be unqualified. After sorting the degree of difference, select the agent nodes with the smallest degree of difference as candidate execution groups; Analyze the distribution characteristics of the evaluation values ​​of agent nodes in the candidate execution group. If multiple nodes have similar differences, conflict detection is performed. The detection is achieved by calculating the overlap of resource requirements. Among the nodes with high overlap, the nodes with faster cooperative response speed are retained. The selected candidate execution groups are sorted in ascending order of their degree of difference to generate a comprehensive fit index ranking. The ranking results include the three-dimensional evaluation value of each node and the reasons for the difference, which serve as the basis for decision-making on task pre-assignment.

8. The method of claim 1, wherein, The process involves each agent executing corresponding atomic task units in parallel according to the task allocation relationship in the dynamic task graph. During execution, task execution status information and intermediate data output are exchanged in real time through a preset publish-subscribe communication method, including: Each intelligent agent node locks its own atomic task unit according to the allocation relationship in the dynamic task graph, obtains the operation instructions, parameter constraints and the identifier of the preceding task unit of the atomic task unit, starts the dynamic resource pre-allocation process, reserves processing resources, storage resources and communication bandwidth according to the task complexity, and negotiates the data receiving window with the preceding task node in sync. Send a data subscription request to the agent node corresponding to the preceding task unit marked in the dynamic task graph to establish a dynamic communication link based on task dependency; When the front-end task unit generates intermediate data, it receives the data through the dynamic communication link, performs feature comparison and context association verification on the data, and generates a data quality assessment result. Configure input parameters according to the parameter requirement list, execute the operation logic corresponding to the atomic task unit, adjust the number of execution threads and the size of data processing blocks according to the real-time resource load, continuously collect execution progress, resource consumption and environmental parameters during execution, and generate a time-series state sequence containing multi-dimensional state parameters. The time-series state sequence and intermediate data output are pushed to the distributed message bus. The coordinating node and the intelligent agent node that subscribes to the distributed message bus adopt a differentiated receiving strategy according to their own task priorities. After receiving, they return a receiving feedback containing data integrity and processing suggestions to the publishing node.

9. A task collaboration processing apparatus characterized by comprising: include: processor; And a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 8.

Images