Multi-scene business data intelligent analysis and auxiliary decision method based on AI agent

Abstract

This application discloses a method for intelligent analysis and decision support of multi-scenario business data based on AI agents, belonging to the field of data intelligent analysis technology. It includes acquiring task description data, analysis result types, and a list of capability tags for each AI agent in a business scenario analysis task; decomposing the task description data and analysis result types into multiple sub-task nodes, and calculating the capability compatibility data between each AI agent and the sub-task nodes; comparing the capability compatibility data with a preset compatibility threshold to determine whether a matching completed node or a capability gap node is identified; generating execution sequence configuration data based on the determination data of the matching completed node, and invoking supplementary execution strategies based on the determination data of the capability gap node to obtain supplementary execution result data; integrating the execution sequence configuration data and the supplementary execution result data to obtain and output the decision support result data. This application has the effect of improving the efficiency and reliability of multi-agent collaborative analysis and decision support in multiple scenarios.

Description

Technical Field

[0001] This application relates to the field of data intelligence analysis technology, and in particular to a method for intelligent analysis of multi-scenario business data and auxiliary decision-making based on AI intelligent agents. Background Technology

[0002] As enterprises deepen their digital transformation, the demand for data intelligence analysis is becoming increasingly urgent in diverse business scenarios such as supply chain management, risk control, customer insights, and equipment operation and maintenance. AI agents, with their outstanding capabilities in natural language understanding, anomaly detection, trend prediction, and decision reasoning, are gradually becoming important tools to assist in business analysis across various scenarios. Different AI agents possess different strengths due to differences in training data, algorithm architecture, and optimization goals.

[0003] It is worth noting that different business scenarios naturally have different objectives for analysis results. For example, early warning scenarios prioritize timeliness, prediction scenarios prioritize robustness in trend fitting, and decision-making scenarios prioritize logical completeness and sufficient evidence. When an agent independently handles the needs of multiple scenarios, a mismatch in capabilities can easily occur. For instance, using an agent skilled in numerical prediction for rule validation tasks can lead to an increased false alarm rate; similarly, using an agent lacking attribution capabilities for anomaly root cause analysis can result in superficial analysis conclusions.

[0004] However, existing solutions generally treat each AI agent as an independently callable functional unit, allocating tasks using a fixed workflow. They simply match pre-defined agents based on task type, failing to break down the analysis task into sub-stages. This results in an inability to accurately identify the differentiated requirements of each sub-stage for agent capabilities, leading to capability mismatches between agents and sub-tasks. While some solutions attempt to use capability tags for matching, their matching logic is mostly static keyword comparison, lacking a quantitative evaluation mechanism that incorporates the agent's historical performance. This makes it difficult to adaptively adjust to changes in task scenarios, resulting in consistently low matching accuracy. When a sub-task node experiences a capability gap due to insufficient capabilities of existing agents, existing solutions lack effective gap-filling strategies, causing the analysis chain to break down. The intermediate analysis results from each sub-task node cannot be fully aggregated and integrated, ultimately compromising the reliability of the decision-making results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a method for intelligent analysis and decision support of multi-scenario business data based on AI agents.

[0006] The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents provided in this application includes the following steps:

[0007] Step S1: Obtain the task description data, analysis result type, and capability tag list of each AI agent for the current business scenario analysis task;

[0008] Step S2: Based on the task description data and analysis result type of the current business scenario, identify and decompose the current business scenario analysis task to obtain multiple sub-task nodes; based on the capability tag list of each AI agent, calculate the capability compatibility between each AI agent and the multiple sub-task nodes to obtain capability compatibility data.

[0009] Step S3: Compare the capability matching data with a preset matching threshold to obtain matching status determination data for each sub-task node; if the capability matching data corresponding to a sub-task node exceeds the preset matching threshold, the sub-task node is determined to be a matching completed node; if the capability matching data corresponding to a sub-task node does not exceed the preset matching threshold, the sub-task node is determined to be a capability gap node.

[0010] Step S4: Based on the matching status determination data of the matching completed nodes, generate execution sequence configuration data; based on the matching status determination data of the capability gap nodes, call the preset supplementary execution strategy to perform execution processing on the capability gap nodes to obtain supplementary execution result data; integrate the execution sequence configuration data and the supplementary execution result data to obtain auxiliary decision result data, and output the auxiliary decision result.

[0011] In summary, this application includes at least one of the following beneficial technical effects:

[0012] 1. This application breaks down a complete business analysis task into multiple sub-task nodes with sequential dependencies and identifies corresponding capability requirement tags. It then combines the historical execution accuracy, execution timeliness, and the number and weighted fusion of capability tag matching quantities and weights of each AI agent to calculate capability fit. This allows the computer system to accurately match agents and sub-tasks based on quantitative evaluation during the task scheduling phase, reducing the number of invalid matching attempts. Compared to traditional solutions that only match preset agents based on task type, leading to repeated retries due to capability mismatches, this application's refined sub-task decomposition and multi-dimensional weighted fit calculation reduce the number of agent matching iterations. Furthermore, by integrating historical performance and tag coverage quantitative evaluation, it reduces the analysis interruption rate caused by capability mismatches, improving the task completion rate of multi-agent collaborative analysis. In addition, based on the topological sorting of the sub-task node dependency graph G, it identifies parallel-executable node groups, enabling the computer system to simultaneously launch dependent agents for concurrent execution, shortening the overall response time of multi-agent collaborative analysis.

[0013] 2. This application distinguishes between completed nodes and nodes with capability gaps by comparing capability matching data with a preset matching threshold. It then adaptively and dynamically adjusts the matching threshold based on historical delay and quality failure statistics, enabling the computer system to continuously optimize threshold settings during operation. This reduces the number of subtask node formation delays and execution quality failures caused by improper threshold settings. For nodes with capability gaps, intelligent agent combination invocation or capability expansion strategies are used to include them in the formation execution sequence instead of skipping them, ensuring the topological integrity of the analysis link and reducing the analysis link interruption rate caused by local capability gaps. Furthermore, the application weights and integrates the execution results of all nodes based on sample size and feature completeness, and calculates confidence scores for hierarchical encapsulation based on historical prediction deviation records. This allows the computer system to automatically process outputs according to confidence levels, reducing the invalid calculations and manual review overhead caused by low-confidence auxiliary decision-making results directly entering downstream decision-making stages. This improves the scheduling efficiency and result reliability of the computer system in multi-scenario business data analysis. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a method for intelligent analysis and decision support of multi-scenario business data based on AI agents, according to an embodiment of this application.

[0016] Figure 2 This is the core data flow diagram of the multi-scenario business data intelligent analysis and auxiliary decision-making method based on AI intelligent agents in the embodiments of this application.

[0017] Figure 3 It is the adaptive dynamic adjustment loop diagram of the preset fit threshold in step S3.

[0018] Figure 4 This is a schematic diagram of the dependency relationship graph G between subtask nodes.

[0019] Figure 5 This is a system architecture diagram of intelligent analysis and decision support for multi-scenario business data based on AI agents, according to an embodiment of this application. Detailed Implementation

[0020] The following description, in conjunction with the implementation of the present invention, is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention, and all such modifications and additions should fall within the protection scope of the present invention.

[0021] Example

[0022] This application is applicable to task scenarios in diverse business contexts such as supply chain management, risk control, customer insight, and equipment operation and maintenance, where AI agents need to be scheduled to complete data intelligent analysis and assist decision-making. It is particularly suitable for situations where a single business analysis task involves multiple processing stages, each stage has different requirements for the capabilities of the AI ​​agents, and the system has multiple AI agents with different capability focuses available for invocation. Its core idea is not to treat each AI agent as an independently callable functional unit and allocate tasks as a whole using a fixed process orchestration method, but rather to synchronously obtain task description data, analysis result types, and a list of capability tags for each AI agent after the business analysis task is submitted. First, a primary judgment of task decomposition and capability matching is performed. After identifying capability gap nodes, supplementary execution processing is carried out on these gap nodes through AI agent combination invocation or targeted capability expansion strategies, thereby obtaining a complete execution result covering all sub-task nodes. Based on this, the confidence level of the assist decision-making results is assessed and graded and packaged. With this approach, even if there is no single agent in the current agent pool capable of independently completing all sub-tasks, the system can still obtain a complete analysis chain through refined sub-task decomposition and differentiated gap-filling mechanisms, reducing problems such as analysis interruptions, unreliable results, and repeated task retries caused by agent capability mismatch.

[0023] This application discloses a method for intelligent analysis and decision support of multi-scenario business data based on AI agents.

[0024] Reference Figure 1-5 A multi-scenario business data intelligent analysis and decision support method based on AI intelligent agents includes the following steps:

[0025] Step S1: Obtain the task description data, analysis result type, and capability tag list of each AI agent for the current business scenario analysis task.

[0026] Furthermore, task description data is obtained through natural language parsing tasks, and the relevant business scenario category and source of original business data are identified to obtain scenario task identification data;

[0027] The analysis result types of business scenarios are obtained based on scenario task recognition data;

[0028] The types and completion quality evaluations of the historical tasks executed by the AI ​​agent are obtained, and the types and completion quality evaluations of the historical tasks executed by the AI ​​agent are pre-labeled to obtain a list of capability tags for each AI agent.

[0029] Based on the scene task identification data, analysis result types, and capability tag list, time alignment processing is performed to obtain the basic data for agent scheduling.

[0030] In one specific implementation, step S1 is the information preparation stage of the entire multi-scenario business data intelligent analysis and auxiliary decision-making method. Its goal is to fully acquire the three types of basic information required to drive subsequent steps before the task is formally decomposed and the intelligent agent is matched.

[0031] Specifically, the following example uses an inventory anomaly early warning task in a supply chain management scenario as a typical implementation example. In this example, the current business scenario analysis task is to conduct an early warning analysis of the inventory backlog risk of a regional warehousing center within the next 72 hours. The task description data includes: a task request text submitted in natural language, the business scenario category identifier to which the task belongs, and the data objects involved in the task. Specifically, the task request text submitted in natural language is: Please analyze the probability of inventory backlog for SKU number SKU-2026-XXXX in Warehouse A of East China from August 1st to August 3rd, 2026, and provide the threshold conditions for triggering the early warning and restocking suggestions; the business scenario category identifier is: the inventory early warning subclass in supply chain analysis; the data objects involved in the task are: the warehouse's historical inbound and outbound flow table, real-time inventory snapshot stream, and associated promotional plan table. The analysis result type is an early warning output, which requires outputting the confidence probability of the abnormal event, the triggering conditions, and the corresponding emergency dispatch suggestions.

[0032] When submitting analysis requests, the analysis objectives are usually described in natural language. The natural language description contains multi-dimensional information such as the target of the task, the scope of data involved, the business scenario, and the expected granularity of analysis. It needs to be structured through natural language parsing before it can be used for subsequent task decomposition and agent matching.

[0033] The natural language parsing task employs an intent recognition and slot filling method based on a pre-trained large language model. The specific construction process is as follows: BERT is used as the base pre-trained encoder. Two self-supervised tasks, Masked Language Modeling (MLM) and Next Sentence Prediction (NSP), are performed on a large-scale general corpus to enable the model to learn deep bidirectional language representation capabilities. Based on this, a joint modeling framework is constructed to address the intent recognition and slot filling requirements of business scenarios. This framework shares the same Transformer encoder and connects to two task output branches: the intent recognition branch extracts the hidden states of sentence-level CLS tokens, maps them through a fully connected layer, and outputs the probability distribution of the sentence belonging to each predefined intent category via Softmax; the slot filling branch connects the hidden states of each input token to an independent classifier, using the BIO sequence labeling system to output the slot label for each token. The training phase employs a pre-training + fine-tuning paradigm. Based on the pre-trained BERT model, supervised fine-tuning is performed using manually annotated business statement data that simultaneously labels intent categories and corresponding slot sequences. The optimization objective is a weighted sum of the loss functions for intent recognition and slot filling branches. The core training parameter configuration is: learning rate... to The batch size is 16 to 64, the number of training rounds is 3 to 10, the optimizer is AdamW, and the maximum sequence length is set to 128 to 512 to adapt to the actual length of the business statement. During the inference phase, after the user inputs a natural language description, the model outputs the intent category and the extracted slot key-value pairs in real time, transforming the unstructured natural language into structured data for subsequent subtask node decomposition and capability matching.

[0034] Compared to keyword matching methods based on rule dictionaries, large language models can understand the semantic business meaning rather than just relying on literal keywords, and have a stronger generalization ability for similar tasks with large differences in expression. Compared to traditional text classification models, intent recognition combined with slot filling can extract key entities such as time range, region, and business objects in the task while recognizing the task category, providing more refined structured information for subsequent scene recognition.

[0035] Natural language task description based on user input Analyze and extract the task theme. Core analysis objectives Data objects involved With time range Structured slots are provided, and the current task is mapped to a predefined set of business scenario categories based on the parsing results. Examples include supply chain analysis, risk control, customer insights, and equipment operation and maintenance, which result in scenario category identification. In addition, identifying the sources of raw business data required for the current task. Data acquisition paths, such as database tables, real-time data streams, or historical report files, should be clearly defined. The above analysis results should be summarized to obtain the scene task recognition data. This serves as the contextual basis for subsequently obtaining the analysis result types and capability tag list.

[0036] Continuing with the supply chain management example above, the user-input natural language task description q is: Analyze the inventory backlog risk in Warehouse A of East China over the next three days. Based on historical inbound and outbound data and promotional plans, output a list of SKUs with a backlog probability exceeding 80% and corresponding restocking suggestions. After natural language parsing, the extracted structured slots include: Task Topic =Inventory backlog risk analysis; core analysis objectives = Issue warnings and generate inventory adjustment suggestions; data objects involved ={Historical Inbound / Outbound Transaction Record, Real-time Inventory Snapshot Transaction, Promotional Plan Report}; Time Range =August 1, 2026 to August 3, 2026; Business Scenario Category Identifier =Supply Chain Analysis - Inventory Warning; Source of Original Business Data ={Database table: Real-time data stream: inventory_realtime; file: promotion_plan_2026Q3.xlsx. The resulting scene task recognition data. This provides a clear contextual basis for confirming the type of analysis results and decomposing subtasks.

[0037] The analysis result type describes the expected output form of this business analysis task. Different business scenarios have fundamentally different requirements for the form of analysis results: early warning scenarios require outputting the confidence probability and triggering conditions of abnormal events; prediction scenarios require outputting numerical trend sequences with confidence intervals; root cause analysis scenarios require outputting the ranking of factor contributions and logical causal chains; and decision-making suggestion scenarios require outputting structured action plans and supporting evidence. The importance of this distinction lies in the fact that even if two tasks have similar natural language descriptions, if the analysis result types are different, the requirements for the agent's output capabilities may be completely different. Incorrect result type identification will lead to an overall deviation in the subsequent capability matching direction.

[0038] Based on scene task recognition data Scene category identifier With core analysis objectives Retrieve and confirm the analysis result type of the current task from the predefined analysis result type mapping table. The analysis result type mapping table is indexed by business scenario category and analysis target semantics, storing the standard result types corresponding to each combination. It is pre-built and regularly maintained by domain experts. For standard scenarios with clear correspondences in the mapping table, the corresponding analysis result type is returned directly; for emerging or cross-scenario composite tasks, the type is determined based on... The semantic vectors are compared with the semantic similarity of the result type descriptions in the mapping table, and the one with the highest similarity is taken as the analysis result type for this task. This ensures the accuracy of result type identification in standard scenarios and its robustness in novel scenarios.

[0039] Taking the aforementioned supply chain management scenario as an example, in the predefined analysis result type mapping table, the combination of supply chain analysis - inventory warning and core analysis target warning with generation of transfer suggestions corresponds to the analysis result type. For early warning and decision-making outputs, the standard output format includes: the backlog probability value and confidence interval for each SKU within the next 72 hours; the quantitative threshold conditions for triggering an early warning; and a structured restocking suggestion text, including the source warehouse to be restocked, the target warehouse to be restocked, the suggested restocking quantity, and the basis for the suggestion. If the current task is high-value customer churn prediction in a customer insight scenario, the corresponding analysis result type is a predictive trend output, requiring the output of a time series of customer churn probabilities with confidence intervals. If it is production line failure root cause analysis in an equipment operation and maintenance scenario, the corresponding analysis result type is an attribution explanation output, requiring the output of the contribution ranking of failure factors and the logical causal chain. This mapping table ensures the accuracy of identifying the analysis result type in different scenarios.

[0040] The capability tag list is a structured representation describing the historical performance capabilities of each AI agent across different task types. It serves as the core input for subsequent capability fit calculations on the agent side. The capability tag list does not rely on the subjective statements of agent developers but is generated entirely based on objective performance data from each agent's historical task executions, thus ensuring the objectivity and traceability of capability assessments.

[0041] For each AI agent registered in the system ,in , Given the total number of agents, obtain their historical task execution records and extract two key types of information from the records: one is the distribution of historical task types, i.e., the agents'... The first aspect is the types of sub-tasks undertaken and their corresponding execution frequencies, reflecting the scope of its tasks; the second aspect is the completion quality evaluation, which is the quality score given by downstream nodes or manual review processes after the completion of each historical task. ,in The historical task is assigned a number, reflecting its execution quality level in the corresponding task type.

[0042] Based on the above two types of information, the intelligent agent Perform pre-labeling of capability tags. The labeling logic is as follows: for the intelligent agent... In a certain task type The execution frequency exceeds the preset minimum sample size. And the corresponding average of the completed quality evaluation Exceeding the preset lower quality limit In this case, the task type Labeled as an intelligent agent Effective capability labels, and with As the capability weight of this label Task types with insufficient execution frequency or substandard quality will not be labeled as valid capability tags to avoid erroneous overestimation of agent capabilities due to insufficient sample size or unstable historical performance. After labeling, the agent is obtained. Ability Tag List Each entry consists of a capability label type and a corresponding capability weight. The capability label lists of all M agents are aggregated to obtain a system-level capability label list set. .

[0043] The specific implementation process of the pre-calibration process is as follows: First, for each AI agent in the system... The historical task execution records are structured and parsed to extract the task type identifier for each record. The execution completion timestamp, output result data, and completion quality evaluation given by downstream nodes or manual review. First, construct a standardized data table with task type as the key and all historical quality evaluation sequences under that type as the value. Second, for each task type... Statistical intelligent agents Cumulative execution frequency on this type And calculate the arithmetic mean of all historical quality ratings under this type. Subsequently, a dual threshold determination is performed: if and Then the task type Labeled as an intelligent agent Effective capability labels, and with As the capability weight of this label If any condition is not met, no labeling will be performed. A minimum sample size is preset. The value ranges from 20 to 50, with a preset lower quality limit. The values ​​range from 0.75 to 0.85. Both are calibrated by domain experts based on business fault tolerance requirements before system deployment and are periodically reviewed and adjusted during operation as historical data accumulates.

[0044] Furthermore, to avoid capability assessment bias caused by a single quality evaluation dimension, a time-decrease factor is introduced when calculating capability weights. This applies to task types. The following are various historical quality evaluations According to the execution completion time and the current calibration time time interval Assign decay weight ,in The attenuation coefficient, ranging from 0.01 to 0.05, corresponds to a half-life of approximately 14 to 69 days, ensuring that recent performance has a greater impact on capability weighting than long-term performance. Weighted average quality assessment after time attenuation:

[0045]

[0046] by Alternate arithmetic mean It participates in the above dual threshold determination and serves as the final capability weight. This time-lapse decay mechanism ensures that the capability label list can dynamically reflect the agent's current actual capability level, avoiding labeling distortion caused by excellent early historical performance but recent capability degradation.

[0047] After calibration, the effective capability tags and their weights of each AI agent are categorized and summarized according to task type to generate a structured capability tag list. And record the timestamp of the list's generation. With the historical data time window on which it is based When a new AI agent is introduced into the system, if its historical task execution records are insufficient (i.e., records of all task types), the system may fail to detect it. If the target sample size is not specified, it will be marked as "to be observed" and will be labeled and included in the scheduling candidate pool after the historical sample size meets the threshold requirement.

[0048] Taking the aforementioned supply chain management scenario as an example, suppose there are three AI agents registered in the system: [Agent Name] As an inventory forecasting agent based on a time-series forecasting model, its historical execution record shows that it has been executed 150 times on time-series forecasting tasks, with an average completion quality score of 0.92. Therefore, its capability label is: intelligent agent For an anomaly detection agent based on a rule engine, given that it performs anomaly detection tasks 200 times and has an average completion quality score of 0.88, its capability label is defined as follows: intelligent agent For a decision suggestion generation agent based on a large language model, given that the agent performs the decision suggestion generation task 80 times with an average completion quality score of 0.85, its capability label is defined as follows: If an agent performs a specific task only 5 times, which is below the preset minimum sample size... Even if its average quality rating is 0.90, it will not be labeled as a valid capability label in order to avoid capability assessment bias due to insufficient sample size.

[0049] Scene task recognition data Analysis result types Collection of Ability Tag List These three data sets are generated by different processing stages and may differ in generation time and data version. For example, the capability tag list may come from historical data from the previous scheduling cycle, while the scene task identification data reflects the latest task information at the current moment. If used directly without alignment, the historical performance data of the agent on which the capability assessment is based may not match the actual capability status of the agent at present, affecting the accuracy of the matching results.

[0050] Therefore, time alignment is performed on the above three types of data: based on the current task submission time. To standardize the time baseline, a list of capability tags was inspected. Generation timestamp ,like Exceeding the preset maximum allowable time deviation This triggers the recalibration process of the capability tag list, regenerating it using the latest historical data prior to the current moment. Ensure that the capability tag list and the current task belong to the same valid time window. Align the scene task identification data with the correct time. Analysis result types Collection of Ability Tag List Encapsulated as basic data for agent scheduling .

[0051] Taking the aforementioned supply chain management scenario as an example, at the current task submission time The intelligent agent in the system at 14:00 on July 27, 2026. Capability tag list generation timestamp The time deviation is 10:00 AM on July 27, 2026. The time deviation did not exceed the preset maximum allowable time deviation within the hour. For hours, the existing list of capability tags will be used directly; if The time is 10:00 AM on July 20, 2026, with a time deviation of 168 hours, exceeding... If this occurs, a recalibration process will be triggered, regenerating the data based on the historical execution data of the latest week prior to 14:00 on July 27, 2026. This ensures that the capability assessment reflects the agent's current actual capability status.

[0052] Step S2: Based on the task description data and analysis result type of the current business scenario, identify and decompose the current business scenario analysis task to obtain multiple sub-task nodes; based on the capability tag list of each AI agent, calculate the capability compatibility between each AI agent and the multiple sub-task nodes to obtain capability compatibility data.

[0053] Furthermore, based on the intelligent agent scheduling basic data, the various processing links in the current business scenario analysis task are identified to obtain initial task node data;

[0054] The initial task node data is analyzed for sequential dependencies to obtain multiple sub-task nodes with sequential dependencies.

[0055] Obtain the target output type corresponding to each sub-task node, and based on the target output type corresponding to each sub-task node, determine the capability requirement tag corresponding to each sub-task node, and obtain the capability requirement data of each sub-task node.

[0056] The capability requirements data of each subtask node are compared with the capability tag list of each AI agent to calculate the capability compatibility data between each AI agent and the subtask node.

[0057] Furthermore, obtain the execution accuracy and execution timeliness of the AI ​​agent in historical tasks of the same category as the current subtask node to obtain historical performance data;

[0058] Obtain the number and weight of matches between the AI ​​agent's capability tags and the capability requirement data of the current subtask node to obtain tag matching data;

[0059] The historical performance data and the tag matching data are weighted and fused to obtain the capability fit between the AI ​​agent and the sub-task node;

[0060] Multiple AI agents are compared with the same subtask node to obtain capability fit data. The capability fit data is then normalized and sorted to obtain a capability fit data set.

[0061] In one specific implementation, step S2 undertakes two core tasks: first, it breaks down the complete business scenario analysis task into multiple sub-task nodes with sequential dependencies according to the processing stages, and determines the required capability tags for each sub-task node; second, it quantitatively compares the capability requirements of each sub-task node with the capability tag list of each AI agent to calculate the capability fit data. The quality of step S2 directly determines the granularity of subsequent agent matching. If the task breakdown is too coarse, the differentiated requirements of different sub-stages for agent capabilities will be merged and masked, leading to capability mismatch; if the breakdown is too fine, the complexity of managing the sequential dependencies between sub-task nodes increases sharply, increasing scheduling overhead.

[0062] From the basic data of agent scheduling Read scene task recognition data and the type of analysis results This involves identifying the processing steps for the current business scenario analysis task. While analysis tasks across different business scenarios share a common processing structure, typically including standard steps such as data acquisition and cleaning, feature extraction and transformation, pattern recognition or predictive modeling, result interpretation and attribution, and decision suggestion generation, the specific content and required capabilities of each step vary across different scenarios. Scene category identifier With core analysis objectives The system retrieves a standard processing flow template matching the current scenario from the task processing template library and instantiates the processing steps in the template as the initial task node set for the current task. ,in Given the initial total number of task nodes, we obtain the initial task node data.

[0063] The task processing template library is constructed based on a combination of domain expert knowledge summarization and historical task execution pattern extraction. Specifically, the execution paths of completed historical business analysis tasks are traced back to extract common structures in each task's processing stages, such as data cleaning, feature engineering, pattern recognition, result interpretation, and decision suggestions. The frequency of occurrence and execution order preference of each stage under different scenario categories are statistically analyzed to obtain candidate processing flow patterns. Domain experts review the candidate patterns, confirming the definition of the target output type for each stage, the data dependencies between stages, and the set of capability requirement tags required for each stage. These are then standardized to form template entries. Each template entry... Includes: a collection of applicable scenario categories Core analysis target keyword set Standard processing sequence Target output types for each stage Inter-stage dependency matrix Recommendation capability requirement tag set for each stage And the weighting of accuracy and timeliness in this scenario. For cross-scenario composite tasks or emerging business scenarios, it supports the derivation construction based on similar scenario templates: using the existing template with the highest semantic similarity as the base, domain experts can add or delete specific processing steps to generate new template entries, which are marked as "derived templates" for subsequent task verification.

[0064] Template retrieval employs a two-level matching strategy. The first level is exact matching: based on scene task recognition data. Scene category identifier With core analysis objectives Search the template library for templates that meet the requirements. and The first level is template entries; if a unique match exists, the standard template is returned directly. The second level is semantic similarity matching: when an exact match yields no results or multiple results conflict, the template is returned. The semantic vector of the template is compared with the core analysis target description vector of each template entry in the template library using cosine similarity calculation. The template with the highest similarity exceeding the preset matching threshold is selected. The template is used as the matching result; if the highest similarity is still lower than If this is not the case, a manual review process is triggered, where domain experts temporarily configure a processing flow for this new task combination or create a new template entry. A matching template is then retrieved. Then, its standard processing sequence. Instantiated as the initial set of task nodes for the current task And based on the specific data object of the current task Time range and types of analysis results Instantiate and populate the parameterized fields in the template to adapt the general template to the current specific task.

[0065] The template library is continuously updated and maintained during system operation. Whenever a business analysis task is completed and its execution path differs significantly from the original template, the system records the deviation data between the actual execution path and the original template, summarizes it periodically, and submits it to domain experts for review. New processing modes that are confirmed to be representative are added to the template library as new template entries or revised versions of existing templates, ensuring that the template library can continuously cover emerging analysis needs as business evolves.

[0066] Initial task node set There are sequential execution orders and data dependencies among the nodes in the process. For example, the feature extraction node can only execute after the data cleaning node is completed, and the input of the prediction modeling node depends on the output of the feature extraction node. If these dependencies are not clearly defined, the startup conditions and data transmission paths of each node cannot be determined during the subsequent execution of the intelligent agent formation, leading to confusion or deadlock in the execution chain. A sequential dependency analysis is performed on the initial task node data. Specifically, based on the input data requirements and output data types of each node, each pair of nodes is checked to see if there is a dependency relationship where the output of node A is a necessary input for node B. If so, a directed dependency edge is established from A to B. After completing the dependency check for all node pairs, each node and its dependencies are organized using a directed acyclic graph (DAG). Nodes in the DAG are sub-task nodes, and directed edges represent data flow directions, resulting in multiple sub-task nodes with sequential dependencies. The dependency graph G is used. The reason for choosing a directed acyclic graph instead of a nonlinear sequence to represent the dependency is that some subtask nodes do not have dependencies and can be executed in parallel to shorten the overall analysis time. A linear sequence forces sequential execution, which will artificially introduce unnecessary waiting. A DAG can naturally describe the actual situation where parallel execution and sequential constraints coexist.

[0067] For multiple subtask nodes Each node in ( ), confirm its target output type in the analysis link. This refers to the output format that should be provided upon completion of this node, such as a cleaned dataset, anomaly score sequences, trend predictions, root cause attribution reports, or decision recommendation text. Target Output Type There are two types of sources of confirmation: for the type of analysis results The target output type of the directly corresponding terminal node is... For intermediate processing nodes, the target output type is determined by reverse derivation from the input requirements of downstream nodes.

[0068] Based on the target output type of each subtask node Retrieve and confirm the set of capability requirement tags needed to complete this type of output from the capability tag system. The capability requirement tagging system is consistent with the tagging system used in the agent capability tagging list to ensure semantic correspondence of tags during subsequent comparisons. For standard nodes with clearly defined target output types, capability requirement tags can be directly retrieved from the tagging system mapping table; for composite nodes that cross scenarios, the capability requirement tags are determined based on the target output type. The semantic vectors are compared with the semantic similarity of the descriptions of each tag in the tag system, and the top similarity vectors are selected. Each tag serves as the set of capability requirement tags for that node. The capability requirement tag sets for each node are then aggregated to obtain the capability requirement data for each sub-task node. This is used for subsequent capability compatibility calculations.

[0069] Compatibility Quantitative description of AI intelligent agents In undertaking sub-task nodes The overall competence at that time serves as the direct numerical basis for subsequent matching status determination. The calculation of capability fit comprehensively considers information from two dimensions: first, the agent's actual performance in historical tasks of the same category as the current subtask, reflecting its verified capability level in similar tasks; second, the degree of label coverage between the agent's capability labels and the capability requirement labels of the current subtask, reflecting the direct matching degree between its capabilities and the requirements of the current task.

[0070] For the acquisition of historical performance data, extract the intelligent agent In the subtask node Two key metrics for historical tasks of the same category: execution accuracy and the percentage of historical tasks in the same category whose outputs meet quality standards. And the timeliness of execution of the proportion of similar tasks completed within the specified time limit in the past. The reason for choosing these two metrics is that execution accuracy reflects the quality of the agent's output, while execution timeliness reflects the time it takes to start downstream nodes without delaying their startup in the task scheduling chain. Together, they determine the overall usability of the agent in the actual analysis chain. and We perform a weighted fusion to obtain the historical performance score:

[0071]

[0072] Among them, the weighting coefficient The relative priorities of accuracy and timeliness in the current business scenario determine that early warning scenarios have high requirements for timeliness. Choose a smaller value, such as 0.4; prediction and root cause analysis scenarios have higher accuracy requirements. Take a larger value, such as 0.7. The specific values ​​are pre-configured in the task processing template library according to the scene category, along with the scene category identifier. Retrieve all at once.

[0073] For the calculation of label matching data, the agent Ability Tag List With sub-task nodes Ability requirement tag set Compare and count the number of matched tags.

[0074] ,in For intelligent agents The set of tag types for ability tags is used, and the total matching weight score is obtained by summing the ability weights of the matching tags. The tag matching score, which takes into account both the number of hits and the historical quality weight of the hit tags, is defined as follows:

[0075]

[0076] in For nodes Total number of capability requirement tags For intelligent agents exist The theoretical maximum weighted total score across all demand tags is used for normalization, making... .

[0077] Historical performance score Tag matching score By performing weighted fusion, an intelligent agent is obtained. With sub-task nodes Compatibility:

[0078]

[0079] Among them, the fusion weight This reflects the relative importance of historical performance and label matching in the comprehensive evaluation. When the historical task execution records of each agent in the system are abundant, the historical performance data is more reliable. Take a larger value, such as 0.6; when the system lacks sufficient historical records for newly introduced agents, the weight of label matching should be appropriately increased. Choose a smaller value, such as 0.4, to avoid underestimating the actual capabilities of the new agent due to insufficient historical sample size.

[0080] For all Individual agents and all Capability fit of individual subtask nodes Normalization is performed to make the fit scores of agents under the same subtask node comparable under the same units. Then, the scores are sorted in descending order of fit to obtain the capability fit dataset. .

[0081] Following the previous supply chain management example, the current inventory anomaly alert task, after template retrieval and dependency analysis, is broken down into five sub-task nodes with sequential dependencies: Data cleaning, target output: cleaned dataset, capability requirement label = {data cleaning}; Feature extraction and seasonal decomposition; target output: standardized temporal feature vector; capability requirement labels: {feature engineering, temporal transformation, seasonal decomposition}; dependencies. ; Inventory time-series forecast, target output: 72-hour inventory forecast sequence, capacity requirement tag = {time-series forecast}, dependency ; Anomaly detection, target output: anomaly score sequence, capability requirement label = {anomaly detection}, dependencies ; Early warning decision generation, target output: early warning decision-making output, capability requirement tags = {decision suggestion generation, supply chain domain knowledge}, dependencies The registered agents in the system include: Temporal prediction agent, capability tag: temporal prediction 0.92; Anomaly detection agent, capability tags: anomaly detection 0.88, data cleaning 0.85; Decision-making agent, capability tag: decision-making suggestion generation 0.85; Feature engineering agent, capability tags: Feature engineering 0.90, Temporal transformation 0.88. Calculations show that... and compatibility , and compatibility , and compatibility , and compatibility , and compatibility The quantification result directly determines the matching status of each node in subsequent S3.

[0082] Step S3: Compare the capability matching data with a preset matching threshold to obtain matching status determination data for each sub-task node; if the capability matching data corresponding to a sub-task node exceeds the preset matching threshold, the sub-task node is determined to be a matching completed node; if the capability matching data corresponding to a sub-task node does not exceed the preset matching threshold, the sub-task node is determined to be a capability gap node.

[0083] Furthermore, the highest capability fit for each AI agent in the capability fit dataset is obtained to get the optimal fit dataset;

[0084] Each optimal matching data in the optimal matching data set is compared with a preset matching threshold one by one to determine whether there is an AI agent that exceeds the preset matching threshold, and the matching status judgment data of the subtask node is obtained.

[0085] If it exists, the subtask node is determined to be a matching completion node; if it does not exist, the subtask node is determined to be a capability gap node, and the capability requirement data of the subtask node is recorded.

[0086] Furthermore, the method for confirming the preset fit threshold includes counting the number of times the subtask node formation was delayed due to the fit threshold being set too high in historical tasks, and obtaining delay statistics.

[0087] The number of times subtask execution quality failed due to excessively low matching threshold settings in historical tasks was counted, and statistical data on substandard quality was obtained.

[0088] Based on delay statistics and substandard quality statistics, the preset fit threshold is adaptively adjusted to obtain the updated fit threshold.

[0089] The updated fit threshold will be used as the new preset fit threshold for matching determination of the next subtask node.

[0090] In one specific implementation, step S3 is the judgment and triage stage in the entire agent matching process. (Capability compatibility dataset) For each subtask node, a quantified score of the fit metric for all candidate agents is provided. Step S3 then makes a definitive determination for each subtask node to determine whether an agent with sufficient fit can fulfill the task. The key to accurate determination lies in setting a reasonable fit threshold: a threshold that is too high will incorrectly exclude many capable agents, creating unnecessary capability gaps; a threshold that is too low will classify incapable agents as fully matched, introducing execution quality risks. Therefore, step S3 introduces an adaptive threshold adjustment mechanism based on historical task statistics, allowing the fit threshold to be continuously optimized as system operating experience accumulates.

[0091] Data set of capability fit Each subtask node in From all The optimal fit for a node is obtained by extracting the highest fit score from the fit scores of each agent for that node.

[0092]

[0093] Optimal fit This represents the subtask node in the current agent pool with the closest capabilities. The highest degree of matching that a demanding agent can achieve is a sufficient statistical measure to determine whether the system's current capabilities can cover the demand of that node. (This will be repeated in the original text.) The optimal fit order of each subtask node is used to obtain the optimal fit dataset. .

[0094] The optimal fit data set Optimal fit of each node With preset fit threshold Compare and judge one by one: If Then determine the subtask node. There exists a capable intelligent agent that will Mark the node as a completed match and record the corresponding optimal agent number. As the pre-selected execution agent for this node; if If no node is found in the current pool of agents that is fully capable of performing the subtask, then it is determined that there is no such node in the current pool of agents. The intelligent agent will Mark the node as a capacity gap node and record the capacity requirement data for that node. Best fit with the current This is used to formulate subsequent supplementary execution strategies. The judgment results of all sub-task nodes are summarized to obtain the matching status judgment data for each sub-task node. ,in .

[0095] Preset fit threshold initial value The system is calibrated based on historical execution data distribution during the initial deployment phase, rather than simply relying on experience. Specifically, data evaluated during the system trial operation phase is collected. Each historical subtask node, based on the minimum business quality standard As a boundary, the compatibility of capabilities at each historical node is determined. Divided into positive sample set With negative sample set The execution quality meets the standards. The execution quality did not meet the standards. .

[0096] Calculate the mean of the ability fit of the positive sample set with standard deviation Take the baseline initial threshold ,in For scenario risk tolerance coefficient: timeliness is prioritized for early warning scenarios. For prediction and root cause analysis scenarios, accuracy is prioritized. Up to version 1.5, the highest requirement for logical integrity in decision-making suggestion scenarios is... Up to version 2.0. If the negative sample set is not empty, then... With negative sample mean The weighted median is used as the correction benchmark to ensure that the initial threshold is located within a reasonable boundary between the positive and negative sample distributions.

[0097] Further introduce scene correction coefficient Adapting to the baseline accuracy requirements of different business scenarios: For example, in the supply chain management scenario... = 1.00, risk control scenario takes = 1.10, Customer Insight Scenario Taking = 0.95, taken in equipment operation and maintenance scenarios = 1.05. Initial threshold after scene correction. Within the permitted range Internal truncation serves as a preset fit threshold during system startup. The initial value is used for matching the first sub-task node. As the system runs, this initial value is then incorporated into the matching process based on delay statistics. Statistics on substandard quality An adaptive dynamic adjustment process is implemented, forming a threshold management system based on data calibration, scenario correction, and online optimization. This is based on delay statistics. Statistics on substandard quality For the preset matching threshold Perform adaptive adjustments:

[0098]

[0099] in To adjust the step size and control the magnitude of each threshold update, take... If the value is too large, the threshold will oscillate repeatedly between the two directions and fail to converge. This fixed step size has been calibrated experimentally under normal system load to ensure that the single threshold adjustment amplitude is stable within the total dynamic range of the threshold. Within this range, to prevent oscillations and ensure convergence speed. To prevent smoothing terms with a denominator of zero, a smaller positive number is used, such as... The adjustment logic of this formula is as follows: when the number of substandard quality events exceeds the number of formation delays, it indicates that the current threshold is too low and too many substandard agents have been accepted, requiring an increase in the threshold; when the number of formation delays exceeds the number of substandard quality events, it indicates that the current threshold is too high and too many actually usable agents have been excluded, requiring a decrease in the threshold; when the two are close, the adjustment amount approaches zero, and the threshold remains stable. The adjusted fit threshold. Within the allowable threshold range The threshold is truncated to prevent it from being adjusted to an meaningless extreme value, and the updated threshold is used for matching determination of the next subtask node.

[0100] Taking the aforementioned supply chain management example, a preset fit threshold is set. After initial calibration and scene correction, the value is 0.82. The optimal fit of each of the five subtask nodes is compared one by one: Optimal fit Once a node is determined to be a completed match, a pre-selected intelligent agent is executed. ; Optimal fit If a node is identified as a capacity gap, its capacity requirement data will be recorded. ={Feature Engineering, Temporal Transformation, Seasonal Decomposition}; Optimal fit Once a node is determined to be a completed match, a pre-selected intelligent agent is executed. ; Optimal fit Once a node is determined to be a completed match, a pre-selected intelligent agent is executed. ; Optimal fit If a node is identified as a capacity gap, its capacity requirement data will be recorded. ={Decision suggestion generation, supply chain domain knowledge}. This yields matching status determination data. ,in and As intermediate and terminal gap nodes, excluding them from the formation sequence will directly lead to Due to lack The output feature vector cannot be started, so it must be included in the formation and a supplementary execution strategy must be triggered later.

[0101] Step S4: Based on the matching status determination data of the matching completed nodes, generate execution sequence configuration data; based on the matching status determination data of the capability gap nodes, call the preset supplementary execution strategy to perform execution processing on the capability gap nodes to obtain supplementary execution result data; integrate the execution sequence configuration data and the supplementary execution result data to obtain auxiliary decision result data, and output the auxiliary decision result.

[0102] Furthermore, confirm the matching completion nodes recorded in the matching status judgment data of the sub-task nodes, sort the matching completion nodes according to the sequential dependency relationship between the sub-task nodes, and obtain the formation order data set.

[0103] The execution sequence configuration data set is obtained by matching the subtask content of the AI ​​agent corresponding to the completed node in the configuration formation order data set, and matching the data transmission relationship between the completed node and the adjacent node.

[0104] Based on the execution sequence configuration data set, the node order is confirmed, and the AI ​​agent corresponding to each matched node is controlled to execute the assigned sub-tasks in the order of the nodes. The intermediate analysis results generated by each node are passed to the next node along the formation order data to obtain the formation execution result data.

[0105] Furthermore, the capability gap nodes and corresponding capability requirement data recorded in the matching status determination data are confirmed;

[0106] Based on the capability requirement data, determine whether there is a combination of AI agents that can cover the capability requirement data among all AI agents, and confirm the supplementary execution strategy; if there is, use the calling strategy of the combination of AI agents that can cover the capability requirement data as the supplementary execution strategy for the capability gap node; if there is not, use the strategy of expanding the capabilities of the AI ​​agent with the highest compatibility among the AI ​​agents matched with the capability gap node as the supplementary execution strategy for the capability gap node.

[0107] The capability gap nodes are processed based on the aforementioned supplementary execution strategy to obtain supplementary execution result data for the capability gap nodes.

[0108] Furthermore, based on the formation execution result data and supplementary execution result data, the sample size and feature completeness of the original data were confirmed;

[0109] Based on the original data sample size and feature completeness, the formation execution result data and supplementary execution result data are weighted and integrated to obtain auxiliary decision-making result data;

[0110] Identify prediction deviation records for similar historical tasks, and calculate the confidence score of the auxiliary decision-making result data based on these records; compare the confidence score with a preset multi-level confidence interval to determine the confidence level of the auxiliary decision-making result data; classify and encapsulate the auxiliary decision-making result data according to the confidence level and output it.

[0111] In one specific implementation, step S4 is the execution and integration phase of the entire method, divided into two sub-processes: the formation execution of all sub-task nodes, and the integrated output of the formation execution results. Matching completed nodes are directly executed by the corresponding AI agent, while capability gap nodes are executed by a pre-determined supplementary execution strategy. The design rationale is that in real-world business scenarios, it is almost impossible for all sub-task nodes to find ideal matching agents; capability gaps in some nodes are normal rather than abnormal. If capability gap nodes are excluded from the formation execution sequence, their downstream nodes will be unable to execute due to a lack of necessary input, leading to an interruption in the analysis chain. Therefore, the system must include capability gap nodes in the topological sorting and data transmission chain of the sub-task dependency graph, ensuring the completeness of input data for each node and the integrity of the analysis chain even when gap nodes and completed nodes coexist.

[0112] Data determined from matching status Extract all subtask node sets This includes the set of subtask nodes marked as completed matches. and the set of subtask nodes marked as capability gaps Based on the subtask node dependency graph G constructed in step S2, perform a process on all subtask nodes. Each node in the process undergoes topological sorting to obtain the execution order of nodes that satisfies the sequential dependency constraints, i.e., the queuing order data. The reason for covering all subtask nodes in the topological sorting is that although the capability gap node has no direct matching by a single agent, it still undertakes specific data processing functions as a node in the DAG, and its output is a necessary input for downstream nodes. If the capability gap node is excluded from the topological sorting, the downstream matching completed nodes will be unable to start due to the lack of input, resulting in a break in the analysis link.

[0113] The result of topological sorting is a partial order: for node pairs with direct dependencies, the upstream node must be executed before the downstream node; for nodes without dependencies, parallel execution is allowed to shorten the overall analysis time.

[0114] Based on the formation order data, for each subtask node Configure its execution content and data transfer relationships between nodes: For nodes that have completed matching Configure the corresponding intelligent agent The subtask content includes a description of the specific analysis operations the agent needs to perform, input data specifications, and output data specifications; for capability gap nodes... Configure corresponding supplementary execution strategies, including strategy type, execution subject, and output data specifications, to ensure that this node can produce intermediate analysis results with the same structured format as the matched nodes during queue execution. Data transfer relationships define the relationships between each node. How the output data is passed to its downstream neighboring nodes as input includes the data type, field mapping relationships, and format conversion rules. The formation order, the execution content of each node, and the data transmission relationships are encapsulated together to obtain the execution sequence configuration data. .

[0115] Based on execution sequence configuration data The subtask nodes are executed sequentially according to the topological sorting: for nodes that have completed matching, the corresponding agent is activated to execute its assigned subtask; for nodes with capability gaps, a pre-determined supplementary execution strategy is invoked for processing, and the resulting supplementary execution results are used as the intermediate analysis results for that node. For node groups that can be executed in parallel, the corresponding execution units are activated synchronously and executed concurrently. The next layer of dependent nodes is activated only after all nodes in the group have completed their execution. After each node has completed its execution, the resulting intermediate analysis results are... The data is passed down along the data transmission relationship to downstream nodes as input data. As each node in the formation sequence completes sequentially, the intermediate analysis results converge layer by layer along the dependency graph G, ultimately yielding the formation execution result data covering all subtask nodes. .

[0116] Generate execution sequence configuration data Previously, supplementary execution strategies needed to be determined in advance for each capability gap node. Data was determined based on the matching status. Extract the set of all subtask nodes marked as capability gaps. And the capacity requirement data recorded at each gap node. .

[0117] For each capacity gap node First, determine whether there are labels in the current agent pool that can cover all the capability requirements of this node through combination. The combination of intelligent agents. The specific judgment logic is as follows: iterate through the capability tag sets of each intelligent agent in the agent pool, and check if there exists a union of the capability tag sets of several intelligent agents that can cover... All tags in, i.e., do they exist? Make The search process begins with the agent with the highest capability label coverage and greedily expands, prioritizing the agent that can add the most uncovered capability labels each time, until complete coverage is achieved or exhaustive search of all combinations fails. The reason for using a greedy search instead of exhaustive search of all agent combinations is that when the number of agents... When the subsets are large, the computational complexity of exhaustively searching for all subsets is O(n log n). This is unacceptable in real-time task scheduling scenarios; greedy search has a provable approximation ratio guarantee for the cover set problem, and can achieve... We find a near-optimal coverage combination with a complexity that meets real-time requirements.

[0118] If it exists that can completely cover it. Combination of intelligent agents If the combined strategy is invoked, it will be identified as the supplementary execution strategy for the gap node: the capability requirement label will be used. The assignment is based on the coverage label range of each agent. Each agent is responsible for the sub-requirement portion covered by its capability label. After each agent executes independently, the analysis results are merged to jointly complete the gap node. The analysis task.

[0119] If there is no way to completely cover it The combination of intelligent agents then adopts a capability expansion strategy: from the capability compatibility data set Extracting gap nodes The most compatible intelligent agent ,Right now The corresponding intelligent agent undergoes targeted capability extension processing. Specific methods of capability extension include: […]. Injection and The system does not cover the specific knowledge or tool access capabilities corresponding to the tags, thus expanding its capabilities to include the range of needs for handling gap nodes.

[0120] The injection of specialized knowledge is achieved using a retrieval-enhanced RAG (Research-Based Gaining) mechanism. Specifically, this is implemented for capability gap nodes. Ability requirement tag set The agent was not used A subset covered by existing capability tags Retrieve a set of specialized knowledge documents from the domain knowledge base that semantically match the subset. The retrieval process employs semantic similarity matching based on vector embedding: ... The semantic vectors of each tag in the document are compared with the semantic vectors of pre-stored document blocks in the knowledge base using cosine similarity calculation, and the similarity scores are ranked. =5 document blocks During the execution of the gap node When performing analysis tasks, The specialized knowledge document in the context is concatenated with the original task input as a context prefix to form an enhanced input prompt, enabling the agent to... Without altering the underlying model parameters, this method temporarily acquires the domain knowledge required to handle the gap node. The advantage of this RAG injection approach is that it eliminates the need for time-consuming and resource-intensive retraining or fine-tuning of the agent; it achieves instant capability adaptation simply through context expansion. Furthermore, the injected knowledge precisely corresponds to the needs of the current gap node, avoiding interference from irrelevant knowledge.

[0121] The injection of the tool's calling capability is implemented using a function registration and dynamic binding mechanism. Specifically, for The system requires capability tags that support specific computing tools or external interfaces. For example, time series decomposition requires calling the STL decomposition tool, association rule mining requires calling the Apriori algorithm library, and geospatial analysis requires calling the GIS service interface. The system maintains a registry of tool functions. Each utility function The corresponding capability labels, input parameter formats, and output data formats are clearly labeled. This is done to determine the capabilities of the intelligent agent. When expanding capabilities, from Search coverage A subset of utility functions Furthermore, the calling interface, parameter specifications, and return format descriptions of this subset of tool functions are dynamically bound to the intelligent agent in a structured description format. The execution environment of the intelligent agent. At the execution gap node When performing analysis tasks, if it is identified that a need For an operation corresponding to a certain capability tag, the system automatically generates a call request based on the dynamically bound tool description. After system verification, the corresponding tool function is executed, and the returned result is incorporated into the analysis process. This dynamic binding mechanism enables the intelligent agent to... Without altering its own algorithm architecture, the range of executable operation types can be expanded by calling external tools.

[0122] After the capability expansion is completed, the system records the set of specialized knowledge documents injected during this expansion. and the bound subset of utility functions , and gap node Execution result Related storage. If subsequent historical task backtesting evaluations show that the extended agent performs better on similar gap nodes... If the threshold for valid capability tags is reached, the corresponding capability tag will be... Formal inclusion in intelligent agents Ability Tag List This enables the transformation from temporary capability expansion to permanent capability accumulation.

[0123] After the capability expansion is completed, it will be handled by the expanded version. Execute the gap node separately The analysis task. The reason for choosing the agent with the highest fit for capability expansion instead of random selection is that the agent with the highest fit has the closest needs to the gap node in terms of existing capabilities, requires the smallest increment of capabilities, has the lowest expansion cost, and has the most guaranteed execution quality after expansion.

[0124] The confirmed supplementary execution strategy is bound to the capability gap node and incorporated into the execution sequence configuration data. The execution content field of this node is used for the formation to execute.

[0125] The processing logic for capability gap nodes in the formation execution is as follows: when the formation execution order reaches this node, the pre-bound supplementary execution strategy is invoked; if it is an agent combination invocation strategy, the capability requirement tag is added. The tasks are assigned according to the coverage label range of each agent. Each agent is responsible for the sub-requirement part covered by its capability label. After each agent executes independently, the analysis results are merged to jointly complete the analysis task of the gap node. The resulting merged result serves as the intermediate analysis result for that node. If it is a capability expansion strategy, then it will be directed to the agent with the highest compatibility. Injection and For nodes whose labels are not covered by the algorithm, the extended agent will perform the analysis task for that missing node independently, and the result will serve as the intermediate analysis result for that node. The intermediate analysis results The data is then passed down to downstream nodes along the data transmission relationship to ensure that downstream nodes receive complete input data.

[0126] After all subtask nodes have been executed, the formation execution result data is available. The system already naturally covers all execution results from matched nodes and capability gap nodes, eliminating the need to merge them with separate supplementary execution result data. Before outputting auxiliary decision-making results, the system first considers the original data sample size of each node. With feature completeness Confirm the data quality weights of each execution result. The more abundant the sample size and the higher the feature completeness of the node's execution results, the higher their weight is assigned during integration. This ensures that the results of sub-tasks with better data quality contribute more to the final decision support. Feature completeness is a key factor in this process. Defined as the proportion of valid feature fields in the output of this node to the total number of required feature fields. The results of all nodes are weighted. Weighted aggregation is performed to obtain data for decision support. .

[0127] Before outputting the auxiliary decision-making results, it is necessary to... Conduct a confidence assessment to inform decision-makers of the reliability of the current decision-making aids. (Based on scenario category identification.) Extract prediction deviation records from the execution records of the same historical tasks, calculate the statistical distribution of the deviation between the auxiliary decision-making results and the actual business results in similar historical tasks, and then use the current... Based on the probability density estimate of the resulting features under this distribution, the confidence score is calculated. Confidence scores Compared with the preset multi-level confidence interval, The confidence levels of the decision support results are compared and determined by classifying them as high confidence, medium confidence ([0.5, 0.8)) and low confidence ([0, 0.5)). The decision support results are then classified and packaged according to their confidence levels. High confidence results are directly output as complete decision suggestions. Medium confidence results are accompanied by uncertainty explanations and suggestions for manual review. Low confidence results are marked with data quality risks and their direct use in decision-making is restricted. Finally, the decision support results are output.

[0128] Taking the aforementioned supply chain management example, all sub-task nodes Perform topological sorting based on DAG dependencies to obtain the formation execution order. . The node corresponding to the completed match Perform data cleaning, taking the original inbound / outbound transaction log and inventory snapshot stream as input, and outputting the cleaned dataset. And transmitted to ; For nodes with capacity gaps, the pre-determined supplementary execution strategy is to target the node with the highest compatibility. Capability expansion: via the RAG mechanism Inject seasonal decomposition-specific knowledge documents, from the expanded Perform feature extraction and seasonality decomposition, input Output standardized time series feature vectors And transmitted to ; Depend on To perform inventory time-series forecasting, input... Output 72-hour inventory forecast series And transmitted to ; Depend on Perform anomaly detection, input Output anomaly scoring sequence And transmitted to ; For nodes with capacity gaps, the supplementary execution strategy is to... Capability expansion: Dynamically integrate supply chain inventory early warning rule engine tools and inject supply chain domain knowledge, resulting in an expanded system. Execute early warning decision generation, input The output includes an early warning decision text containing the backlog probability, trigger threshold, and restocking suggestions. .

[0129] After all nodes have completed execution, the formation execution result data is available. All subtasks have been covered. The integration weights were determined based on the original data sample size and feature completeness of each node. to For nodes that have been matched, all feature completeness scores are above 0.90 and are assigned higher weights. and Due to capability expansion, the feature completeness levels are 0.85 and 0.80 respectively, and although they are assigned moderate weights, they are still included in the integration. The weighted integration yields the auxiliary decision-making result data. Confidence scores are further calculated based on prediction deviation records of similar historical inventory early warning tasks. It is in the high confidence interval. Therefore, The system performs high-confidence hierarchical packaging and directly outputs complete early warning decision recommendations, including: a list of SKUs in Warehouse A of East China with a backlog probability of more than 80% in the next 72 hours, the quantitative threshold conditions for triggering the early warning, and specific quantity and route recommendations for transferring goods to the target warehouse.

[0130] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0131] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0132] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents, characterized in that: Includes the following steps: Step S1: Obtain the task description data, analysis result type, and capability tag list of each AI agent for the current business scenario analysis task; Step S2: Based on the task description data and analysis result type of the current business scenario, identify and decompose the current business scenario analysis task to obtain multiple sub-task nodes; Based on the capability tag list of each AI agent, calculate the capability compatibility between each AI agent and the multiple sub-task nodes to obtain capability compatibility data. Step S3: Compare the capability matching data with a preset matching threshold to obtain matching status determination data for each sub-task node; if the capability matching data corresponding to a sub-task node exceeds the preset matching threshold, the sub-task node is determined to be a matching completed node; if the capability matching data corresponding to a sub-task node does not exceed the preset matching threshold, the sub-task node is determined to be a capability gap node. Step S4: Based on the matching status determination data of the matching completed nodes, generate execution sequence configuration data; based on the matching status determination data of the capability gap nodes, call the preset supplementary execution strategy to perform execution processing on the capability gap nodes to obtain supplementary execution result data; The auxiliary decision result data is obtained by integrating the execution sequence configuration data and the supplementary execution result data, and the auxiliary decision result is output.

2. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain task description data through natural language parsing, identify the business scenario category and the source of the original business data, and obtain scenario task identification data; Step S12: Obtain the analysis result type of the business scenario based on the scenario task recognition data; Step S13: Obtain the types and completion quality evaluations of the historical tasks executed by the AI ​​agent, and pre-calibrate the types and completion quality evaluations of the historical tasks executed by the AI ​​agent to obtain a list of capability tags for each AI agent; Step S14: Perform time alignment processing based on scene task identification data, analysis result type, and capability tag list to obtain basic data for agent scheduling.

3. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 2, characterized in that, Step S2 includes the following steps: Step S21: Based on the intelligent agent scheduling basic data, identify each processing link in the current business scenario analysis task to obtain initial task node data; Step S22: Perform a sequential dependency analysis on the initial task node data to obtain multiple sub-task nodes with sequential dependencies; Step S23: Obtain the target output type corresponding to each sub-task node, and based on the target output type corresponding to each sub-task node, determine the capability requirement label corresponding to each sub-task node, and obtain the capability requirement data of each sub-task node. Step S24: Compare the capability requirement data of each subtask node with the capability tag list of each AI agent, and calculate the capability compatibility data between each AI agent and the subtask node.

4. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 3, characterized in that, Step S24 includes the following steps: Step S241: Obtain the execution accuracy and execution timeliness of the AI ​​agent in historical tasks of the same category as the current subtask node, and obtain historical performance data; Step S242: Obtain the number of matches and the matching weights between the AI ​​agent's capability tags and the capability requirement data of the current subtask node, and obtain the tag matching data; Step S243: Perform weighted fusion processing on the historical performance data and the tag matching data to obtain the capability fit between the AI ​​agent and the sub-task node; Step S244: Compare multiple AI agents with the same subtask node to obtain capability matching data, normalize and sort the capability matching data to obtain a capability matching data set.

5. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 4, characterized in that, Step S3 includes the following steps: Step S31: Obtain the highest capability fit for each AI agent in the capability fit data set to obtain the optimal fit data set; Step S32: Compare each optimal matching data in the optimal matching data set with a preset matching threshold one by one to determine whether there is an AI agent that exceeds the preset matching threshold, and obtain the matching status judgment data of the subtask node. Step S33: If it exists, determine that the subtask node is a matching completion node; if it does not exist, determine that the subtask node is a capability gap node, and record the capability requirement data of the subtask node.

6. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 5, characterized in that, The methods for confirming the preset compatibility threshold include: The number of times the subtask node formation was delayed due to excessively high matching threshold settings in historical tasks was counted to obtain delay statistics. The number of times subtask execution quality failed due to excessively low matching threshold settings in historical tasks was counted, and statistical data on substandard quality was obtained. Based on delay statistics and substandard quality statistics, the preset fit threshold is adaptively adjusted to obtain the updated fit threshold. The updated fit threshold will be used as the new preset fit threshold for matching determination of the next subtask node.

7. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 6, characterized in that, Based on the matching status determination data of the matched nodes, the execution sequence configuration data is generated, including the following steps: Confirm the matching completion nodes recorded in the matching status judgment data of the subtask nodes, sort the matching completion nodes according to the sequential dependency relationship between all subtask nodes, and obtain the formation order data set. The subtask content of each subtask node corresponding to the AI ​​agent in the configuration formation order data set, as well as the data transmission relationship between each subtask node and adjacent nodes, are used to obtain the execution sequence configuration data set; among them, the subtask content of the AI ​​agent corresponding to the matching completed node is configured, and the corresponding supplementary execution strategy is configured for the capability gap node. Based on the execution sequence configuration data set, the node order is confirmed. For nodes that have completed matching, the corresponding AI agent is controlled to execute the assigned sub-tasks according to the node order. For nodes with capability gaps, the corresponding supplementary execution strategy is invoked for execution processing. The intermediate analysis results generated by each sub-task node are passed to the next node along the formation order data to obtain the formation execution result data.

8. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 7, characterized in that, The determination of the supplementary execution strategy corresponding to the capability gap node includes the following steps: Confirm the capability gap nodes and corresponding capability requirement data recorded in the matching status determination data; Based on the capability requirement data, determine whether there is a combination of AI agents among all AI agents that can cover the capability requirement data. If such a strategy exists, the invocation strategy of the AI ​​agent combination that can cover the capability requirement data will be used as the supplementary execution strategy for the capability gap node; if such a strategy does not exist, the strategy of expanding the capabilities of the AI ​​agent with the highest compatibility among the AI ​​agents matched with the capability gap node will be used as the supplementary execution strategy for the capability gap node.

9. The method for intelligent analysis and decision support of multi-scenario business data based on AI intelligent agents according to claim 8, characterized in that... The auxiliary decision-making result data is obtained and output based on the formation execution result data, including the following steps: Based on the formation execution result data and supplementary execution result data, the original data sample size and feature completeness were confirmed; Based on the original data sample size and feature completeness, the formation execution result data and supplementary execution result data are weighted and integrated to obtain auxiliary decision-making result data; Identify prediction deviation records for similar historical tasks, and calculate the confidence score of the auxiliary decision-making result data based on these records; compare the confidence score with a preset multi-level confidence interval to determine the confidence level of the auxiliary decision-making result data; classify and encapsulate the auxiliary decision-making result data according to the confidence level and output it.

Images