Method and system for constructing post agent based on employee work flow extraction

Abstract

The application provides a post intelligent agent construction method and system based on employee work flow extraction, comprising: collecting real work flow data of employees in a certain post, including: complete conversation records, continuous task steps, tool calling chains, parameter configurations and operating system interaction behaviors generated in the actual office, operating system and tool calling process; preprocessing the collected real work flow data of employees in a certain post; carrying out process disassembly, skill extraction and executable rule extraction on the pretreated real work flow data to generate structured post skills and process rules; encapsulating the generated structured post skills and process rules into standardized Skill units; the standardized Skill units can be directly loaded into an intelligent agent framework; when employees in a certain post perform tasks, the intelligent agent automatically identifies the current task, calls the corresponding standardized Skill unit, and realizes process guidance, template filling and tool calling.

Description

Technical Field

[0001] This invention relates to the field of enterprise digitalization and artificial intelligence application technology, specifically to a method and system for constructing job-specific intelligent agents based on employee workflow extraction. Background Technology

[0002] In enterprise digitalization and intelligent office scenarios, professional positions such as software development and data governance involve complex business processes, numerous scenario branches, and dense system interaction links. The core business knowledge of these positions has long relied on the experience accumulated by senior employees. Existing methods for building intelligent agents for these positions mostly rely on manually sorting out business processes and manually configuring business rules and skill templates. This is not only costly and time-consuming, but also prone to overlooking real operational details and failing to adapt to employees' actual work habits. As a result, the generated intelligent skills are coarse-grained, lack versatility, and have low adaptability for implementation.

[0003] Meanwhile, traditional intelligent solutions adopt static and fixed rule systems, which cannot perceive business process iterations, employee operating habit updates and new business scenarios in real time. They lack the ability to automatically mine processes, extract rules and continuously iterate, making it difficult to adapt to the dynamic changes in job requirements. Ultimately, this results in industry pain points such as the disconnect between job intelligent assistance capabilities and real business, poor automation implementation, and difficulty in standardizing and reusing enterprise job knowledge.

[0004] Patent document CN121436927A (application number: 202511703010.6) discloses a method, device, and server for implementing an intelligent agent for employee positions, involving the field of artificial intelligence technology. The method includes: using a constructed intelligent agent for employee positions to capture multi-source data on tasks from various systems and construct a job knowledge graph; extracting key information from each multi-source data for task analysis and matching, matching the corresponding job profile based on the associated person, and determining the processing subject; for sub-tasks with clear rules, automatically processing them by calling a rule engine and automated execution module through the intelligent agent for employee positions; for tasks requiring manual processing, pushing them to the corresponding employee's visual workbench through a collaborative push module and setting status tracking; and monitoring the processing status in real time, triggering warnings for tasks that fail to execute or are about to expire. This patented process relies on manual pre-setting and cannot automatically extract skills and rules from employees' actual workflows. Furthermore, it lacks a closed loop of "employee practice - effect evaluation - iterative optimization," preventing the intelligent agent from continuously evolving with business changes; it does not deeply adapt to professional positions such as software development and data governance, resulting in coarse skill granularity that is difficult to cover complex tasks.

[0005] Patent document CN120317282A (application number: 202510800093.4) ​​discloses a method and system for building and training AI agents in a work interaction scenario. The method includes: identifying the access positions of the AI ​​agent; configuring the workflow of each access position; generating a workflow chain; clustering the access positions into several categories; collecting initial data from each category; and configuring attribute information, wherein the attribute information includes at least: type and source; and creating branch nodes corresponding one-to-one with each category. This patent's workflow chain requires manual configuration and does not utilize daily employee input for automatic generation; skill extraction relies on static data and is not optimized based on actual execution results; and it does not form a hierarchical architecture from Skill to Job Agent to Organizational-level Digital Assistant. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for constructing job-specific intelligent agents based on employee workflow extraction.

[0007] A method for constructing a job-specific intelligent agent based on employee workflow extraction, provided by the present invention, includes: Step S1: Collect real workflow data of employees in a certain position, including: complete session records, continuous task steps, tool call chains, parameter configurations, and operating system interaction behaviors generated during actual office work, operating system, and tool calls; Step S2: Preprocess the collected real workflow data of employees in a certain position to obtain preprocessed real workflow data; Step S3: Perform process decomposition, skill extraction, and executable rule extraction on the preprocessed real workflow data to generate structured job skills and process rules; Step S4: Encapsulate the generated structured job skills and process rules into standardized Skill units; the standardized Skill units can be directly loaded into the agent framework; Step S5: When an employee in a certain position performs a task, the intelligent agent automatically identifies the current task, calls the corresponding standardized Skill unit, and realizes process guidance, template filling and tool calling.

[0008] Preferably, step S2 includes: performing data classification and collection, dirty data filtering, time-series unified sorting, business tag labeling, and fragmented data regularization and compression processing on the collected real workflow data of employees in a certain position to obtain a standardized workflow; The data classification and collection involves dividing the collected data into two categories: structured data from operating system logs and natural language process text from employees. The dirty data filtering process removes data with no business value, including debugging and testing operations, abandoned and interrupted operations, and garbled blank text. The unified temporal sorting is based on timestamps and semantic temporal terms to restore the actual execution order of tasks and isolate different business sessions. The business tag annotation is a metadata tag that includes business domain, operation type, and business format constraints; The fragmented data regularization and compression involves merging scattered and meaningless minor operations that meet preset requirements and removing isolated and invalid nodes.

[0009] Preferably, step S3 includes: Step S3.1: Based on the operation sequence, branch conditions, and independent operation actions, the standardized workflow is broken down into the smallest indivisible atomic task nodes, and the upstream and downstream dependencies and branch topology of the nodes are obtained. Step S3.2: Divide each atomic task node into high / medium / low automation nodes according to preset requirements; Step S3.3: For highly automated nodes, extract atomic skills with single responsibilities based on operational behavior and business system call characteristics; Step S3.4: Based on the original parameters, interaction logs, and branching logic, automatically extract data / material admission rules, execution timing and dependency rules, format conversion and output specification rules, interaction and API call rules, and exception handling and branching rules for each atomic skill; Step S3.5: Summarize the process topology, atomic skills, and supporting rules to generate structured job skills and process rules; The data / material access rules include rules that constrain task input content, file format, data source, and access permissions. The execution timing and dependency rules include: rules that constrain the order of task execution and the dependency relationships between upstream and downstream nodes; The format conversion and output specification rules include: rules for constraining the output content format, field standards, naming rules, and validation and conversion rules for layout requirements; The interaction and API call rules include: interface calls, parameter passing, and connection authentication rules for connecting to third-party business systems, databases, and tool platforms; The exception handling rules include: handling rules for exception scenarios such as task errors, missing data, timeouts, and verification failures.

[0010] Preferably, step S5 includes: Step S5.1: Collect real-time input commands from employees and operating system context, and combine them with historical task samples from the job position to identify the target task to be processed; Step S5.2: Divide the current target task to be processed into multiple atomic subtasks, and label each atomic subtask with multi-dimensional feature tags including business domain, operation action, and interface system; Step S5.3: Search the built-in standardized Skill unit resource library, compare the atomic subtask feature tags with the pre-stored adaptation tags of each standardized Skill unit, and realize one-to-one matching between atomic subtasks and standardized Skill units. Step S5.4: Generate the Skil execution chain based on the upstream and downstream dependencies and conditional branching relationships of atomic subtasks using standardized Skill units that are matched; Step S5.5: Execute the target task by calling the standardized Skill unit based on the generated Skil execution chain.

[0011] Preferably, the method further includes: establishing a multi-dimensional effect evaluation system, including: efficiency indicators, quality indicators, and experience indicators; evaluating the execution results of the current target task based on the established multi-dimensional effect evaluation system to obtain evaluation results; generating optimization suggestions based on the evaluation results; and completing process optimization, skill supplementation, and rule iteration based on the optimization suggestions.

[0012] A job-specific intelligent agent construction system based on employee workflow extraction, provided by the present invention, includes: Module M1: Collects real workflow data of employees in a certain position, including: complete session records, continuous task steps, tool call chains, parameter configurations, and operating system interaction behaviors generated during actual office work, operating system, and tool calls; Module M2: Preprocesses the collected real workflow data of employees in a certain position to obtain preprocessed real workflow data; Module M3: Performs process decomposition, skill extraction, and executable rule extraction on preprocessed real workflow data to generate structured job skills and process rules; Module M4: Encapsulates the generated structured job skills and process rules into standardized Skill units; these standardized Skill units can be directly loaded into the agent framework. Module M5: When an employee in a certain position performs a task, the intelligent agent automatically identifies the current task, calls the corresponding standardized Skill unit, and realizes process guidance, template filling and tool calling.

[0013] Preferably, module M2 includes: performing data classification and collection, dirty data filtering, time-series unified sorting, business tagging, and fragmented data regularization and compression processing on the collected real workflow data of employees in a certain position to obtain a standardized workflow; The data classification and collection involves dividing the collected data into two categories: structured data from operating system logs and natural language process text from employees. The dirty data filtering process removes data with no business value, including debugging and testing operations, abandoned and interrupted operations, and garbled blank text. The unified temporal sorting is based on timestamps and semantic temporal terms to restore the actual execution order of tasks and isolate different business sessions. The business tag annotation is a metadata tag that includes business domain, operation type, and business format constraints; The fragmented data regularization and compression involves merging scattered and meaningless minor operations that meet preset requirements and removing isolated and invalid nodes.

[0014] Preferably, the module M3 includes: Module M3.1: Based on the operation sequence, branch conditions, and independent operation actions, the standardized workflow is broken down into the smallest indivisible atomic task nodes, and the upstream and downstream dependencies and branch topology of the nodes are obtained. Module M3.2: Each atomic task node is divided into high / medium / low automation nodes according to preset requirements; Module M3.3: For highly automated nodes, it extracts atomic skills with single responsibilities based on operational behavior and business system call characteristics; Module M3.4: Based on raw parameters, interaction logs, and branching logic, it automatically extracts data / material admission rules, execution timing and dependency rules, format conversion and output specification rules, interaction and API call rules, and exception handling and branching rules for each atomic skill; Module M3.5: Summarizes process topology, atomic skills, and supporting rules to generate structured job skills and process rules; The data / material access rules include rules that constrain task input content, file format, data source, and access permissions. The execution timing and dependency rules include: rules that constrain the order of task execution and the dependency relationships between upstream and downstream nodes; The format conversion and output specification rules include: rules for constraining the output content format, field standards, naming rules, and validation and conversion rules for layout requirements; The interaction and API call rules include: interface calls, parameter passing, and connection authentication rules for connecting to third-party business systems, databases, and tool platforms; The exception handling rules include: handling rules for exception scenarios such as task errors, missing data, timeouts, and verification failures.

[0015] Preferably, the module M5 includes: Module M5.1: Collects real-time input commands from employees and operating system context, and combines this with historical task samples from the job position to identify the target task to be processed. Module M5.2: Breaks down the current target task into multiple atomic subtasks, and labels each atomic subtask with multi-dimensional feature tags including business domain, operation action, and interface system; Module M5.3: Retrieves the built-in standardized Skill unit resource library, compares the feature tags of atomic subtasks with the pre-stored adaptation tags of each standardized Skill unit, and realizes one-to-one matching between atomic subtasks and standardized Skill units; Module M5.4: Generates Skil execution chains based on standardized Skill units matched according to upstream and downstream dependencies and conditional branching relationships of atomic subtasks; Module M5.5: Executes target tasks by calling standardized Skill units based on the generated Skil execution chain.

[0016] Preferably, the system further includes: establishing a multi-dimensional effect evaluation system, including: efficiency indicators, quality indicators, and experience indicators; evaluating the execution results of the current target task based on the established multi-dimensional effect evaluation system to obtain evaluation results; generating optimization suggestions based on the evaluation results; and completing processes optimization, skills supplementation, and rule iteration based on the optimization suggestions.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention automatically extracts the core skills and processes in employee workflows, avoiding the inefficiency and omissions of manual sorting, and realizing the rapid and standardized accumulation of job knowledge in software development and data governance; 2. The standardized Skill unit of this invention can significantly reduce repetitive work by employees, reduce the learning time for new employees, and improve the execution efficiency and output quality consistency of tasks such as software development and data governance. 3. This invention constructs a closed-loop mechanism of "practice-evaluation-feedback-iteration", enabling the job-specific intelligent agent to continuously evolve with business changes and employee feedback, and adapt to the long-term development needs of the enterprise. 4. This invention upgrades individual job skills to an organizational-level digital assistant, enabling cross-job knowledge sharing and process collaboration, and providing core intelligent support for enterprise digital transformation; 5. This invention enables the automated accumulation, reuse, and continuous iteration of knowledge in core enterprise positions, significantly improving the work efficiency and service quality of positions such as software development and data governance, and providing a feasible intelligent auxiliary solution for enterprise digital transformation. Attached Figure Description

[0018] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating the method for constructing job-specific intelligent agents based on employee workflows. Detailed Implementation

[0019] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0020] Example 1 This invention discloses a method and system for constructing job-specific intelligent agents based on employee workflow extraction. By collecting employees' daily work input information and combining it with job workflow, dependent materials, information sources, and output standards, the system automatically extracts job skills and processes to form standardized skill modules. After autonomously assisting employees in practical verification, effect evaluation, and feedback iteration, the system aggregates these modules to form exclusive job-specific intelligent agents for different positions, ultimately building an organization-level digital assistant.

[0021] The method for constructing job-specific intelligent agents based on employee workflow extraction, such as Figure 1 As shown, it includes: Step S1: Collect real workflow data of employees in a certain position, including: complete session records, continuous task steps, tool call chains, parameter configurations, and operating system interaction behaviors generated during actual office work, operating system, and tool calls; Step S2: Preprocess the collected real workflow data of employees in a certain position to obtain preprocessed real workflow data; Specifically, step S2 includes: performing data classification and collection, dirty data filtering, time-series unified sorting, business tag labeling, and fragmented data regularization and compression processing on the collected real workflow data of employees in a certain position to obtain a standardized workflow; The data classification and collection involves dividing the collected data into two categories: structured data from operating system logs and natural language process text from employees. The dirty data filtering process removes data with no business value, including debugging and testing operations, abandoned and interrupted operations, and garbled blank text. The unified temporal sorting is based on timestamps and semantic temporal terms to restore the actual execution order of tasks and isolate different business sessions. The business tag annotation is a metadata tag that includes business domain, operation type, and business format constraints; The fragmented data regularization and compression involves merging scattered and meaningless minor operations that meet preset requirements and removing isolated and invalid nodes.

[0022] Step S3: Perform process decomposition, skill extraction, and executable rule extraction on the preprocessed real workflow data to generate structured job skills and process rules; Specifically, step S3 includes: Step S3.1: Based on the operation sequence, branch conditions, and independent operation actions, the standardized workflow is broken down into the smallest indivisible atomic task nodes, and the upstream and downstream dependencies and branch topology of the nodes are obtained. Step S3.2: Divide each atomic task node into high / medium / low automation nodes according to preset requirements; Step S3.3: For highly automated nodes, extract atomic skills with single responsibilities based on operational behavior and business system call characteristics; Step S3.4: Based on the original parameters, interaction logs, and branching logic, automatically extract data / material admission rules, execution timing and dependency rules, format conversion and output specification rules, interaction and API call rules, and exception handling and branching rules for each atomic skill; Step S3.5: Summarize the process topology, atomic skills, and supporting rules to generate structured job skills and process rules; The data / material access rules include rules that constrain task input content, file format, data source, and access permissions; for example, in the file-merger skill, only .docx / ..xlsx format files within the document library are allowed to be read, and illegal format files are denied access.

[0023] The execution timing and dependency rules include: rules that constrain the order of task execution and the dependency relationship between upstream and downstream nodes; for example, in a data governance scenario, "data extraction" must be executed first, followed by "data cleaning", and if the preceding task fails, the current Skill will be terminated.

[0024] The format conversion and output specifications include: rules for constraining the output content format, field standards, naming rules, and formatting requirements; for example, the report-export skill automatically names the merged document as "XX Department-Weekly Report-YYYYMMDD.pdf" and verifies the compliance of the page layout.

[0025] The interaction and API call rules include: interface calls, parameter passing, and connection authentication rules for connecting to third-party business systems, databases, and tool platforms; for example, the jira-processor skill, which pulls Jira task data according to a specified account and interface address, and limits the request frequency and parameter format.

[0026] The exception handling rules include rules for handling exception scenarios such as task errors, missing data, timeouts, and verification failures. For example, the notion-updater skill will retry twice if the database connection times out; if the retry fails, it will trigger log recording and push a manual reminder.

[0027] Step S4: Encapsulate the generated structured job skills and process rules into standardized Skill units; the standardized Skill units can be directly loaded into the agent framework; The standardized Skill unit is the smallest functional unit for implementing the capabilities of this invention. It possesses the characteristics of independent operation, reusability, and composability. When executing tasks, the standardized Skill unit provides standardized process guidance, template filling, operation prompts, and format validation, reducing the threshold for manual operation and minimizing human error. For example, the email-processor skill automatically parses email content and prompts for steps to extract table data. For highly repetitive subtasks with clearly defined rules, the standardized Skill unit can directly call built-in rules and interfaces to complete the operation fully automatically without manual intervention. Examples include periodically fetching Jira tasks, automatically merging weekly reports, and synchronizing data to Notion.

[0028] Step S5: When an employee in a certain position performs a task, the intelligent agent automatically identifies the current task, calls the corresponding standardized Skill unit, and realizes process guidance, template filling and tool calling.

[0029] Specifically, step S5 includes: Step S5.1: Collect real-time input commands from employees and operating system context, and combine them with historical task samples of the job to identify the target task to be processed; among which, historical task samples of the job include: historical workflow samples, job profiles, and a library of commonly used task tags for employees.

[0030] Step S5.2: Divide the current target task to be processed into multiple atomic subtasks, and label each atomic subtask with multi-dimensional feature tags including business domain, operation action, and interface system; Step S5.3: Retrieve the built-in standardized Skill unit resource library, and compare the atomic subtask feature tags with the pre-stored adaptation tags of each standardized Skill unit through semantic parsing, keyword matching, and scene feature comparison to achieve one-to-one matching between atomic subtasks and standardized Skill units. Step S5.4: Generate the Skil execution chain based on the upstream and downstream dependencies and conditional branching relationships of atomic subtasks using standardized Skill units that are matched; Step S5.5: Based on the generated Skil execution chain, call the standardized Skill unit to execute the target task; In this embodiment, based on the generated Skil execution chain, according to the automation level, select one of two modes to call the standardized Skill unit: manual assistance prompt or fully automatic execution; Among them, the manual assistance mode is suitable for complex, low standardization tasks: (1) Display the complete process steps, input and output requirements, and format specifications corresponding to the current task; (2) Perform format verification, content correction, and operation reminders in real time; (3) Record each step of the employee's operation behavior and intervention nodes, and generate workflow logs. The fully automatic execution mode is suitable for highly repetitive, rule-clear tasks: (1) Automatically connect and run all Skill units according to the time sequence rules, interface rules, and branch rules encapsulated inside the Skill; (2) Automatically complete data reading, processing, system interaction, and result output; (3) Trigger the built-in exception handling rules to deal with timeout, error and other issues; (4) After the task is completed, output the execution results and running logs.

[0031] For example, the process of placing an order in a shopping cart: When a user initiates an order from the shopping cart, the system first checks the inventory of the items in the cart and the discounts offered. If the inventory is insufficient, the order is terminated and the user is notified. If the inventory is sufficient, a master order is generated, the corresponding inventory is deducted, the coupon is locked, and then payment verification is initiated. If the payment is successful, a logistics message is pushed, and finally, an order record ledger is generated.

[0032] Post-processing: Clean the text / operation logs such as "shopping cart inventory verification, order generation, inventory deduction, payment and logistics push", and extract actions, branches, and timing. Identify segmentation breakpoints: Segment tasks based on chronological order, insufficient / sufficient inventory branches, and independent operation verbs; Atomic decomposition: Divide into single-responsibility subtasks such as checking inventory, generating orders, and deducting inventory, and mark serial / branch dependencies; Subtask tagging: Tag the operation type, business domain, and automation level; The output is a structured subtask chain. After decomposition, each independent subtask is matched one-to-one with the corresponding atomic Skills of shopping cart, order, and inventory, forming a complete automated execution chain. The core constraint of decomposition is that each subtask contains only one type of operation, without mixing multiple types of business logic, ensuring accurate matching of independent Skill units.

[0033] Specifically, the method further includes: establishing a multi-dimensional effect evaluation system, including efficiency indicators, quality indicators, and experience indicators; evaluating the execution results of the current target task based on the established multi-dimensional effect evaluation system to obtain evaluation results; generating optimization suggestions based on the evaluation results; and completing process optimization, skill supplementation, and rule iteration based on the optimization suggestions.

[0034] In this embodiment, the effect evaluation and feedback iteration is a closed-loop process of quantitative evaluation → problem identification → tiered optimization → version update, which is divided into 6 specific execution stages, illustrated with examples: Step 1: Comprehensive Data Collection; During the operation of the standard Skill Unit and the Role Agent, three types of data are automatically collected: Efficiency data: task duration, tool call duration, number of manual interventions; Quality data: Output error rate, number of non-compliant formats, number of validation failures; Behavioral data: new employee operation steps, temporary processing logic, and abnormal handling methods.

[0035] Step 2: Multi-dimensional indicator evaluation; According to the preset evaluation system, the collected data is quantitatively scored to determine operational defects: Low efficiency: determined to be due to redundant process nodes and unreasonable execution logic; Low quality: determined to be due to loopholes or incomplete constraints in the existing rules; Frequent manual intervention: This indicates that existing skills do not cover the new operational scenarios.

[0036] For example, the weekly report upload task frequently has format errors, and the assessment determines that "the output format validation rules are incomplete".

[0037] Step 3: Problem Classification and Localization; Based on the evaluation results, three optimization directions were identified: process optimization, skills enhancement, and rule iteration. Process optimization: The problem lies in the task execution order, node redundancy, and unreasonable dependencies; Skill Supplement: Existing atomic skills cannot cover the newly added operation scenarios and subdivided actions; Rule iteration: Existing access, format, interface, and exception rules have vulnerabilities and lag behind business changes.

[0038] Step 4: Feedback to the FDE Workflow Analyzer for secondary analysis; All assessment reports, new workflow logs, and temporary employee operation logic will be fed back into the core analytics module: For process optimization: break down the workflow, remove redundant nodes, and adjust the execution sequence and dependencies; Regarding skill enhancement: Extract new atomic operations from the newly added actions to generate new Skill modules; For rule iteration: abstract new validation rules, exception handling logic, and interface parameter rules from new business constraints and abnormal scenarios.

[0039] Step 5: Skill / Process Version Update; Regenerate and update the corresponding Skill units using skill-generator.py: Process adjustment: Update the scheduling sequence and dependencies between Skills; Skills Supplement: Added independent atomic skills and incorporated them into the job agent combination system; Rule iteration: Adding, modifying, and replacing old executable rules in the existing Skill.

[0040] Step 6: Gray-scale rollout and a new cycle; The optimized Skill / process will be rolled out in a phased, gradual manner to test its effectiveness on a small scale. Full release after verification; New operational data is collected again, and the process continues in a continuous cycle.

[0041] The method also includes: combining and encapsulating multiple Skill units for the same position to form a dedicated job-specific intelligent agent (such as a data governance agent or a data analysis agent) with complete job task processing capabilities; Organizational-level digital assistant upgrade: Connect intelligent agents from various positions to a unified organizational-level platform to achieve cross-position knowledge sharing, process collaboration and task scheduling, ultimately forming a digital assistant for the entire organization to support the overall digital operation of the enterprise.

[0042] In this embodiment, the atomic Skill units are independently decoupled and can be reused in different positions and processes. Multiple Skills for the same position can be freely combined and quickly encapsulated into a position-specific Agent, significantly reducing the cost and time required to build the job-specific intelligent agent. For example, the file-uploader skill can be used simultaneously for multiple scenarios such as report uploading and document archiving.

[0043] The present invention also provides a job-specific intelligent agent construction system based on employee workflow extraction. The job-specific intelligent agent construction system based on employee workflow extraction can be implemented by executing the process steps of the job-specific intelligent agent construction method based on employee workflow extraction. That is, those skilled in the art can understand the job-specific intelligent agent construction method based on employee workflow extraction as a preferred embodiment of the job-specific intelligent agent construction system based on employee workflow extraction.

[0044] Example 2 Example 2 is a preferred example of Example 1. A method for constructing a job-specific intelligent agent based on employee workflow extraction, provided by the present invention, includes: A backend Java developer who mainly develops the order module.

[0045] 1. Automatic data collection of the entire order workflow (Java data collection service); Automatically capture real operational data for two types of orders: Development and Operation Log: Backend Development Daily Operations - Order Table SQL Query / Add, Order Status Interface Integration, Order Timeout Scheduled Task Writing, Order Refund Branch Logic Development, Order Message Queue Sending and Receiving, Order Excel Export Operation; Business operation session logs: Test / operations personnel handle orders: record the entire process of order creation, payment verification, inventory deduction, after-sales refund, logistics synchronization, and manual review of abnormal orders; Java scheduled tasks clean logs, filter debugging test data, and output standardized order sequence workflows with timestamps, call order, and input parameters; Enterprise static specifications (order field standards, refund approval process, export file format) are only used as data filtering tags and do not participate in the generation of execution logic.

[0046] 2. WorkflowAnalyzer automatically breaks down order processes (Java parsing engine); The system automatically decomposes the order sequence operations into the smallest atomic operations: the original complete order process is: create an order → verify payment → deduct inventory → push logistics messages → enter the order ledger into the warehouse → automatically cancel if payment is not made within the time limit → archive the refund order; the analyzer automatically splits it into independent atomic nodes: 1) order creation interface call; 2) payment status verification; 3) inventory deduction SQL execution; 4) MQ message push; 5) order data entry into the warehouse; 6) timeout scheduled task judgment; 7) refund order generation and archiving; it automatically identifies branch logic: payment success triggers the delivery process, payment failure triggers order cancellation, and after-sales application enters the refund branch; it automatically marks the upstream and downstream dependencies of each node and generates a structured order process JSON.

[0047] 3. Multi-dimensional quantitative evaluation, extracting order-specific skills + five categories of executable rules; Java's built-in scoring model (operation repeatability, rule fixity, system API availability) automatically determines automation capabilities, generating atomic skills and machine-executable rules for the order domain. Automatically generate atomic skills for order processing (single responsibility, reusable): order-creator: Creates the master order; payment-checker: Payment status verification; stock-deductor: Deduction of order inventory; order-mq-pusher: Order logistics message push; order-sql-saver: Order ledger entry; order-timeout-handler: Orders not paid within the timeout period will be automatically cancelled; refund-order-generator: Generates and archives after-sales refund orders; Automatically extract executable rules across the entire order chain (no manual coding required). Data access rules: Only allow the input of valid orders with a status of "pending payment / paid / after-sales", and filter out invalid empty orders and orders with abnormal amounts; Execution sequence dependency rule: Order creation must be completed before inventory deduction can be performed; if inventory deduction fails, the order process will be terminated and logistics will not be pushed. Output format specification rules: Order export files should be uniformly named date-order ledger.xlsx, and should always include the order number, amount, user ID, and logistics status fields; System API interaction rules: Order payment interface address, inventory database connection parameters, MQ queue name, and scheduled task execution cycle are automatically extracted and fixed; Abnormal branch flow rules: Orders are automatically rolled back when inventory is insufficient; payment interfaces are retried twice after a timeout, and consecutive failures generate abnormal orders that are pushed for manual review; refunds exceeding the threshold trigger approval blocking.

[0048] 4. Automatically generate standardized skill code skeletons for orders (Java calling Python scripts) The Java backend passes the structured data of order skills and rules to skill-generator.py; Automatically generate the SKILL.md configuration file for order-related skills; The output is a Java code template that can be directly integrated into the intelligent agent framework: it includes a complete code structure for order input parameter validation, MyBatis database operations, MQ sending, scheduled tasks, and exception handling; Individual order skills can be freely combined to create a complete order processing workflow.

[0049] 5. Order business data feedback and automatic iterative updates (closed loop); When employees use the Order Skill to handle business / development iterations, the backend tracking points collect new scenarios such as new installment orders, coupon-deducted orders, and batch after-sales refunds, and automatically feed the logs back to WorkflowAnalyzer. The system re-decomposes the process, automatically adds new skills such as coupon-order-processor and batch-refund-handler, and synchronously updates order verification and branch judgment rules. It also performs canary release updates to adapt to order business iterations, without requiring manual rule maintenance throughout the entire process.

[0050] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0051] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for constructing a job-specific intelligent agent based on employee workflow extraction, characterized in that, include: Step S1: Collect real workflow data of employees in designated positions, including: complete session records, continuous task steps, tool call chains, parameter configurations, and operating system interaction behaviors generated during actual office work, operating system, and tool calls; Step S2: Preprocess the collected real workflow data of employees in a certain position to obtain preprocessed real workflow data; Step S3: Perform process decomposition, skill extraction, and executable rule extraction on the preprocessed real workflow data to generate structured job skills and process rules; Step S4: Encapsulate the generated structured job skills and process rules into standardized Skill units; the standardized Skill units can be directly loaded into the agent framework; Step S5: When an employee in a certain position performs a task, the intelligent agent automatically identifies the current task, calls the corresponding standardized Skill unit, and realizes process guidance, template filling and tool calling.

2. The method for constructing a job-specific intelligent agent based on employee workflow extraction according to claim 1, characterized in that, Step S2 includes: performing data classification and collection, dirty data filtering, time-series unified sorting, business tag labeling, and fragmented data regularization and compression processing on the collected real workflow data of employees in a certain position to obtain a standardized workflow; The data classification and collection involves dividing the collected data into two categories: structured data from operating system logs and natural language process text from employees. The dirty data filtering process removes data with no business value, including debugging and testing operations, abandoned and interrupted operations, and garbled blank text. The unified temporal sorting is based on timestamps and semantic temporal terms to restore the actual execution order of tasks and isolate different business sessions. The business tag annotation is a metadata tag that includes business domain, operation type, and business format constraints; The fragmented data regularization and compression involves merging scattered and meaningless minor operations that meet preset requirements and removing isolated and invalid nodes.

3. The method for constructing a job-specific intelligent agent based on employee workflow extraction according to claim 1, characterized in that, Step S3 includes: Step S3.1: Based on the operation sequence, branch conditions, and independent operation actions, the standardized workflow is broken down into the smallest indivisible atomic task nodes, and the upstream and downstream dependencies and branch topology of the nodes are obtained. Step S3.2: Divide each atomic task node into high / medium / low automation nodes according to preset requirements; Step S3.3: For highly automated nodes, extract atomic skills with single responsibilities based on operational behavior and business system call characteristics; Step S3.4: Based on the original parameters, interaction logs, and branching logic, automatically extract data / material admission rules, execution timing and dependency rules, format conversion and output specification rules, interaction and API call rules, and exception handling and branching rules for each atomic skill; Step S3.5: Summarize the process topology, atomic skills, and supporting rules to generate structured job skills and process rules; The data / material access rules include rules that constrain task input content, file format, data source, and access permissions. The execution timing and dependency rules include: rules that constrain the order of task execution and the dependency relationships between upstream and downstream nodes; The format conversion and output specification rules include: rules for constraining the output content format, field standards, naming rules, and validation and conversion rules for layout requirements; The interaction and API call rules include: interface calls, parameter passing, and connection authentication rules for connecting to third-party business systems, databases, and tool platforms; The exception handling rules include: handling rules for exception scenarios such as task errors, missing data, timeouts, and verification failures.

4. The method for constructing a job-specific intelligent agent based on employee workflow extraction according to claim 1, characterized in that, Step S5 includes: Step S5.1: Collect real-time input commands from employees and operating system context, and combine them with historical task samples from the job position to identify the target task to be processed; Step S5.2: Divide the current target task to be processed into multiple atomic subtasks, and label each atomic subtask with multi-dimensional feature tags including business domain, operation action, and interface system; Step S5.3: Search the built-in standardized Skill unit resource library, compare the atomic subtask feature tags with the pre-stored adaptation tags of each standardized Skill unit, and realize one-to-one matching between atomic subtasks and standardized Skill units. Step S5.4: Generate the Skil execution chain based on the upstream and downstream dependencies and conditional branching relationships of atomic subtasks using standardized Skill units that are matched; Step S5.5: Execute the target task by calling the standardized Skill unit based on the generated Skil execution chain.

5. The method for constructing a job-specific intelligent agent based on employee workflow extraction according to claim 4, characterized in that, The method further includes: establishing a multi-dimensional effect evaluation system, including efficiency indicators, quality indicators, and experience indicators; evaluating the execution results of the current target task based on the established multi-dimensional effect evaluation system to obtain evaluation results; generating optimization suggestions based on the evaluation results; and completing process optimization, skill supplementation, and rule iteration based on the optimization suggestions.

6. A job-specific intelligent agent construction system based on employee workflow extraction, characterized in that, include: Module M1: Collects real workflow data of employees in designated positions, including: complete session records, continuous task steps, tool call chains, parameter configurations, and operating system interaction behaviors generated during actual office work, operating system, and tool calls; Module M2: Preprocesses the collected real workflow data of employees in a certain position to obtain preprocessed real workflow data; Module M3: Performs process decomposition, skill extraction, and executable rule extraction on preprocessed real workflow data to generate structured job skills and process rules; Module M4: Encapsulates the generated structured job skills and process rules into standardized Skill units; these standardized Skill units can be directly loaded into the agent framework. Module M5: When an employee in a certain position performs a task, the intelligent agent automatically identifies the current task, calls the corresponding standardized Skill unit, and realizes process guidance, template filling and tool calling.

7. The job-specific intelligent agent construction system based on employee workflow extraction according to claim 6, characterized in that, The module M2 includes: performing data classification and collection, dirty data filtering, time-series unified sorting, business tag labeling, and fragmented data regularization and compression processing on the collected real workflow data of employees in a certain position to obtain a standardized workflow; The data classification and collection involves dividing the collected data into two categories: structured data from operating system logs and natural language process text from employees. The dirty data filtering process removes data with no business value, including debugging and testing operations, abandoned and interrupted operations, and garbled blank text. The unified temporal sorting is based on timestamps and semantic temporal terms to restore the actual execution order of tasks and isolate different business sessions. The business tag annotation is a metadata tag that includes business domain, operation type, and business format constraints; The fragmented data regularization and compression involves merging scattered and meaningless minor operations that meet preset requirements and removing isolated and invalid nodes.

8. The job-specific intelligent agent construction system based on employee workflow extraction according to claim 6, characterized in that, The module M3 includes: Module M3.1: Based on the operation sequence, branch conditions, and independent operation actions, the standardized workflow is broken down into the smallest indivisible atomic task nodes, and the upstream and downstream dependencies and branch topology of the nodes are obtained. Module M3.2: Each atomic task node is divided into high / medium / low automation nodes according to preset requirements; Module M3.3: For highly automated nodes, it extracts atomic skills with single responsibilities based on operational behavior and business system call characteristics; Module M3.4: Based on raw parameters, interaction logs, and branching logic, it automatically extracts data / material admission rules, execution timing and dependency rules, format conversion and output specification rules, interaction and API call rules, and exception handling and branching rules for each atomic skill; Module M3.5: Summarizes process topology, atomic skills, and supporting rules to generate structured job skills and process rules; The data / material access rules include rules that constrain task input content, file format, data source, and access permissions. The execution timing and dependency rules include: rules that constrain the order of task execution and the dependency relationships between upstream and downstream nodes; The format conversion and output specification rules include: rules for constraining the output content format, field standards, naming rules, and validation and conversion rules for layout requirements; The interaction and API call rules include: interface calls, parameter passing, and connection authentication rules for connecting to third-party business systems, databases, and tool platforms; The exception handling rules include: handling rules for exception scenarios such as task errors, missing data, timeouts, and verification failures.

9. The job-specific intelligent agent construction system based on employee workflow extraction according to claim 6, characterized in that, The module M5 includes: Module M5.1: Collects real-time input commands from employees and operating system context, and combines this with historical task samples from the job position to identify the target task to be processed. Module M5.2: Breaks down the current target task into multiple atomic subtasks, and labels each atomic subtask with multi-dimensional feature tags including business domain, operation action, and interface system; Module M5.3: Retrieves the built-in standardized Skill unit resource library, compares the feature tags of atomic subtasks with the pre-stored adaptation tags of each standardized Skill unit, and realizes one-to-one matching between atomic subtasks and standardized Skill units; Module M5.4: Generates Skil execution chains based on standardized Skill units matched according to upstream and downstream dependencies and conditional branching relationships of atomic subtasks; Module M5.5: Executes target tasks by calling standardized Skill units based on the generated Skil execution chain.

10. The job-specific intelligent agent construction system based on employee workflow extraction according to claim 9, characterized in that, The system also includes: establishing a multi-dimensional effect evaluation system, including efficiency indicators, quality indicators, and experience indicators; evaluating the execution results of the current target task based on the established multi-dimensional effect evaluation system to obtain evaluation results; generating optimization suggestions based on the evaluation results; and completing optimization suggestions including process optimization, skill supplementation, and rule iteration.

Images