Digital employee automatic construction method based on business system self-learning

Abstract

The invention discloses a digital employee automatic construction method based on business system self-learning, and the method comprises the steps: receiving a natural language demand description of a business system through a self-learning integrated agent, calling a large language model to carry out intention understanding, and generating a structured tool development plan; the self-learning integrated agent adopts a three-stage self-learning strategy of exploration-learning-optimization to generate a parameterized script; the self-learning integrated agent encapsulates the optimized parameterized script into a standardized tool conforming to the MCP protocol specification, and encapsulates the structured business data model into a data structure definition document of the tool; the self-learning integrated agent registers the standardized tool to an MCP tool registration center; a digital employee agent loads a tool and business data structure definition from an MCP tool registration center, intelligent tool calling is carried out by adopting an'intention-planning-execution 'three-layer decision-making architecture, business task execution is completed, and the construction threshold and cost of digital employees are greatly reduced.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and enterprise digital transformation technology, specifically to a method for automating the construction of digital employees based on self-learning of business systems. Background Technology

[0002] Digital employees are intelligent agent applications that can understand business needs, perform business tasks, and interact naturally with users, much like human employees. Traditional methods of building digital employees primarily rely on manual development. Developers need to manually write business operation scripts, encapsulate API interfaces, develop tool plugins, and then integrate these tools into the intelligent agent framework. This approach not only has a long development cycle but also a high technical threshold, requiring developers to master multiple aspects such as business processes, automation technologies, and intelligent agent development. When the business system interface or processes change, manual adaptation and code modification are required, leading to continuously accumulating maintenance costs.

[0003] In recent years, breakthroughs in Large Language Model (LLM) technology have brought new possibilities to digital employees. LLM-based agents possess powerful natural language understanding and reasoning capabilities, enabling them to understand user intent, plan execution steps, and invoke external tools. However, tools in existing agent development frameworks still require manual development and registration, and cannot be automatically generated from business systems. This makes the expansion of digital employee capabilities heavily reliant on development resources, hindering rapid responses to changing business needs. Existing technologies suffer from three main core problems: First, digital employee development is inefficient, typically taking 1-2 months from requirement submission to deployment, failing to meet enterprises' rapid deployment requirements. Second, tool development lacks standardization; tools created by different developers vary in format, lacking a unified tool management and invocation mechanism, making tools difficult to reuse and share. Third, agent technology is disconnected from business operational capabilities; while agents possess powerful dialogue and reasoning abilities, they are insufficient in autonomously acquiring operational capabilities from business systems, limiting the practical value of digital employees. Summary of the Invention

[0004] The purpose of this invention is to provide an automated method for building digital employees based on self-learning of business systems. This method utilizes a self-learning integrated intelligent agent to automatically learn the operation methods of the business system, autonomously generate business operation tools, and standardize and encapsulate them, enabling the digital employee intelligent agent to quickly acquire business capabilities. The self-learning integrated intelligent agent possesses autonomous learning and automatic debugging capabilities. It can automatically understand the operational logic by observing and analyzing the business system interface, and automatically identify and repair anomalies during execution. This allows for the integration of existing business systems with digital employees at extremely low cost, thus solving the problems mentioned in the background section.

[0005] This invention provides the following technical solution: an automated construction method for digital employees based on self-learning of business systems, comprising a self-learning integrated intelligent agent and a digital employee intelligent agent, the method comprising the following operational steps:

[0006] Step S1: The self-learning integrated intelligent agent receives the natural language requirement description of the business system, calls the large language model for intent understanding, extracts the target business system address, operation sequence, key data fields and expected output format, and generates a structured tool development plan.

[0007] Preferably, invoking the large language model for intent understanding includes: the large language model identifying key information in the business system requirement description: the name of the target business system, the type of operation to be performed, and the data objects involved; decomposing the business process into ordered atomic operation steps, and identifying the key data fields involved in the operation, their data types, and constraints.

[0008] Step S2: The self-learning integrated intelligent agent adopts a three-stage autonomous learning strategy of "exploration-learning-optimization". It autonomously explores the operation path of the business system, learns the semantics and constraint relationships of business data fields, builds a structured business data model, and automatically debugs and optimizes the generated scripts to generate parameterized scripts.

[0009] Preferably, the self-learning integrated intelligent agent adopts a three-stage autonomous learning strategy of "exploration-learning-optimization," including: in the exploration stage, starting the browser's automated control tool, autonomously exploring the operation path and interaction logic of the business system through interface interaction, and synchronously capturing operation sequences and page states; in the learning stage, using a large language model to conduct in-depth analysis of the exploration results, autonomously learning the semantics and constraint relationships of business data fields, and constructing a structured business data model; and in the optimization stage, automatically debugging and parameterizing the generated scripts, including automatically identifying and repairing location failures, automatically handling abnormal processes, automatically optimizing execution paths, and generating parameterized scripts with generalization and self-healing capabilities.

[0010] Step S3: The self-learning integrated agent encapsulates the optimized parameterized script into a standardized tool conforming to the MCP protocol specification, and at the same time encapsulates the structured business data model into the tool's data structure definition document.

[0011] Preferably, the encapsulated standardized tools include: generating tool description JSON, parameter schema definition, execution logic, and error handling mechanism; and encapsulated data structure definition document: used to describe the business data fields, field types, constraints, and relationships between fields involved in the tool.

[0012] Step S4: The self-learning integrated intelligent agent registers the standardized tools to the MCP tool registry center for unified management and reuse sharing.

[0013] Preferably, the self-learning integrated intelligent agent registers standardized tools to the MCP tool registry center by: uploading tool metadata and business data structure definition documents, configuring service access addresses and authentication methods, setting access permissions and visibility scope, establishing a version management mechanism, and configuring health checks and performance monitoring indicators.

[0014] Step S5: The digital employee intelligent agent loads the tool and business data structure definition from the MCP tool registry center, and uses a three-layer decision architecture of "intent-planning-execution" to intelligently call tools and complete the execution of business tasks.

[0015] Preferably, the three-tiered decision-making framework of "intent-planning-execution" includes:

[0016] Intent Layer: The digital employee intelligent agent receives the user's natural language instructions, calls the large language model for semantic understanding, extracts business entity information and operation intent, performs structured processing on the extracted business data according to the business data structure definition, and persists it to a document-based database.

[0017] Planning layer: Based on structured intent, the digital employee intelligent agent filters candidate tools from the list of loaded tools using semantic matching algorithms, analyzes the dependencies and data flow requirements between tools, and generates tool call sequences and parameter mapping schemes;

[0018] Execution Layer: The digital employee intelligent agent executes tasks through task queue management and scheduling tools. The task queue maintains a list of tasks to be executed, task execution status, and dependencies between tasks, supporting both serial and parallel execution of tasks. The execution layer queries and loads the structured business data required for the current task from a document-oriented database, passes the business data as parameters to the MCP tool for execution, updates the task status and business data in the database after obtaining the execution results, and retryes or rolls back abnormally according to a preset strategy when the tool fails to execute.

[0019] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0020] (1) In terms of the efficiency of digital employee development, this invention achieves full-process automation from capability requirements to tool generation through self-learning integrated intelligence, compressing the original 1-2 months of digital employee development work to 1-2 days, improving efficiency by more than 90%. Enterprises do not need to equip themselves with professional automation development teams. Business personnel can complete tool generation by describing the capabilities required for digital employees in natural language, which greatly reduces the technical threshold and significantly reduces development costs.

[0021] (2) In terms of tool standardization, this invention adopts the MCP protocol as a unified tool specification. All generated tools naturally conform to the standard and can be seamlessly called by any digital employee application that supports MCP. Through a unified tool registry center, tool reuse and sharing are realized, avoiding redundant development. One tool can be assigned to multiple digital employees, greatly improving tool utilization.

[0022] (3) In terms of intelligence, the self-learning integrated intelligence of this invention can not only automatically record operation scripts, but also intelligently optimize scripts through the reasoning ability of large language models, including removing redundant operations, selecting stable positioning strategies, adding exception handling, and achieving parameterization and generalization. The stability of the tool after intelligent optimization is improved by more than 40%, enabling digital employees to have more reliable business execution capabilities.

[0023] (4) In terms of ease of application, this invention makes expanding the capabilities of digital employees extremely simple; users only need to select the required tools from the tool registry and configure prompts. Business personnel do not need to write code and can build digital employees with complete business capabilities in just 1-2 days, which is dozens of times more efficient than the 1-2 month development cycle of traditional methods. Enterprises can quickly deploy multiple digital employees, each responsible for different business areas.

[0024] (5) In terms of scalability, this invention adopts a modular and loosely coupled architecture design, allowing for flexible replacement and expansion of each component. It supports various intelligent agent frameworks, browser tools, large language models, and tool registry solutions, enabling enterprises to customize deployments according to their own needs. Digital employees can continuously expand their capabilities by adding new tools without modifying the core code, adapting to the continuous evolution of business.

[0025] (6) In terms of operating costs, this invention significantly reduces the cost of using large models through the "learn once, reuse continuously" mechanism. The parameterized scripts generated after the self-learning integrated agent completes business learning can be repeatedly and directly executed by the digital employee agent without repeatedly calling the large model for inference in each business operation. Compared with the solution that requires the large model to participate in the entire process for each execution, this invention can reduce the token consumption of a single business execution by more than 80%. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0027] Figure 1 This is a schematic diagram of the overall process of the automated construction method for digital employees based on self-learning of business systems provided in this embodiment of the invention;

[0028] Figure 2This is a schematic diagram of the self-learning integrated intelligent workflow provided in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the digital employee intelligent agent architecture provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1

[0032] Combination Figures 1-3 As shown, this invention provides the following technical solution: an automated construction method for digital employees based on self-learning of business systems, comprising a self-learning integrated intelligent agent and a digital employee intelligent agent, wherein the self-learning integrated intelligent agent has autonomous learning capabilities, can automatically learn the operation methods of the business system, autonomously generate business operation tools and perform standardized encapsulation, and also has automatic debugging capabilities, automatically identifying problems and repairing and optimizing them during tool generation and execution; the digital employee intelligent agent executes business tasks by calling these tools. The specific operation steps of this invention include:

[0033] Step S1: The self-learning integrated agent receives the capability requirement description of the digital employee agent and calls the large language model for intent understanding.

[0034] In this embodiment, the specific method of calling the large language model for intent understanding includes: receiving a natural language description of the business system's requirements, calling the large language model API for deep semantic analysis; using the large language model to identify key information in the requirements, including the name of the target business system, the type of operation to be performed, and the data objects involved; extracting the operation sequence and decomposing the business process into ordered atomic operation steps; identifying the key data fields involved in the operation and their data types and constraints; clarifying the expected output format requirements; and generating a structured tool development plan based on the extracted information.

[0035] Step S2: The self-learning ensemble agent adopts a three-stage autonomous learning strategy of "exploration-learning-optimization" to generate parameterized scripts:

[0036] In this embodiment, the self-learning integrated intelligent agent adopts a three-stage autonomous learning strategy of "exploration-learning-optimization". In the exploration stage, the browser automation control tool is launched to autonomously explore the operation path and interaction logic of the business system through interface interaction, and simultaneously capture operation sequences and page states. In the learning stage, the large language model is used to conduct in-depth analysis of the exploration results, autonomously learn the semantics and constraint relationships of business data fields, and construct a structured business data model. In the optimization stage, the generated script is automatically debugged and parameterized reconstructed, including automatically identifying and repairing location failures, automatically handling abnormal processes, and automatically optimizing execution paths, generating parameterized scripts with good generalization and self-healing capabilities.

[0037] For example, the "exploration-learning-optimization" three-stage self-directed learning strategy specifically includes:

[0038] Exploration Phase: The self-learning integrated intelligent agent launches the browser automation control tool, configures browser parameters and navigates to the target system, handles the login authentication process, and explores the operation path of the business system autonomously through interface element recognition and interaction according to the generated tool development plan. It uses multiple positioning strategies to accurately locate page elements, executes corresponding operations while simultaneously capturing the operation sequence, records the target element, operation type, input data, page state changes and corresponding page DOM snapshots for each step of the operation, and generates the original operation script.

[0039] Learning Phase: The self-learning integrated intelligent agent inputs the original scripts, DOM snapshots, and page state change information obtained from exploration into the large language model for in-depth analysis. It autonomously learns the semantic meaning, data type, constraints, and relationships between business data fields, understands the preconditions and post-operations of business operations, constructs the structured business data model required by the current business needs, and identifies the core business operations and auxiliary operations in the script, distinguishing between necessary execution steps and optimizable steps.

[0040] Optimization Phase: Based on the business knowledge acquired during the learning phase, the self-learning integrated intelligent agent automatically debugs and refactors the original script, automatically identifies and fixes element location failures, automatically handles page loading delays and asynchronous operations, automatically optimizes execution paths to improve efficiency, replaces identified business data fields with parameterized variables, removes redundant auxiliary operations, establishes a multi-strategy locator alternative chain, adds intelligent exception handling and automatic retry mechanisms, converts extracted data into structured JSON format, and generates parameterized scripts with good generalization and self-healing capabilities.

[0041] In this embodiment, the self-learning integrated agent possesses automatic debugging capabilities: during script generation, the self-learning integrated agent automatically verifies and executes the generated script, monitoring for anomalies during execution; when element location failure is detected, it automatically analyzes changes in the page's DOM structure and attempts to use alternative location strategies or dynamically generate new location expressions; when an operation execution timeout is detected, it automatically adjusts the waiting strategy, optimizing it from fixed waiting to conditional waiting or intelligent waiting; when a business process anomaly is detected, it automatically analyzes the cause of the anomaly, attempts to perform repair operations, or rolls back to the previous stable state for re-execution; the automatic debugging process iterates until the script can stably complete the business process, and the anomaly handling experience accumulated during debugging is encoded into the final script, improving the tool's robustness and self-healing capabilities.

[0042] Step S3: The self-learning integrated agent encapsulates the optimized parameterized script into a standardized tool that conforms to the MCP protocol specification.

[0043] In this embodiment, the specific methods of standardizing tool encapsulation include: self-learning integrated intelligence generates tool description documents based on the MCP protocol specification, defining the tool name, function description, parameter schema, and execution interface; the structured business data model built in the analysis phase of the adopted "exploration-learning-optimization" three-stage self-learning strategy is encapsulated into the tool's data structure definition, describing the business data fields, field types, constraints, and relationships between fields involved in the tool; the parameterized script is encapsulated into a service that can be called through standard protocols, implementing the tool's execute method, adding error handling mechanisms and logging functions; and detailed tool usage documentation and calling examples are generated, enabling digital employees to understand and correctly use the tool.

[0044] The final output includes a tool description JSON, parameter schema definitions, execution logic, error handling, and usage documentation. It also encapsulates the structured business data model built during the learning phase into a data structure definition document for the tool, enabling it to be invoked by any digital employee that supports the MCP protocol.

[0045] Step S4: The self-learning integrated intelligent agent registers the standardized tools to the MCP tool registry center for unified management and reuse sharing.

[0046] In this embodiment, the specific methods for tool registration and management include: self-learning integrated smart connection to the MCP tool registration center service, uploading tool metadata and description information, as well as the business data structure definition model corresponding to the tool; configuring the tool's service access address and authentication method; setting the tool's access permissions and visibility scope; establishing a version management mechanism to support the tool's iterative updates; and configuring health checks and performance monitoring indicators to ensure the stable operation of the tool.

[0047] Step S5: The digital employee intelligent agent loads available tools and corresponding business data structure definitions from the MCP tool registry center, automatically injects the data structure definitions into the prompt words, and adopts a three-layer decision architecture of "intent-planning-execution" to realize intelligent tool invocation. Through the business data structure memory mechanism, the extracted business data is persistently stored in a document-type database. During execution, structured data is loaded from the database and passed to the tool. The execution of tasks is managed through a task queue, and finally the results are fed back to the user, possessing complete business execution capabilities.

[0048] In this embodiment, the specific methods for endowing digital employees with capabilities include: users defining the role and scope of responsibility of digital employees by writing prompts, and selecting and configuring the tools that the digital employee needs to use from the MCP tool registry center; the system automatically loads tool descriptions, call interfaces and corresponding business data structure definition models, and automatically injects the business data structure definition into the prompts of digital employees, enabling digital employees to have a structured understanding of data in the business domain; users interact with digital employees through natural language, and digital employees understand user intentions, plan execution paths, select appropriate tools, execute business processes and provide feedback on execution results based on a large language model, thereby completing business tasks like human employees.

[0049] For example, the digital employee intelligent agent adopts a three-layer decision-making architecture of "intent-planning-execution" to realize intelligent tool invocation, and realizes persistent data management through a business data structure memory mechanism, specifically including:

[0050] Intent Layer: The digital employee intelligent agent receives natural language instructions from users, calls a large language model for semantic understanding, extracts business entity information and operational intent from user statements, identifies the business goals that users expect to achieve, the data objects involved, the operation types and constraints, generates a structured intent representation, and, based on the business data structure definition injected in the prompt words, performs structured processing on the extracted business data and persists it to a document-oriented database (such as MongoDB), establishing an association index between business data and the current task.

[0051] Planning Layer: Based on the structured intent output by the intent layer, the digital employee intelligent agent filters candidate tools related to the current intent from the list of loaded tools using a semantic matching algorithm, analyzes the dependencies between tools and data flow requirements, and generates tool call sequences and parameter mapping schemes. When the business process involves multiple tools, the planning layer determines the order of tool calls and the correspondence between the output of the preceding tool and the input of the subsequent tool.

[0052] Execution Layer: The digital employee intelligent agent executes tasks based on the invocation scheme generated by the planning layer through task queue management and scheduling tools. The task queue maintains a list of tasks to be executed, task execution status, and dependencies between tasks, supporting both serial and parallel execution. The execution layer sequentially retrieves tasks from the task queue, queries and loads the structured business data required by the current task from the document-oriented database, passes the business data as parameters to the corresponding MCP tool for execution, updates the task status and business data in the document-oriented database after obtaining the execution result, and passes the result to subsequent dependent tasks. When a tool fails to execute, the execution layer retryes or rolls back according to a preset strategy to ensure the integrity and consistency of the business process. After all tools have completed execution, the execution layer summarizes the execution results and converts them into natural language to provide feedback to the user.

[0053] Example 2

[0054] In this embodiment, the official vehicle application function of a university information portal system is used as an example to give the following complete implementation process. The official vehicle application system contains 9 required form fields, covering a variety of common form controls such as text input, number input, date and time selection, radio buttons, and check boxes, which is highly representative.

[0055] Step 1: Understanding Business Intent and Operational Planning.

[0056] In this embodiment, the user submits a request to the self-learning integrated intelligent agent: I need to automatically fill out and submit an application form for official vehicles, including the number of passengers, the number of students, the vehicle type, the source of funding, the schedule, and the location information. The system address is the school's information portal website.

[0057] The self-learning ensemble agent invokes a large language model for intent understanding. The large language model, through deep semantic analysis, returns a structured operation plan, which mainly includes the following key information:

[0058] Target business system: University Information Portal Service Center;

[0059] Authentication method: CAS Single Sign-On system;

[0060] Operation sequence: CAS login → Navigation service center → Open application form → Fill in form → Submit form → Confirm result;

[0061] Input parameters: reason for vehicle use, number of passengers, number of students, source of funding, vehicle type selection, departure time, end time, pick-up location, and arrival location;

[0062] Expected output: A structured JSON object containing a success flag and a message.

[0063] Step 2: Perform automated recording and intelligent optimization.

[0064] In this embodiment, the self-learning integrated agent adopts a three-stage script generation strategy of "recording-analysis-optimization".

[0065] Recording Phase: The self-learning integrated agent invokes the browser automation control tool to launch a browser instance. The browser navigates to the target system, identifies the CAS single sign-on system, automatically completes the authentication process, and after successful login, the agent executes business operations step by step according to the tool development plan: navigating to the service center page, locating and clicking the "Official Vehicle Application" link, waiting for the new tab to open, and locating the form page. The self-learning integrated agent uses multiple positioning strategies to accurately locate nine form fields, including text input boxes, numeric input boxes, date and time pickers, radio buttons, and checkboxes, and fills in test data in sequence. After completion, the submit button is clicked, the confirmation dialog box is processed, and submission is completed. Throughout the entire operation, the system synchronously captures the operation sequence, records the target element, operation type, input data, and corresponding page DOM snapshot for each step, and generates the original operation script.

[0066] Analysis Phase: The self-learning integrated agent inputs the recorded raw script and DOM snapshot information into the large language model for deep analysis. The large language model identifies the business data fields involved in the script, extracts field names, data types, constraints, and relationships between fields, and constructs a structured business data model for official vehicle applications. This model contains 9 business data fields: reason for vehicle use (string type), number of passengers (integer type, range 1-50), number of students (integer type, must be less than or equal to the number of passengers), funding source (enumeration type, selectable values ​​are off-campus special funds / on-campus special funds / public funds), vehicle type selection (array type, supports multiple selections), departure time and arrival time (date and time type, arrival time must be later than departure time), pick-up location and arrival location (string type). Simultaneously, the model identifies core business operations and auxiliary operations in the script, distinguishing between necessary execution steps and optimizable steps.

[0067] Optimization Phase: Based on the business data model and operation classification results obtained in the analysis phase, the self-learning ensemble agent reconstructs and optimizes the original script. The model performs the following optimization tasks: identifies and merges repetitive waiting operations, changing fixed-time waiting to intelligent waiting; replaces identified business data fields with parameterized variables, such as replacing hard-coded values ​​like student ID "20231001" and location "Xiaobaihua Theater" with function parameters; optimizes the element locator, using multiple locator methods such as role, text, and placeholder, and establishes a locator alternative chain; adds a retry mechanism of up to 5 times and exception handling logic for critical operations; encapsulates the complex operations of the date and time selector into independent functions; and converts the extracted data into structured JSON format. The optimized script has better stability and generalization ability.

[0068] Step 3: Standardize and package the tools.

[0069] In this embodiment, the self-learning integrated intelligent agent generates a tool description based on the MCP protocol specification, defines the tool name as "submit-car-apply", and describes its function as "submitting an application for official vehicle use, automatically filling in the form and completing the submission process".

[0070] For example, the parameter schema definition generated by the official vehicle application system includes 9 required parameters and 2 optional parameters. The required parameters include reason (string, reason for vehicle use), people_count (integer, number of passengers), student_count (integer, number of students), funding_source (string enumeration, funding source, optional values ​​are "external special fund" / "internal special fund" / "public funds"), car_types (string array, vehicle type selection, optional values ​​are "car" / "commercial vehicle" / "minibus" / "bus" / "freight vehicle"), start_time and end_time (string, format YYYY-MM-DDHH:MM), origin and destination (string, location). Optional parameters include username and password for CAS authentication.

[0071] In this embodiment, the execution method of the official vehicle application system tool receives a parameter object, internally calls an optimized parameterized script, launches the browser, completes CAS login, navigates to the form page, fills in nine fields, submits the form, verifies the result, and finally returns a structured execution result. A multi-layered exception handling mechanism is added to capture various errors such as network errors, element location failures, login timeouts, and form verification failures, returning standardized error messages. Detailed tool usage documentation is generated, including parameter descriptions, calling examples, and return formats.

[0072] For example, the official vehicle application system encapsulates the structured business data model built in the analysis phase of step two into a data structure definition model for the tool, which describes the nine business data fields involved in the tool, field types, constraints, and relationships between fields, enabling digital employees to understand the data structure of this business domain.

[0073] Step 4: Manage the entire lifecycle of the tool.

[0074] In this embodiment, the self-learning integrated agent connects to the MCP tool registry center, uploads tool metadata, including tool name, description, version number (v1.0), author information, creation time, business data structure definition model, etc., configures the tool's service access address, and supports both STDIO and SSE transmission methods.

[0075] The access permissions for the official vehicle application system tool are set to be visible only within the organization. Tags such as "Official Vehicle" and "Form Automation" are added for easy retrieval. A health check mechanism is configured to periodically verify tool availability. Version management is established to support subsequent feature iterations and bug fixes. Upon successful registration, the tool receives a unique tool identifier, "submit-car-apply".

[0076] Step 5: Digital Employee Intelligent Agent Collaboration.

[0077] In this embodiment, the user writes prompts for the digital employee agent and defines its role as "Official Vehicle Application Assistant," whose responsibility is to assist the user in completing administrative tasks such as applying for official vehicles and booking meeting rooms. The user browses the list of available tools from the MCP registry center, searches using the keyword "official vehicle," finds the submit-car-apply tool, reviews its detailed description and usage examples, and then adds the tool to the agent configuration.

[0078] For example, the official vehicle application system deploys a digital employee intelligent agent, which automatically loads the description information, call interface, and corresponding business data structure definition model of the selected tool, and automatically injects the business data structure definition into the prompts of the digital employee, enabling the digital employee to have a structured understanding of the business data of the official vehicle application. Users interact with the intelligent agent through natural language: "I need to leave from school at 10 am tomorrow to attend a meeting at the municipal government, and I expect to return at 6 pm. I need a bus for 30 people, including 20 students, and the funds will be from a special off-campus expenditure."

[0079] In this embodiment, the digital employee intelligent agent uses a three-layer decision-making architecture of "intent-planning-execution" to process user requests:

[0080] Intent Layer: The agent receives the user's natural language instructions, invokes a large language model for semantic understanding, and identifies that this is a request for official vehicle use. The agent extracts business entity information from the user's statement: 30 passengers, 20 students, bus type, funding source from an off-campus project, departure time 10:00 AM tomorrow, return time 6:00 PM, boarding location at the school, destination at the municipal government, and adds the reason for the use as "attending a municipal government meeting." Based on the business data structure definition injected in the prompt, the agent structures the extracted business data and persists it to a document-oriented database (such as MongoDB), establishing an index linking the business data to the current task.

[0081] Planning layer: Based on the structured intent output by the intent layer, the agent filters candidate tools from the list of loaded tools using a semantic matching algorithm, determines the use of the submit-car-apply tool, generates a tool invocation scheme and parameter mapping, and converts the extracted information into the parameter format required by the tool.

[0082] Execution Layer: The agent executes tasks through the task queue management tool, queries and loads the structured business data required for the current task from the document-oriented database, and passes the business data as parameters to the submit-car-apply tool for execution. After the tool returns a successful result, the agent updates the task status in the document-oriented database and converts the structured data into natural language to provide feedback to the user: "Your official vehicle application has been successfully submitted. Application details: Depart from the school to the municipal government at 10:00 AM tomorrow and return at 6:00 PM, using a bus, for a total of 30 people (including 20 students), funded by off-campus special expenditures. The system will arrange vehicles as soon as possible, please pay attention to the approval progress."

[0083] The entire interaction process is natural and smooth. Users do not need to understand the specific form fields and filling rules. They can complete complex business processes through simple natural language descriptions.

[0084] Example 3

[0085] This embodiment demonstrates how to use a digital employee agent to invoke multiple MCP tools to complete complex cross-system business processes, achieving end-to-end automation of cross-system procurement approval processes. Assume the enterprise has generated and registered the following three MCP tools through steps S1-S4:

[0086] erp-purchase-order-create: Create a purchase order in the ERP system.

[0087] oa-approval-submit: Submit the approval process to the OA system.

[0088] supplier-system-notify: Sends order notifications to the supplier system.

[0089] The user interacts with the purchasing assistant AI: "Please purchase 100 A4 cardboard boxes. The supplier is XX Company. The budget is 5,000 yuan. It needs Manager Wang's approval."

[0090] In this embodiment, the digital employee intelligent agent uses a three-layer decision-making architecture of "intent-planning-execution" to process user requests:

[0091] Intent Layer: The digital employee agent invokes a large language model for semantic understanding, extracting business entities (item: A4 cardboard box, quantity: 100, supplier: XX company, budget: 5000 yuan, approver: Manager Wang) and operational intent (purchase application and approval) from the user's expression. Based on the procurement business data structure definition injected into the prompts, the agent performs structured processing on the extracted business data and persistently stores it in a document-oriented database.

[0092] Planning Layer: Based on the structured intent output from the intent layer, the digital employee agent selects three relevant tools from the list of loaded tools, analyzes the dependencies and data flow requirements between the tools, determines the calling order as first creating an order, then submitting for approval, and finally notifying the supplier, and establishes parameter mapping relationships (the order number output of the first tool is mapped to the order number input of the latter two tools), generating a tool calling sequence.

[0093] Execution Layer: The digital employee agent adds the three tool invocation tasks to the task queue and executes them sequentially according to their dependencies. The execution layer retrieves tasks from the task queue in turn and loads structured business data from the document-oriented database: First, it calls the erp-purchase-order-create tool to create a purchase order in the ERP system, passing in parameters such as item name, quantity, supplier, and budget, obtaining the order number and updating it in the database; then, it calls the oa-approval-submit tool to submit an approval application, loading the order number from the database and specifying Manager Wang as the approver; finally, it calls the supplier-system-notify tool to send an order notification to the supplier system, passing in the order number and supplier identifier.

[0094] During execution, the digital employee agent monitors the execution status of each tool. If a step fails, it retryes or rolls back according to a preset strategy. After all tools have executed successfully, the digital employee agent updates the task status in the document database, summarizes the execution results, and provides feedback to the user: "Purchase order has been created (order number: PO202401001), the approval process has been submitted to Manager Wang, and the supplier has received the order notification. You can check the approval progress and order status through the system."

[0095] It achieves end-to-end automation of cross-system business processes. The entire process requires only a single natural language description from the user, eliminating the need for manual operation across multiple systems.

[0096] Performance Verification: As can be seen from the above embodiments, the automated digital employee construction method based on self-learning intelligent agents provided by this invention successfully realizes the complete process from capability requirement description to automatic tool generation, and then to rapid deployment of digital employees. Compared with the prior art, this invention has the following significant advantages:

[0097] First, the efficiency of building digital employees has been significantly improved. In Example 2, the construction of the official vehicle application assistant took only 3 hours (2 hours for tool generation and 1 hour for digital employee configuration), which is more than 95% more efficient than the traditional manual development method (which usually takes 1-2 months).

[0098] Second, the tools are highly standardized. All generated tools conform to the MCP protocol specification, and the submit-car-apply tool can be called by any digital employee application that supports MCP, achieving true tool reuse and interoperability.

[0099] Third, the stability of the tool has been significantly improved. The intelligently optimized script has capabilities such as parameterization, multi-strategy positioning, exception handling, and intelligent retries. In actual tests of official vehicle application scenarios, the success rate reached 98%, and the stability was improved by more than 40%.

[0100] Fourth, the deployment threshold for digital employees has been significantly lowered. Business personnel do not need programming skills; they only need to write prompts and select tools to build digital employees. Example 3 demonstrates how a combination of tools can empower digital employees with cross-system collaboration capabilities, truly achieving business automation.

[0101] Fifth, it supports complex business scenarios. Digital employees can chain together multiple tools to automate end-to-end business processes. The form filling in Example 1 involves 9 fields and 6 control types, fully verifying the completeness and versatility of the method. This invention can be applied to multiple industries such as education, healthcare, finance, government affairs, and manufacturing, providing technical support for various enterprises to quickly build digital employees.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatically building digital employees based on self-learning of business systems, characterized in that: This includes self-learning integrated intelligent agents and digital employee intelligent agents, and the method includes the following operational steps: Step S1: The self-learning integrated intelligent agent receives the natural language requirement description of the business system, calls the large language model for intent understanding, extracts the target business system address, operation sequence, key data fields and expected output format, and generates a structured tool development plan. Step S2: The self-learning integrated intelligent agent adopts a three-stage autonomous learning strategy of "exploration-learning-optimization". It autonomously explores the operation path of the business system, learns the semantics and constraint relationships of business data fields, builds a structured business data model, and automatically debugs and optimizes the generated scripts to generate parameterized scripts. Step S3: The self-learning integrated agent encapsulates the optimized parameterized script into a standardized tool conforming to the MCP protocol specification, and at the same time encapsulates the structured business data model into the tool's data structure definition document; Step S4: The self-learning integrated intelligent agent registers the standardized tools to the MCP tool registry center for unified management and reuse. Step S5: The digital employee intelligent agent loads the tool and business data structure definition from the MCP tool registry center, and uses the "intent-planning-execution" three-layer decision architecture to intelligently call tools and complete the execution of business tasks.

2. The method for automatically constructing digital employees based on self-learning of business systems according to claim 1, characterized in that: The self-learning integrated intelligent agent possesses autonomous learning capabilities and automatic debugging capabilities; The self-learning capability is used to automatically learn the operation methods of business systems, autonomously generate business operation tools, and standardize and encapsulate them. The automatic debugging capability is used to automatically verify and execute the generated script during the script generation process. When an element location failure is detected, it automatically tries alternative location strategies. When an operation execution timeout is detected, it automatically adjusts the waiting strategy. When a business process anomaly is detected, it automatically analyzes the cause and attempts to fix it. The debugging process is iterative until the script runs stably. The digital employee agent executes business tasks by calling tools.

3. The method for automatically constructing digital employees based on self-learning of business systems according to claim 2, characterized in that: Step S1, which involves calling a large language model for intent understanding, includes: Large language models identify key information in the natural language of business systems: the name of the target business system, the type of operation to be performed, and the data objects involved. The business process is broken down into ordered atomic operation steps, and the key data fields involved in the operation, their data types, and constraints are identified.

4. The method for automatically constructing digital employees based on self-learning of business systems according to claim 3, characterized in that: In step S2, the self-learning ensemble agent adopts a three-stage autonomous learning strategy of "exploration-learning-optimization," including: During the exploration phase, the browser automation control tool is activated to autonomously explore the operation path and interaction logic of the business system through interface interaction, and simultaneously capture the operation sequence and page state. During the learning phase, the large language model is used to conduct in-depth analysis of the exploration results, autonomously learn the semantics and constraint relationships of business data fields, and build a structured business data model. During the optimization phase, the generated scripts are automatically debugged and parameterized refactored, including automatically identifying and fixing location failures, automatically handling abnormal processes, and automatically optimizing execution paths, generating parameterized scripts with generalization and self-healing capabilities.

5. The method for automatically constructing digital employees based on self-learning of business systems according to claim 4, characterized in that: The exploration phase also includes: the self-learning integrated intelligent agent starts the browser automation control tool, attempts to autonomously explore the operation path of the business system through interface element recognition and interaction, synchronously captures the operation sequence, records the target element, operation type, input data, page state changes and corresponding page DOM snapshots of each operation, and generates the original operation script.

6. The method for automatically constructing digital employees based on self-learning of business systems according to claim 5, characterized in that: The learning phase also includes: the self-learning integrated agent inputs the original script and DOM snapshot information obtained from exploration into the large language model for in-depth analysis, autonomously learns the semantic meaning, data type, constraints and relationships between fields of business data, and constructs a structured business data model.

7. The method for automatically constructing digital employees based on self-learning of business systems according to claim 6, characterized in that: The optimization phase also includes: the self-learning integrated intelligent agent automatically debugging and reconstructing the original script, replacing business data fields with parameterized variables, removing redundant operations, establishing a multi-strategy locator alternative chain, adding intelligent exception handling and automatic retry mechanisms, and generating a parameterized script with self-healing capabilities.

8. The method for automatically constructing digital employees based on self-learning of business systems according to claim 7, characterized in that: The standardized tools encapsulated in step S3 include: generating a tool description JSON, parameter schema definition, execution logic, and error handling mechanism; Encapsulated data structure definition document: used to describe the business data fields, field types, constraints, and relationships between fields involved in the tool.

9. The method for automatically constructing digital employees based on self-learning of business systems according to claim 8, characterized in that: In step S4, the self-learning integrated intelligent agent registers the standardized tools to the MCP tool registry center, including: uploading tool metadata and business data structure definition documents, configuring service access addresses and authentication methods, setting access permissions and visibility scope, establishing a version management mechanism, and configuring health checks and performance monitoring indicators.

10. The method for automatically constructing digital employees based on self-learning of business systems according to claim 9, characterized in that: Step S5 employs a three-tiered decision-making framework of "intent-planning-execution," which includes: Intent Layer: The digital employee intelligent agent receives the user's natural language instructions, calls the large language model for semantic understanding, extracts business entity information and operation intent, performs structured processing on the extracted business data according to the business data structure definition, and persists it to a document-based database. The document-based database is used to persistently store business data, establish an index linking business data and tasks, and support structured querying and updating of business data. Planning layer: Based on structured intent, the digital employee intelligent agent filters candidate tools from the list of loaded tools using semantic matching algorithms, analyzes the dependencies and data flow requirements between tools, and generates tool call sequences and parameter mapping schemes; Execution layer: The digital employee intelligent agent executes tasks through task queue management and scheduling tools. The task queue maintains a list of tasks to be executed, task execution status, and dependencies between tasks, supporting both serial and parallel execution of tasks. The execution layer queries and loads the structured business data required for the current task from the document-based database, passes the business data as parameters to the MCP tool for execution, updates the task status and business data in the database after obtaining the execution result, and retryes or rolls back abnormally according to the preset strategy when the tool fails to execute.

Images