Method and system for arranging agent process workflow

Abstract

The invention provides a method for arranging an agent process workflow, and belongs to the field of workflow arrangement. The method defines a service through an agent and arranges a workflow execution flow direction, and comprises the following steps: S1, selecting to access a self-defined process agent; s2, a flow starting operation related flow variable is introduced, and a flow starting flow is called; s3, judging whether flow starting is successful or not, and if yes, turning to S3; if not, whether the number of times of starting the execution process reaches a threshold value or not is judged, if yes, the step S5 is executed, and if not, the step S2S3 is executed, and the process is executed; s4, judging whether the process is executed successfully or not, and if yes, turning to S4; if not, whether the number of times of flow execution reaches a threshold value or not is judged, if yes, the step S5 is executed, and if not, the step S3 is executed again; s4, calling an API (Application Program Interface), and accessing an intelligent agent; and S5, ending. According to the method, a natural language-driven process is realized and a use threshold is reduced; and the flow process engine and the embedded architecture are decoupled, and the flow diagram is decoupled, so that the extensibility of the process service is improved.

Description

Technical Field

[0001] This invention relates to the field of workflow orchestration technology, specifically to a method and system for orchestrating intelligent agent workflows. Background Technology

[0002] While lightweight workflow services built on the Flowable workflow engine currently in the industry perform well in terms of performance and flexibility, they still have the following shortcomings that affect workflow orchestration: 1. Significant limitations: The flowable workflow engine is an open-source tool from the community, lacking in-depth case studies of complex enterprise-level scenarios. Some advanced features need to be explored independently, and scenario integration is quite limited. 2. High complexity: There are many process configuration items (such as EL expressions and various event listening mechanisms). Beginners need to learn the BPMN2.0 standard in depth to efficiently integrate the flowable process engine. 3. Low scalability: The flowable process engine is driven by flowcharts, which are compiled and fixed according to the BPMN2.0 standard. If there are any changes, the flowcharts need to be recompiled. Summary of the Invention

[0003] The purpose of this invention is to solve the problems existing in the prior art and provide a method for workflow orchestration based on flowable intelligent agents, so as to solve the defects of flowable process engines in terms of limitations, complexity and extensibility with a flexible orchestration structure.

[0004] This invention is achieved through the following technical solution: The first aspect of the present invention provides a method for orchestrating intelligent agent workflows, the method comprising: S1. Select to access a custom process agent; S2: Pass in the relevant process variables for starting and running the process, and call flowable to start the process; determine whether the process started successfully. If it did, proceed to S3; if not, determine whether the number of times the process has been started has reached the threshold. If it did, proceed to S5; if not, return to S2. S3, Execute the process; Determine if the process execution was successful. If yes, proceed to S4; if no, determine if the number of times the process has been executed has reached the threshold. If yes, proceed to S5; if no, return to S3. S4. Call the API to connect to the smart agent; S5, End.

[0005] Furthermore, the access methods in S1 include: directly importing process types and building through intelligent agent tools; directly importing process types involves calling the four underlying process types of Flowable; building through intelligent agent tools involves orchestrating process intelligent agents based on the business defined by the intelligent agent.

[0006] Furthermore, the rules for orchestrating the process intelligence agent include: Process nodes enable approval by one or more people by setting task assignment attributes; conditional expressions are set on the sequence flow, and the approval conditions of process nodes are dynamically configured through these expressions; process nodes use process-level global variables for dynamic setting, and variables can be set at any node through execution listeners or task listeners; process nodes implement Java delegate API calls, which can store process relationships and query related relationships; each process definition has a unique key, each process instance has a unique ID, processes execute independently, and sub-processes or other processes are started by calling the API; a unique identifier is set after each process starts successfully.

[0007] Furthermore, the four process types include: unconditionally controlled serial process; conditionally controlled serial process; conditionally controlled parallel process; and serial process with subprocesses. The execution flow of the unconditional control serial process is a serial line, and the process is executed sequentially until the end. The conditional control serial process is based on the unconditional control serial process, with the addition of conditional control between two processes, and the process switches to different processes based on the judgment result. The conditionally controlled parallel process includes a main path and a parallel part consisting of multiple parallel branches. The parallel branches are executed simultaneously, and the parallel part is completed after all parallel branches have been executed. The serial process of the sub-process includes a main path serial line, and one or more sub-processes are started at the nodes of the serial line.

[0008] Furthermore, in the serial process with sub-processes, the flowable interface includes an API for the parent-child process ID relationship of the serial process with sub-processes; the parent process can query information of all sub-processes through the relationship API; and the sub-processes can query information of their parent process through the relationship API.

[0009] Furthermore, the operation of initiating the process in S2 includes: S21. Construct business data and process data into structured data; S22. Import the structured data into the large language model; S23. Execute startup and return startup result.

[0010] Furthermore, while importing the structured data into the large language model, relevant prompt words are constructed. The prompt words are constructed through a prompt word template, which includes: the functional roles of the large language model, the structured data, the MySQL table and field attributes, and the requirements for the execution behavior of the large language model.

[0011] Furthermore, the operation of S23 is to execute prompt words of the large language model, generate SQL statements, execute SQL statements, and return and store the execution results.

[0012] Furthermore, the operation of S3 also includes the process agent calling the process engine listening API to notify the process agent of the execution result of the process node, the process node ending and waiting, and selecting to execute the next node based on the execution result.

[0013] A second aspect of the present invention provides a system for orchestrating intelligent agent workflows to perform the method, the system comprising: The client module is used to visualize and operate various workflows, interact with users, and pass data to other modules; The large language model module provides reasoning capabilities and converts natural language into SQL statements, executing SQL statements; The process intelligence agent module is used to build process intelligence agents and call large language models; The process engine module is used to complete various processes by calling process agents through APIs for initiation, approval, and monitoring. The logging module is used to record API calls and various operation logs during the execution of the process engine module and client module. The data module is used to persistently store page data, process data, log data, and business data generated during the execution of the process engine module.

[0014] The beneficial effects of this invention are: the intelligent agent uses a large language model to convert the user's natural language intent, realizes natural language-driven processes, and lowers the threshold for use; the flowable process engine is decoupled from the embedded architecture, and the intelligent agent can independently execute various programming languages ​​or call interfaces, so that it no longer depends solely on a certain programming language; multiple intelligent agents cooperate to execute processes, decoupling the flowchart from the flowable process engine, which greatly improves the scalability of process services. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the design concept of the intelligent agent orchestration workflow of this invention. Figure 2 This is a schematic diagram of the workflow execution for an example of an unconditionally controlled serial process in this invention; Figure 3This is a schematic diagram illustrating the workflow execution of an example of conditionally controlled serial processes in this invention. Figure 4 This is a schematic diagram of workflow execution, illustrating an example of conditionally controlled parallel processes in this invention. Figure 5 This is a schematic diagram illustrating the workflow execution of a serial process with sub-processes in this invention. Figure 6 This is an architecture diagram of the system in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings: This invention utilizes process intelligence agents to orchestrate workflows.

[0017] To orchestrate workflows using process intelligence agents, it is necessary to define the process execution procedure. This invention designs four commonly used process types, such as... Figure 2 , Figure 3 , Figure 4 , Figure 5 As shown, but not limited to, the nodes and condition controls defined in the diagram. The four flow types include: unconditional control serial flow; conditional control serial flow; conditional control parallel flow; and serial flow with sub-flows.

[0018] In this invention, the above-mentioned process intelligent agent orchestration workflow must follow the following process rules: 1. All process nodes support approval by one or more people. One or more people can approve by setting the task assignment attribute. At the same time, multi-instance is used to realize scenarios where multiple people need to approve.

[0019] 2. The approval conditions for process nodes can be dynamically configured. Specifically, it supports setting conditional expressions on the sequence flow. The expressions can be dynamically configured, and the approval conditions of the process can be automatically controlled by passing in the conditions. It supports, but is not limited to, conditions such as one vote to pass, unanimous vote to pass, and half vote to pass. The approval result is passed through process variables.

[0020] 3. The process nodes support dynamically setting variables, using process-level global variables, and each node variable takes effect. Specifically, variables can be set at any node, through execution listeners or task listeners, or directly when the node is completed.

[0021] 4. The process nodes support API calling capabilities, which can be achieved by implementing Java delegates or using expressions to call Flowable's API, such as the underlying Flowable API.

[0022] 5. Each process executes independently, with start and end nodes, yet it can be associated with and referenced by other independent processes. Specifically, process relationships can be stored in API calls, and these relationships can also be retrieved through API calls. Each process definition has a unique key, and each process instance has a unique ID. Sub-processes can be started or other processes can be called within a process by calling the API; furthermore, process variables or custom tables are used to store the relationships between processes.

[0023] 6. Each process is assigned a unique identifier after it is successfully initiated, called the process ID.

[0024] 7. Requirements for adding / deleting nodes and controlling conditions in the workflow, including: ① If no nodes are added in a parallel process, add nodes as needed; if nodes are added in a parallel process, adding a node in a parallel block adds a parallel branch, and it must be ensured that the branch can complete execution; adding nodes in a non-parallel block is the same as in a serial process, i.e., add nodes as needed. Specifically, a parallel gateway is used to create parallel branches: when adding a node, if it is added after the parallel gateway, a new branch needs to be added, and it must be ensured that the branch can merge into the same parallel gateway.

[0025] ② When deleting a node, it is necessary to ensure that there is at least one node besides the start and end points; if there is a conditional judgment, it is necessary to ensure that there are at least two nodes. Deleting a node within a parallel block must ensure that the parallel branch can complete its execution. Specifically, when deleting a node, the process structure needs to be checked. If the node being deleted is in a parallel branch, it must be ensured that the branch cannot merge. The node is deleted through the model API, and the connection lines are automatically adjusted.

[0026] ③ Conditional controls can flow to nodes or other conditional controls, but the process must be able to complete after adding or removing conditional controls. ④ Whether adding or removing nodes or using conditional control, the process can be completed regardless of which execution direction it follows. Conditional control is usually implemented by an exclusive gateway. An exclusive gateway can connect multiple sequence flows, each of which can have conditions set, these conditions can be dynamically modified, and at least one default flow direction (unconditional) can be ensured to guarantee the completion of the process.

[0027] ⑤ A complete process must have start and end nodes to ensure that the process can be completed. After the start node, there can be a node or a condition control. Both nodes and condition controls can lead to the end node to complete the process execution.

[0028] ⑥ There are three types of connections between nodes and conditional controls: First, nodes connect directly to each other; second, a node connects to a conditional control, where the conditional control requires a default execution direction, and other directions execute according to the passed-in conditions; third, conditional controls connect to each other, where both require a default execution condition, and other directions can flow to nodes or other conditional controls. Similarly, all nodes and conditional controls must meet these three requirements to form various workflows. Since nodes (tasks, gateways, etc.) are connected via sequence streams, conditional control can be implemented through exclusive gateways. An exclusive gateway must have a default sequence stream (unconditional), while other sequence streams can have conditions set. Nodes can connect directly to other nodes or to gateways, and gateways can then connect to other gateways or nodes.

[0029] Furthermore, the workflow orchestration using process agents described in this invention defines the overall execution flow of the workflow. In the orchestrated workflow, the execution code tool calls APIs related to the underlying capabilities of Flowable (such as initiation, approval, termination, listening, and logging capabilities), completing the workflow execution flow by calling these APIs. The agent tool connects to the agent, passing in the necessary parameters to enable the agent to perform its related functions; the agent contains the business logic of each process. This flexible orchestration structure addresses the limitations, complexity, and scalability deficiencies of the Flowable workflow engine.

[0030] Example 1: This invention provides a method for orchestrating intelligent agent workflows.

[0031] The overall steps are as follows Figure 1 As shown, it includes: S1. Select Flowchart: Users can choose to integrate a custom process agent. Integration methods include: directly importing the process type, and building it on the page using existing agent tools. Importing a flowchart directly calls the underlying capabilities of Flowable to complete the execution flow of the four flowcharts designed in this invention, and the business is executed according to the flowchart. Building it using existing agent tools completes the process agent orchestration based on the business defined by the agent, and other agents can be integrated into the agent.

[0032] S2. After selecting the flowchart in S1, input the relevant process variables for starting and running the process. These variables include approver variables, process condition control variables, and process attribute variables, all of which need to be input by the agent according to different scenarios. Here, the agent calls the flowable's process start API and obtains the process variable parameters. Based on the parameters, flowable starts the process. Determine whether the process start is successful. If it is, proceed to S3. If not, retry the start. If the start fails after multiple attempts (the threshold for the number of retry attempts can be preset according to business needs), proceed to S5. S3: This node enters the execution phase until the entire process is completed. Determine if the process execution was successful. If successful, proceed to S4; otherwise, execute according to the preset retry method. If it still fails after multiple attempts (the threshold for the number of retries can be preset according to business requirements), proceed to S5.

[0033] S4. The process has been executed successfully. You can call the API and connect to the smart agent here.

[0034] S5, End.

[0035] Specifically, in combination Figure 1 The specific steps of the workflow orchestration technology solution of the present invention are described below: S1, the client module displays various processes constructed in the process intelligence agent module for the user to select. This invention defines four commonly used process intelligence agents, combined with... Figure 2 Figure 3 Figure 4 Figure 5 The following is a detailed explanation: Figure 2 The process shown is an unconditionally controlled serial process, that is, after starting, processes A, B, C, etc. are executed sequentially until the end. The execution direction of this type of process is a serial line, and the number of nodes is not limited to the definition in the figure, and can be increased or decreased according to the use case. Figure 3 The process shown is a conditionally controlled serial process. The execution body of this type of process follows the... Figure 2 The process is the same, in Figure 2 Based on the process, condition control is added, that is, condition judgment is added between two processes, and different processes are entered according to the judgment result. The number of nodes, the number of condition controls, and the position of condition controls are not limited to the definition in the diagram, and can be added or removed according to the use scenario. Figure 4 The illustrated process is a conditionally controlled parallel process. The main execution flow of this type of process is a parallel line. The parallel section supports parallel branches that execute tasks simultaneously. The rule within the parallel section is that parallel branches start simultaneously, and the parallel section is considered complete only after all parallel branches have finished executing. Furthermore, there are also conditionally controlled processes. The number of nodes, the number of parallel processes, the number and position of parallel nodes, and the location of conditional controls are not limited to the definitions in the diagram and can be added or removed according to the use case. Specifically, this includes: 1. If there is no parallel process to add nodes, add them as needed; if there is a parallel process to add nodes, adding a node in a parallel block adds a parallel branch, and it is necessary to ensure that the branch can be executed to completion; adding nodes in non-parallel blocks is the same as in serial processes, that is, add them as needed.

[0036] 2. When deleting a node, if it is necessary to ensure that there is at least one node other than the start and end points, or if there is a conditional judgment, then at least two nodes must be guaranteed. Note that deleting a node within a parallel block must ensure that the parallel branch can complete execution.

[0037] 3. Conditional controls can flow to nodes or other conditional controls, but it is necessary to ensure that the process can be completed after adding or removing conditional controls.

[0038] 4. Whether adding or removing nodes or controlling conditions, the process can be completed regardless of which execution direction it follows.

[0039] Figure 5 The illustrated flow is a serial flow with sub-flows. The execution path of this type of flow (parent flow) is a sequential line. One to N sub-flows can be initiated at each node. The sub-flows can be of the various flow types mentioned above, or they can be custom flow types. The number of nodes, the number of sub-flow nodes, and their positions are not limited to those defined in the diagram and can be increased or decreased according to the use case. Furthermore, the Flowable interface provides an API for establishing parent-child flow ID relationships within the serial flow with sub-flows. The parent flow can query information about all sub-flows through this relationship API; similarly, a sub-flow can query information about its parent flow through the same relationship API.

[0040] The four commonly used process agents defined above can be modified by users according to their scenarios to create new process agents. The execution process of the modified process must follow the defined process rules and be stored in the workflow orchestration tool.

[0041] S2, after S1 completes the process selection, inputs the result into the process engine module. The process engine module calls the flowable API to start the process and pass in the process variables. After the process starts successfully, it needs to store the client input data passed in by S1, as well as the process data generated by flowable (the specific process selected in S1 and related parameters, the data type is also JSON key-value pairs) into the data module. This process calls the large language model module to generate SQL using natural language and store it.

[0042] The implementation process is as follows: S21. Construct business data and process data into structured data that can be recognized by a large language model.

[0043] Specifically, the flowable process data is stored in a database, and the data passed in from S1 is also stored in a database, linked by a unique process ID. The structured data stored in the database here refers to JSON key-value pair data, where the key is the field name and the value is the field value.

[0044] S22. Import the structured data constructed in step 21 into the large language model.

[0045] When importing data, relevant prompts need to be constructed. This invention uses prompt templates to constrain the divergent behavior of the large language model, ensuring it focuses on executing prompt instructions. The template is as follows: (1) Role Role descriptions help large language models better define their functional roles. (2) Data content The structured data constructed in step 21 (3) MySQL table and field attributes This includes descriptions of MySQL database tables and fields.

[0046] (4) Requirements The requirements for the execution behavior of the large language model are described here. The output requirements must include: the output must be SQL type data.

[0047] S23. Execute startup and return startup result.

[0048] Specifically, the large language model executes prompt words by calling a tool, generates SQL, and returns the startup results, which are stored in the data module. The returned result stream then flows to the next node in the workflow.

[0049] If the process starts successfully and the business data and process data are stored successfully, then the data will be transferred to S3.

[0050] In S3, after the process is initiated in S2, the workflow enters the execution phase. This process mainly involves the user approving the workflow at each node of the workflow agent in the client module and generating execution results. After all workflow nodes of the workflow agent have been approved, the workflow agent calls the workflow engine's listening API according to the defined workflow rules. The workflow engine notifies the node of the workflow execution results, and the node ends the waiting phase and moves on to the next node.

[0051] S4 uses the execution result requirements based on the variables input by the execution client module.

[0052] Extract the execution result data of S3, and determine whether the process agent selected by S1 needs to use the execution result. If so, output the execution result and the agent uses the result by calling the API. If the execution result does not need to be used, then flow to the end node S5.

[0053] S5, end, and return the result to the client module.

[0054] Example 2: The present invention also provides a system for orchestrating workflows using flowable intelligent agents.

[0055] The system comprises the following modules, and the relationships between the modules are as follows: Figure 6 As shown: The client module is used to visualize and operate various workflows, interact with users, and pass data to other modules; specifically, it is used for selecting, initiating, and approving workflows, and passing data to the workflow engine module, workflow agent module, and data module.

[0056] The large language model module provides powerful reasoning capabilities and natural language conversion query functions. It can call the data module through database tools to execute SQL. It receives input data from either the process agent module or the data module. The input data is in JSON format, and the output data is formatted according to the prompts. The output is either stored in the data module or circulated within the process agent.

[0057] The process intelligence agent module is used to build various process intelligence agents, utilize the reasoning capabilities of the large language model, and convert natural language into SQL statements according to prompts. Specifically, it receives user interaction input from the client module; receives output results from the large language model module; receives API call results from the process engine module; and receives queried data from the data module. After processing, it continues to transmit the data according to the process intelligence agent workflow.

[0058] The process engine module generates variables required for process initiation, approval, and monitoring. It calls the process intelligence module through API calls for initiation, approval, and monitoring to complete various process workflows and return the results to the process intelligence module.

[0059] The logging module is used to record API calls and various operation logs of the process engine module and client module.

[0060] The data module is used to persistently store page data, process data, log data, etc. generated during the execution of the process engine module, as well as business data that can be called by the process engine module, the large language model module, the log module, etc.

[0061] The above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the technical solutions described in the specific embodiments of the present invention. Therefore, the foregoing description is only preferred and not restrictive.

Claims

1. A method for orchestrating intelligent agent workflows, comprising defining business processes through intelligent agents and orchestrating workflow execution flow, characterized in that: The method includes: S1. Select to access a custom process agent; S2: Pass in the relevant process variables for starting and running the process, and call flowable to start the process; determine whether the process started successfully. If it did, proceed to S3; if not, determine whether the number of times the process has been started has reached the threshold. If it did, proceed to S5; if not, return to S2. S3, Execute the process; Determine if the process execution was successful. If yes, proceed to S4; if no, determine if the number of times the process has been executed has reached the threshold. If yes, proceed to S5; if no, return to S3. S4. Call the API to connect to the smart agent; S5, End.

2. The method for orchestrating intelligent agent workflows according to claim 1, characterized in that: The access methods in S1 include: directly importing process types and building through intelligent agent tools; directly importing process types involves calling the four underlying process types of Flowable; building through intelligent agent tools involves orchestrating process intelligent agents based on the business defined by the intelligent agents.

3. The method for orchestrating intelligent agent workflows according to claim 2, characterized in that: The rules for orchestrating the process agents include: Process nodes enable approval by one or more people by setting task assignment attributes; conditional expressions are set on the sequence flow, and the approval conditions of process nodes are dynamically configured through these expressions; process nodes use process-level global variables for dynamic setting, and variables can be set at any node through execution listeners or task listeners; process nodes implement Java delegate API calls, which can store process relationships and query related relationships; each process definition has a unique key, each process instance has a unique ID, processes execute independently, and sub-processes or other processes are started by calling the API; a unique identifier is set after each process starts successfully.

4. The method for orchestrating intelligent agent workflows according to claim 2 or 3, characterized in that: The four process types include: unconditional control serial process; conditional control serial process; conditional control parallel process; and serial process with sub-processes. The execution flow of the unconditional control serial process is a serial line, and the process is executed sequentially until the end. The conditional control serial process is based on the unconditional control serial process, with the addition of conditional control between two processes, and the process switches to different processes based on the judgment result. The conditionally controlled parallel process includes a main path and a parallel part consisting of multiple parallel branches. The parallel branches are executed simultaneously, and the parallel part is completed after all parallel branches have been executed. The serial process of the sub-process includes a main path serial line, and one or more sub-processes are started at the nodes of the serial line.

5. The method for orchestrating intelligent agent workflows according to claim 4, characterized in that: In the serial flow with sub-flows, the flowable interface includes a parent-child flow ID relationship API; the parent flow can query information about all sub-flows through the relationship API; and the sub-flows can query information about their parent flow through the relationship API.

6. The method for orchestrating intelligent agent workflows according to claim 5, characterized in that: The operations for initiating the process in S2 include: S21. Construct business data and process data into structured data; S22. Import the structured data into the large language model; S23. Execute startup and return startup result.

7. The method for orchestrating intelligent agent workflows according to claim 6, characterized in that: The structured data is imported into a large language model while relevant prompt words are constructed. The prompt words are constructed through a prompt word template, which includes: the functional roles of the large language model, structured data, MySQL tables and field attributes; Requirements for the execution behavior of large language models.

8. The method for orchestrating intelligent agent workflows according to claim 7, characterized in that: The operation of S23 is to execute prompt words of the large language model, generate SQL statements, execute SQL statements, and return and store the execution results.

9. The method for orchestrating intelligent agent workflows according to claim 8, characterized in that: The operation of S3 also includes the process agent calling the process engine listening API to notify the process agent of the execution result of the process node, the process node ending and waiting, and selecting to execute the next node according to the execution result.

10. A system for orchestrating intelligent agent workflows, the system performing the method as described in any one of claims 1 to 9, characterized in that: The system includes: The client module is used to visualize and operate various workflows, interact with users, and pass data to other modules; The large language model module provides reasoning capabilities and converts natural language into SQL statements, executing SQL statements; The process intelligence agent module is used to build process intelligence agents and call large language models; The process engine module is used to complete various processes by calling process agents through APIs for initiation, approval, and monitoring. The logging module is used to record API calls and various operation logs during the execution of the process engine module and client module. The data module is used to persistently store page data, process data, log data, and business data generated during the execution of the process engine module.

Images