Business workflow dynamic arrangement and optimization method based on large model agent

Abstract

The invention provides a business workflow dynamic arrangement and optimization method based on a large model agent, and the method comprises the steps: (1) analyzing the features and operation logic of a multi-element business workflow, and defining a business workflow description language; and (2) decomposing a complex business process construction process, and providing a layered process modeling system to guide a large model agent to construct a workflow model. And (3) constructing a task execution queue based on the task dependency relationship, and dynamically executing and monitoring the workflow. And (4) according to the execution log of the business workflow, discovering the bottleneck, redundancy and abnormity of the workflow model by adopting a process mining algorithm, and proposing a process improvement and optimization strategy. And (5) constructing a collaboration framework of a plurality of large model agents, and realizing dynamic arrangement, monitoring and optimization of the business workflow. According to the method, the defects that the current business workflow construction depends on experience of designers, and effectiveness and function requirements are difficult to guarantee are overcome; the defect that the business workflow is difficult to adapt to dynamic demand changes is overcome.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and business process management technology, specifically relating to a method for dynamic orchestration and optimization of business workflows based on large-scale intelligent agents. Background Technology

[0002] A workflow is a set of collaborative tasks that are executed, fully or partially, according to predefined rules, forming an automated process. The executors of these tasks are typically people, devices, or programs. A workflow model is a way of describing a workflow, defining different attributes for each task, including task name and description, dependencies, control conditions, and resource constraints. Control conditions generally include sequential, loop, and conditional conditions; resource constraints include task deadlines and resource allocation. Resources are the personnel, equipment, or programs needed to complete the task, and resource allocation is the relationship between tasks and resources, which can be optimized using scheduling algorithms.

[0003] In the process of enterprise digital transformation, business process management systems (BPMs) are widely used for process modeling, task execution, and process monitoring. Business workflows, as the core of BPMs, play a crucial role in several key areas. In manufacturing, from raw material procurement and production processing to supply chain sales, complex process connections and production scheduling are involved. Workflow technology ensures close connection and coordination among all links, improving production efficiency. In logistics and transportation, from order processing and warehousing management to transportation and delivery, real-time responses to changes in demand are required. Workflow technology can visualize and dynamically monitor each process node, improving transportation efficiency. The construction and optimization of business workflows are crucial for enterprises to improve business efficiency, reduce operating costs, and enhance core competitiveness, making them key to enterprise digital transformation.

[0004] The most widely used enterprise process modeling method is BPMN (Business Process Management Initiative) modeling language. It requires technical personnel to have professional knowledge of process modeling and manually construct task flows, gateways, and control logic. The workflow model obtained by BPMN modeling can be directly executed by the workflow engine after transformation. That is, the workflow engine can identify the business intent and control logic in the model to automatically run the business process. BPMN contains 5 basic graphical symbols: (1) Flow objects: the basic elements for building business processes, including events, tasks, and gateways; (2) Data: data information in the business process, divided into data input, data output, data objects, and data storage; (3) Connection objects: describing the dependencies and information transmission between flow objects, including sequential flow, message flow, and association; (4) Swimlanes: organizing and classifying cross-departmental flow objects, including pools and swimlanes; (5) Annotations: annotation information related to the business process, which does not affect the control logic of the business process. Using this modeling method, complex business processes can be graphically represented, and different tasks across departments can be flexibly arranged and managed to achieve cross-domain collaboration of workflows.

[0005] Current workflow modeling and optimization processes still face problems and challenges. In terms of process construction, existing business workflows largely rely on the experience of technical experts, making it difficult to guarantee the effectiveness and completeness of functional requirements. Regarding process execution, manual modeling typically uses fixed templates, which cannot cope with complex and dynamic business needs. In actual business operations, external market environments and user demands constantly change, and internal corporate strategy adjustments and business requirement optimizations also impact workflows; therefore, workflows need to have dynamic adjustment capabilities. In terms of process optimization, traditional methods rely on manual analysis of event logs, which is inefficient and easily influenced by subjective factors, making it difficult to accurately locate and analyze problems during process execution, and lacking automated and intelligent process mining tools and optimization methods.

[0006] In recent years, with the development of large language model (LLM) technology, agents based on large models have exhibited a certain degree of human intelligence, capable of using and creating tools, thus enabling intelligent process modeling. Researchers have begun to utilize large model agents to develop natural language-to-workflow systems. For example, some scholars have proposed using Retrieval Augmentation (RAG) technology to improve workflow creation quality, receiving user instructions to retrieve relevant components and generating workflows based on these components. This effectively reduces the illusion problem of large models and allows for workflow generation based on simple business scenarios and requirement descriptions. However, in handling complex instructions and logic, there is still a gap compared to human performance, and single large model agents are insufficient for constructing real-world business workflows. Therefore, multi-agent systems are introduced, utilizing multiple LLM agents with different roles and tasks to facilitate interaction and achieve more complex business workflow construction methods. However, existing multi-agent systems are typically designed for specific tasks and cannot be directly applied to the construction of natural language-to-workflow models. Summary of the Invention

[0007] One objective of this invention is to address the problem of constructing intelligent, dynamically adjustable workflows and fill the gap in research on workflow models in multi-agent systems. This invention proposes a method for dynamic orchestration and optimization of business workflows based on large-scale intelligent agents. This method defines a business workflow description language based on the characteristics and operational logic of diverse business workflows. A hierarchical process modeling system is designed, breaking down complex tasks layer by layer to guide large-scale intelligent agents in constructing workflow models. In workflow execution, task execution queues are constructed based on task dependencies, and the workflow execution process is dynamically monitored. Process mining algorithms are used to analyze the generated workflow event logs, addressing bottlenecks, redundancies, and anomalies. All of the above is achieved through a constructed multi-agent framework, defining the role and tasks of each agent, and enabling dynamic orchestration and optimization of business workflows through collaboration.

[0008] The technical solution of this invention is: a method for dynamic orchestration and optimization of business workflows based on large-scale intelligent agents, comprising the following steps: Step 1) Under the Business Process Modeling Notation (BPMN) standard workflow modeling rules, define the code description language for the business workflow, determine the attributes and elements in the workflow structure, and define the content of elements such as tasks, events, gateways, sequence flows, message flows, and data flows. Step 2) Based on the workflow description language defined above, design a layered workflow modeling system that includes a business intent layer, a process structure layer, a collaborative behavior layer, and a data configuration layer. Integrate a confidence assessment and correction mechanism to transform complex unstructured natural language requirements into a structured and executable workflow model.

[0009] Step 3) Based on the workflow model built in Step 2), construct a task execution queue based on task dependencies to drive and monitor the workflow execution process and output workflow execution logs; Step 4) Analyze the workflow logs using process mining algorithms based on the workflow event logs to identify bottlenecks, redundancies, and anomalies in the process; Step 5) Construct a collaborative framework for multiple large-scale intelligent agents to achieve dynamic orchestration, monitoring, and optimization of the workflow. Define the function and collaboration method of each agent. The process design agent outputs a workflow model that meets business requirements based on user text descriptions and the process modeling system; the execution scheduling agent is responsible for the execution and planning of the process and outputs workflow execution logs; the process optimization agent is responsible for analyzing event logs and resource utilization to improve and optimize the workflow, and iteratively outputs the optimal workflow based on user feedback strategies.

[0010] The beneficial effects of this invention are as follows: 1. This invention starts from the characteristics and operational logic of business workflows, defines a business workflow description language, and proposes a layered process modeling system to guide the construction of large-scale workflow models that fit actual business needs. It solves the problem of current workflow construction relying on the experience of technical experts, achieving intelligence, standardization, and scientification.

[0011] 2. To address the limitations of single-agent capabilities, this invention constructs a collaborative framework for multiple large-scale intelligent agents. This framework enables dynamic collaboration, real-time perception of changes in the business environment and requirements, adjustment of task allocation and workflow execution paths, analysis of execution logs, and feedback of improvement strategies. This significantly enhances the accuracy, dynamic adaptability, and optimization capabilities of business workflow construction. Attached Figure Description

[0012] Figure 1 It is a framework for dynamic orchestration and optimization of business workflows based on large-scale intelligent agents.

[0013] Figure 2 This is a case study of the process for generating large models. Detailed Implementation

[0014] 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.

[0015] First, combine Figure 1-2 As shown, the collaborative framework of multiple large-scale intelligent agents constructed in this invention will be described in detail.

[0016] This invention proposes a dynamic orchestration and optimization method for business workflows based on large-scale intelligent agents. This method achieves closed-loop processing from natural language input to process modeling, execution scheduling, and process optimization by constructing a multi-agent system and a large language model collaboration mechanism. It designs a collaborative framework of three large-scale intelligent agents: process analysis, execution scheduling, and process optimization. The overall framework structure is as follows: Figure 1 As shown.

[0017] The process analysis agent, guided by the defined workflow code description and hierarchical process modeling system, generates a preliminary workflow code model that meets business requirements based on the user-input natural language text description. The output workflow code model serves as the input to the execution scheduling agent. The execution scheduling agent is responsible for executing the workflow, monitoring the status of each task in the workflow, automatically executing task nodes based on task dependencies to complete resource scheduling, and generating event logs containing timestamps, task names, resources, and data. The output event logs and the actual workflow logs serve as the input to the process optimization agent. The process optimization agent analyzes bottlenecks, redundancies, and anomalies in the process operation based on process mining algorithms, provides improvement strategies for process structure and resource utilization, and, combined with user feedback, feeds these improvement strategies back to the process design and execution scheduling agents. Iterative optimization enables the workflow to self-adjust and optimize, and the optimized workflow code model is converted into a graph structure as the final output.

[0018] The multi-agent system constructed by this method completes the construction and optimization of workflow through intelligent collaboration, overcoming the problems of low efficiency in process modeling, strong dependence on manual labor, incomplete functional requirements, and poor dynamic adaptability in existing technologies.

[0019] The technical solution of the present invention will now be described in detail, and its specific implementation steps are as follows: Step 1) Under the BPMN standard workflow modeling rules, define the code description language for the business workflow, determine the attributes and elements in the workflow structure, and define the content of elements such as tasks, events, gateways, sequence flows, message flows, and data flows; To enable large-scale intelligent agents to programmatically model and executablely describe business processes, a programmable structured language is used to describe workflows, transforming the large-scale model diagram generation task into a code generation task. A workflow description language based on JSON structure is designed using Python code, referencing the structure and rules of the BPMN 2.0 standard to define the core elements, behavioral rules, and data interaction patterns of the business process. This language can be generated by the large model from natural language parsing and executed directly by a compiler, or it can be converted into a visual flowchart.

[0020] Workflow attributes include ID, name, task, event, gateway, sequence flow, message flow, data flow, pool, and swimlane. A task represents an executable node, which can be a user task, service task, or script task; an event represents the start, end, intermediate, and exception events of the process, used to trigger the execution of a task; a gateway controls process branches, including parallel gateways, exclusive gateways, containment gateways, and event gateways. Sequence flow is the control flow of the process, containing the connections and dependencies between elements in the element list. Message flow represents message passing and communication between pools, reflecting the cross-domain collaboration characteristic of workflows. Data flow represents the input and output relationships of data information between different tasks. Pools represent processes from different domains or departments, and swimlanes represent different departments or personnel involved in a process, dividing elements and reflecting inter-departmental collaboration. The workflow structure is a dictionary structure type represented in JSON format, containing different keys and values. The core structure includes: Metadata: id, name, description; Collaborating entities: pools, lanes; Tasks; Events; Gateway: gateways; Control flow: sequence_flows; Message flow: message_flows; Data stream: data_associations; This workflow structure includes various attributes and operational logic, which can meet the needs of complex business process modeling. Each element supports additional extended attributes to enhance expressiveness and execution capabilities. For example, different attributes can be set for different types of tasks. For user tasks, roles and deadlines can be specified; for script tasks, program scripts can be set; and for service tasks, API tools can be set. These extended attributes allow for more granular control over process behavior during agent construction, execution, and optimization.

[0021] Step 2) Based on the workflow description language defined above, design a layered workflow modeling system that includes a business intent layer, a process structure layer, a collaborative behavior layer, and a data configuration layer. Integrate a confidence assessment and correction mechanism to transform complex unstructured natural language requirements into a structured and executable workflow model.

[0022] Business workflows typically involve cross-organizational business processes, requiring collaboration among different organizations and departments. They describe the relationships between all activities of multiple participating organizations, involving complex task division and control logic identification. To facilitate the construction of workflow code models by large-scale intelligent agents, a hierarchical process modeling system is proposed. This system simulates the cognitive process of experts from abstract concepts to concrete implementations. By integrating confidence assessment and feedback correction mechanisms, the robustness and accuracy of the workflow generation process are ensured.

[0023] A business workflow modeling hint chain consists of four layers: (1) Business Intent Layer: Extract high-level business objectives, core task behaviors and participating subject roles from the user's natural language description to form the initial intent structure for process modeling, providing a semantic basis for the lower-level process structure modeling.

[0024] (2) Process structure layer: Based on the task sketch output by the intent layer, the tasks are decomposed and classified (such as user tasks, service tasks, etc.), and a process control structure is established according to the semantic dependencies between tasks.

[0025] (3) Collaborative Behavior Layer: Based on the existing control structure, identify the collaborative entities in the process (such as user roles, system modules, external services, etc.), and construct a role division model based on pools and lanes and cross-role communication behaviors.

[0026] (4) Data configuration layer: bind specific data input and output objects, interface parameters and service implementation information to each task to complete the coupling modeling between tasks and data.

[0027] The process of modeling natural language into workflow is called NL-W. The initial input natural language is defined as NL, and the output can be represented as:

[0028] in, It is a JSON object; Large models such as GPT have parameters The following generating functions; A preset Prompt containing a four-layer modeling cue chain; For each output, a structure verification and confidence evaluation will be performed. The structure verification mainly checks the correctness of the output syntax. The verification function is defined as follows: , It performs JSON syntax validation to ensure that the LLM output is a grammatically correct JSON fragment; This involves validating business process modeling rules to ensure the generated workflow model structure is correct and complete. Confidence assessment primarily checks the semantic reliability of the output. Combining model output probabilities with multi-path sampling consistency strategies, a confidence assessment function is defined: , It is the average log probability output by the large model; Multi-path sampling consistency function; It is a weighted average result.

[0029] The language model generates a conditional probability distribution when producing output:

[0030] in, It is a sequence of tokens generated by the model; This represents the total number of tokens output. It is the first A sequence of tokens; It was before A sequence of tokens. Therefore, the average logarithmic generation probability of the output is defined. :

[0031] This metric reflects the average confidence level of the large model output; a higher value indicates stronger semantic consistency of the generated content.

[0032] Multi-path sampling consistency assesses the similarity between multiple responses generated by a large model under the same or similar prompts, quantifying the uncertainty of the large model given an input prompt. It involves sampling from the model. Next, generate candidate outputs:

[0033] This represents the nondeterministic structured output of a large model under the same prompt and input conditions.

[0034] For structured JSON output, the elements can be divided into stream objects and connection objects. Stream objects mainly cover the core execution units in the business process, including tasks, events, and gateways. These are directly categorized into their respective collections by parsing the JSON key-value pairs, thus forming task sets. Event Collection and gateway set Connection objects define the interaction and flow relationships between stream objects, specifically including sequential streams, message streams, and data streams, converting each connection object into a triple. It is represented in the form of, where , Represents the source node of the connection. Represents the target node, and It contains the name attribute information of the stream, which can represent conditions, messages, or data. In this way, a set of sequential streams can be constructed. Message Flow Set MF and data stream set The workflow model is transformed into a union set:

[0035] These collection elements are collectively referred to as :

[0036] Then the first The workflow of the subsampled output is represented as a set. The similarity between two workflows is calculated as follows:

[0037]

[0038]

[0039] in, This represents the total number of elements in the two sets. This is a semantic and frequency-based Sorenson-Disser coefficient adjustment method used to calculate the similarity between two lists. First, semantically similar items in the two lists are matched using a vectorized BERT transformer. The cosine similarity of the two sets of words is then assessed; items exceeding a threshold of 0.7 are considered similar. Finally, the multi-path sampling consistency confidence score is obtained by averaging the pairwise comparisons of the three outputs.

[0040] in The workflow model is represented by the joint set of the first, second, and third sampled outputs, respectively. When the generated output passes the validation function and is determined to be valid... After that, a confidence level assessment is performed. If the assessment result shows that the confidence level is higher than a preset threshold... If the condition is met, the output is considered reliable. If not, the output is considered reliable. and The constraints are detailed business rule validation information and current confidence level. This feedback is integrated into a new Prompt to guide the large model in regenerating its output. This process iterates until the conditions are met or the maximum number of attempts is reached, ultimately generating a complete and accurate workflow model.

[0041] Step 3) Based on the workflow model built in Step 2), output workflow execution logs by driving and monitoring the workflow execution process using task queues; The workflow execution is based on a DAG (Directed Acyclic Graph) structure; therefore, the JSON structure needs to be mapped to a DAG, and the graph structure needs to be constructed using a network interface. The conversion rules are as follows: The `tasks` section comprises all tasks, with node attributes bound to `type` and `lane_id`. `sequence_flows` constructs directed edges, with the edge attribute bound to `name`. Find all nodes with an in-degree of 0 (the starting node), and initialize the execution queue as follows: Build a dynamic queue to be executed The system employs an event-driven mechanism to handle task execution. In each execution loop, the system selects one or more tasks from the queue and submits them for execution.

[0042] Each task has four states: Pending, Running, Success, and Failed. Based on the task state changes, the executableness of subsequent nodes is dynamically determined. For each task... Its predecessor set:

[0043] Task The readiness conditions must meet three conditions: ① ; ②

[0044] ③

[0045] in, For the task Tasks in the precursor set Indicates task The execution status, For the task Input data, This represents the set of all ready data information within the current workflow execution system. For the task Execute the required trigger message, This represents the set of all messages received within the current workflow execution system.

[0046] Condition 1 indicates that task execution requires the successful execution of all predecessor tasks, ensuring the execution order of the process. Condition 2 ensures that all input data required for task execution is successfully transmitted via data stream. For cross-pool collaborative tasks, message streaming is the key method for communication and synchronization; Condition 3 ensures that all trigger messages required for task execution have been successfully received and exist in the received message set. Once the task is ready, it can be added to the execution queue. .

[0047] Task execution supports asynchronous parallelism, and the specific execution function to be called is determined based on the type attribute of each task:

[0048] This formula defines a task. General execution function ,in This represents the input parameters required for task execution. If the task... The attribute type is a service task ( If the task depends on calling an external service interface (API), then the task's completion depends on that API. The attribute type is a script task. The task is completed by executing a predefined script. If the task... The attribute type is user task ( If so, user interaction, approval, or manual operation is required.

[0049] Update node status after call:

[0050] For tasks that fail to execute, a retry, rollback, or manual intervention mechanism should be triggered when the following conditions are met. If the task is successfully executed, it is added back to the queue. After successful execution, the successor nodes of the node are traversed to check if they meet the ready conditions, and subsequent tasks are executed in a loop.

[0051] To support process tracing, performance analysis, and fault tracing, the task execution system needs to record complete logs of the task execution process, and all logs can be constructed into a time-series event stream. .

[0052] Step 4) Analyze the workflow logs using process mining algorithms based on the workflow event logs to identify bottlenecks, redundancies, and anomalies in the process; In a business system, any event 𝑒 is the execution step of its corresponding activity 𝑎∈𝐴 at a certain moment. The execution of each event 𝑒 includes relevant information such as activity name, resources, timestamp, and cost, and is usually defined as a tuple: , Workflow instance ; Execute task name; For timestamps; In execution status; This is the execution function for the task. A trace is a non-empty sequence of events. Where 𝑛 represents the number of events 𝑒 in the trace 𝜎. The event log 𝐿 is a set of traces 𝜎, denoted as .

[0053] First, the initial event log needs to be checked. After preprocessing, including data cleaning, semantic abstraction, and context enhancement, the logs are enhanced. Extract the direct sequence relationships between activities from the logs:

[0054] Indicates activity Adjacent to As previously mentioned, the dependencies between activities can be further defined:

[0055] when A value close to 1 indicates that activity a strongly depends on activity b, suggesting a clear causal relationship; a value close to 0 indicates no clear dependency between them. A dependency matrix can be constructed based on the dependencies between activities. Process models can be built using heuristics, the alpha algorithm, or induction trees. Taking the alpha algorithm as an example, the core idea is based on a set of four-tuple relationships:

[0056] These represent direct causal relationships, mutual exclusion relationships, parallel relationships, and reachability relationships, respectively. Based on activity dependencies, the control flow structure (sequential, parallel, branching, and converging) can be identified and mapped to BPMN elements.

[0057] Consistency checks assess the discrepancy between the process model and the actual system logs, typically involving three quality dimensions: matching, accuracy, and generalization. The core of consistency checks is the alignment principle, which applies to each execution sequence in the event log. The behavioral sequence that the defined model may produce Find the optimal matching path between them. Define a cost function; the movement cost required to align the log trajectory with the model trajectory is:

[0058] It is the log trajectory after movement and trajectory The distance cost required for alignment can be found using the A* search algorithm to find the alignment sequence with the minimum total cost between the log trajectory φ and the model execution path. After sequence matching, the goodness of fit is calculated:

[0059] This is the maximum cost of a complete mismatch in trajectories. Between 0 and 1, the closer to 1, the more consistent they are.

[0060] Performance analysis utilizes timestamps and resource information in the event log to quantitatively evaluate the execution efficiency of the business process model. Bottleneck identification is achieved through queuing theory and flow analysis techniques. By analyzing the execution duration, waiting time, and resource consumption of tasks, the execution efficiency of process activities is quantified. Furthermore, bottleneck links in the critical path can be located through process path frequency analysis.

[0061] Process discovery, consistency checks, and performance analysis are encapsulated into function tools for large model agents to call. By calling the corresponding function tools, process analysis is performed based on the data results output at each step, and resource allocation and optimization strategies are fed back.

[0062] Step 5) Construct a collaborative framework for multiple large-scale intelligent agents to achieve dynamic orchestration, monitoring, and optimization of the workflow. Define the function and collaboration method of each agent. The process design agent outputs a workflow model that meets business requirements based on user text descriptions and the process modeling system; the execution scheduling agent is responsible for the execution and planning of the process and outputs workflow execution logs; the process optimization agent is responsible for analyzing event logs and resource utilization to improve and optimize the workflow, and iteratively outputs the optimal workflow based on human feedback strategies.

[0063] The Process Design Agent (PDA) utilizes the proposed hierarchical modeling system, employing a four-layer cue chain combined with confidence assessment and feedback strategies to decompose, fill in, and generate a four-layer workflow model from unstructured text. It receives optimization suggestions from the Process Design Agent (POA), understands them, and transforms them into structural modifications to the existing workflow model, dynamically adjusting the workflow model based on user needs and feedback.

[0064] The Execution Scheduling Agent (PEA) receives the workflow model generated by the PDA, parses it into a Directed Acyclic Graph (DAG), and executes it efficiently. Based on the dependencies of the DAG, it dynamically schedules and executes each task in the workflow, supporting concurrent execution and error handling mechanisms. It accurately records the execution information of each task (task ID, status, timestamp, input / output, participants, etc.) as data input for the POA. It receives scheduling optimization suggestions from the POA, adjusting task execution priorities, etc., to improve execution efficiency.

[0065] The Process Optimization Agent (POA) continuously monitors the event logs generated by the PEA, using process mining techniques to calculate multi-dimensional performance indicators such as time, cost, and automation rate. It identifies the tasks and paths in the process that are time-consuming, costly, have the highest failure rate, or suffer from uneven resource utilization, uncovering bottlenecks, redundancies, or anomalies in the workflow. The quantitative analysis results from process mining are used as a prompt to the LLM (Workflow Management System), requiring it to perform causal attribution and identify the root causes of problems. Based on the diagnostic results, the LLM generates a series of multi-objective optimization strategies, such as task parallelization, API replacement, and the introduction of manual approval, and estimates their impact on various performance indicators. These strategies, combined with user feedback, serve as the final optimization recommendations. These recommendations are then fed back to the PDA (Process Design Automation System) in natural language to trigger new workflow design iterations.

[0066] This multi-agent framework is a closed-loop iterative process that decomposes complex tasks into LLM agents with different roles, reducing the cognitive burden on individual LLMs. Each agent can be upgraded and optimized independently. For the process design agent, the focus is on process modeling, building a high-quality dataset from natural language to workflow models to ensure the professionalism and accuracy of workflow modeling. Different agents communicate and collaborate through structured information (JSON, logs, optimization suggestions) to achieve workflow design, execution, and optimization.

[0067] Using the process of sales personnel obtaining credit scores from purchasing clients as a case study, a large-scale intelligent agent-driven workflow based on DeepSeek is constructed. This workflow involves three collaborative swimlane pools: the front-end, the bank, and the credit protection agency. Sales personnel initiate credit scoring applications for clients through the front-end, while the bank system communicates with the credit protection agency. The agency performs credit assessments, including primary and secondary scoring, and returns the results to the bank system, which then presents them to the sales personnel. The PDA automatically generates a workflow model containing the three swimlane pools through cue chain reasoning: the front-end includes tasks such as "initiating a request," "querying results," and "process completion"; the bank includes nodes such as "sending a scoring request," "receiving results," "result type validation," "displaying delay information," and "final result output"; and the credit protection agency includes tasks such as "performing primary scoring," "sending scoring results or delay information," and "performing secondary scoring." The automatically generated initial workflow structure achieves a similarity of 0.496 with the standard model, higher than the 0.433 of the manually modeled result. During execution, the system automatically records task execution status, time, data flow, and function call logs. In simulation tests executing multiple scoring requests, PEA found that the secondary scoring task at the credit agency level was the main time-consuming node in the process. The optimization strategy generated by POA based on LLM includes implementing a parallel scoring strategy for the credit agency scoring process, triggering primary scoring and pre-assessment simultaneously, and adjusting the message flow to reduce message latency. Based on POA feedback, PDA automatically restructured the process, updating the control flow logic and message flow of the "delayed scoring processing" sub-process. The optimized workflow similarity improved to 0.594. The final workflow is as follows: Figure 2 As shown, boxes represent tasks, circles represent events, diamonds represent gateways, solid lines represent control flow, and dashed lines represent cross-domain message flow. The flowchart clearly describes the collaborative relationships between multiple participants, enabling dynamic orchestration and optimization of complex cross-domain processes based on large-model intelligent agents.

Claims

1. A method for dynamic orchestration and optimization of business workflows based on large-scale intelligent agents, characterized in that: The steps are as follows: Step 1: Under the Business Process Modeling Notation (BPMN) standard workflow modeling rules, define the code description language for the workflow, determine the attributes and elements in the workflow structure, and define the element content of tasks, events, gateways, sequence flows, message flows, and data flows. Step 2: Based on the workflow description language, design a layered workflow modeling system that includes a business intent layer, a process structure layer, a collaborative behavior layer, and a data configuration layer. Integrate a confidence assessment and correction mechanism to transform complex unstructured natural language requirements into a structured and executable workflow model. Step 3: Construct a task execution queue based on task dependencies to drive and monitor the workflow execution process and output workflow execution logs; Step 4: Based on the workflow event logs, use process mining algorithms to analyze the workflow logs and identify bottlenecks, redundancies, and anomalies in the process; Step 5: Construct a collaborative framework for multiple large-scale intelligent agents to achieve dynamic orchestration, monitoring, and optimization of workflows; define the function and collaboration method of each agent, and design a workflow model that meets business requirements based on user text descriptions and process modeling system. The execution scheduling agent is responsible for the execution and planning of the process and outputs the workflow execution log; the process optimization agent is responsible for analyzing event logs and resource utilization to improve and optimize the workflow, and iteratively outputs the optimal workflow based on user feedback.

2. The method for dynamic orchestration and optimization of business workflow based on a large-scale intelligent agent according to claim 1, characterized in that: In step 1, in order to realize the programmatic modeling and executable description of business processes by the large model intelligent agent, a programmable structured language is used to describe the workflow, and the large model graph generation task is converted into a code generation task. A workflow description language based on JSON structure is designed using Python code. Combining the structure and rules of the BPMN 2.0 standard, it defines the core elements, behavioral rules, and data interaction patterns of business processes. This language is generated by a large model based on natural language parsing and can be directly executed by a compiler or converted into a visual flowchart.

3. The method for dynamic orchestration and optimization of business workflow based on a large-scale intelligent agent according to claim 1, characterized in that: In step 2, the workflow modeling hint chain consists of four layers: Business Intent Layer: Extracts high-level business objectives, core task behaviors, and participating subject roles from users' natural language descriptions to form the initial intent structure for process modeling, providing a semantic foundation for lower-level process structure modeling; Process structure layer: Based on the task sketch output by the intent layer, tasks are decomposed and classified, and a process control structure is established according to the semantic dependencies between tasks; Collaborative Behavior Layer: Based on the existing control structure, identify the collaborative entities in the process and construct a role division model based on pool and lane and cross-role communication behaviors; Data configuration layer: Binds specific data input / output objects, interface parameters and service implementation information to each task, completing the coupling modeling between tasks and data.

4. A method for dynamic orchestration and optimization of business workflow based on a large-scale intelligent agent according to claim 1 or 3, characterized in that: In step 2, the natural language-to-workflow modeling process is referred to as NL-W, and the initial input natural language is defined as NL, with the output represented as: ； in, It is a JSON object; It is a large model in terms of parameters The following generating functions; A preset Prompt containing a four-layer modeling cue chain; For each output, a structure verification and confidence evaluation will be performed. Structure verification checks the syntactic correctness of the output. The verification function is defined as follows: , It performs JSON syntax validation to ensure that the LLM output is a grammatically correct JSON fragment; The business process modeling rules are validated to ensure the generated workflow model structure is correct and complete; the confidence assessment mainly checks the reliability of the output semantics, and combines the model output probability with the multi-path sampling consistency strategy to define the confidence assessment function: , It is the average log probability output by the large model; Multi-path sampling consistency function; It is a weighted average result.

5. The method for dynamic orchestration and optimization of business workflow based on a large-scale intelligent agent according to claim 1, characterized in that: In step 2, the language model generates a conditional probability distribution when producing output: ； in, It is a sequence of tokens generated by the model; This represents the total number of tokens output. It is the first A sequence of tokens; It was before A sequence of tokens.

6. The method for dynamic orchestration and optimization of business workflow based on a large-scale intelligent agent according to claim 1, characterized in that: In step 3, the workflow execution is based on a DAG (Directed Acyclic Graph) structure, requiring the JSON structure to be mapped to a DAG and the graph structure to be constructed using a network interface. The conversion rules are as follows: The `tasks` section comprises all tasks, with node attributes bound to `type` and `lane_id`. `sequence_flows` constructs directed edges, with the edge attribute bound to `name`; Find all nodes with an in-degree of 0, and initially execute the queue as follows: Build a dynamic queue to be executed It uses an event-driven mechanism to handle task execution; In each iteration of the loop, the system selects one or more tasks from the queue and submits them for execution. Each task has four states: Pending, Running, Success, and Failed; based on the changes in task state, the executability of subsequent nodes is dynamically determined; for tasks... Precursor assembly: ； Task The readiness conditions must meet three conditions: Condition 1: ; Condition 2: ; Condition 3: ; in, For the task Tasks in the precursor set Indicates task The execution status, For the task Input data, This represents the set of all ready data information within the current workflow execution system; For the task Execute the required trigger message, This represents the set of all messages received within the current workflow execution system.

7. The method for dynamic orchestration and optimization of business workflow based on a large model intelligent agent according to claim 6, characterized in that: In step 3, task execution supports asynchronous parallelism, and the specific execution function to be called is determined based on the type attribute of each task: ； Define task The general execution function is ,in, These represent the input parameters required for task execution; if the task The attribute type is a service task. The completion of the task depends on the call to the external service interface API; if the task The attribute type is a script task The task is completed by executing a predefined script; if the task The attribute type is user task In this case, user interaction, approval, or manual operation is required. Update node status after call: ； For tasks that fail to execute, a retry, rollback, or manual intervention mechanism should be triggered when the following conditions are met. When the task is successfully executed, the node is re-added to the queue; after the task is successfully executed, the successor nodes of the node are traversed to check whether they meet the ready conditions, and the subsequent tasks are executed in a loop. The task execution system needs to record the entire task execution process in a complete log, and all logs should be constructed into a time-series event stream. .

8. The method for dynamic orchestration and optimization of business workflow based on a large-scale intelligent agent according to claim 1, characterized in that: In step 4, any event 𝑒 in the business system is the execution step of its corresponding activity 𝑎∈𝐴 at a certain moment. ,in, Workflow instance ; Execute task name; For timestamps; In execution status; This is the execution function for the task; a trace is a non-empty sequence of events. Where, 𝑛 is the number of events 𝑒 in the trace 𝜎; the event log 𝐿 is the set of traces 𝜎, denoted as ; Initial event logs are required. After preprocessing, including data cleaning, semantic abstraction, and context enhancement, the logs are enhanced. Extract the direct sequence relationships between activities from the logs: ； Indicates activity Adjacent to Previously, we further defined the dependencies between activities: 。 9. The method for dynamic orchestration and optimization of business workflow based on a large model intelligent agent according to claim 8, characterized in that: In step 4, the consistency check assesses the discrepancy between the process model and the actual system operation logs, typically using three quality dimensions: matching, accuracy, and generalization. This is applied to each execution sequence in the event log. The behavioral sequence that the defined model may produce Find the optimal matching path between them; define the cost function, the movement cost required to align the log trajectory with the model trajectory is: ； It is the log trajectory after movement and the model trajectory The distance cost required for alignment is determined by using the A* search algorithm to find the alignment sequence that minimizes the total cost between the log trajectory 𝜎 and the model execution path. After sequence matching, the goodness of fit is calculated: ； This is the maximum cost of a complete mismatch in trajectories. Between 0 and 1, the closer to 1, the more consistent they are.

10. The method for dynamic orchestration and optimization of business workflow based on a large model intelligent agent according to claim 1, characterized in that: In step 5, the process design agent PDA uses the proposed hierarchical modeling system to decompose, fill in, and generate a four-layer workflow model by combining a four-layer prompt chain with confidence assessment and feedback strategies. It receives optimization suggestions from POA, which are then transformed into structural modifications to the existing workflow model. The workflow model is dynamically adjusted based on user needs and feedback. The execution scheduling agent PEA receives the workflow model generated by the PDA, parses it into a directed acyclic graph (DAG), and executes it efficiently. Based on the dependencies of the DAG, it dynamically schedules and executes each task in the workflow, supporting concurrent execution and error handling mechanisms. It accurately records the execution information of each task as data input for the POA. It also receives scheduling optimization suggestions from the POA, adjusts the priority of task execution, and improves execution efficiency. The process optimization agent POA continuously monitors the event logs generated by PEA and uses process mining technology to calculate multi-dimensional performance indicators, namely time, cost, and automation rate. It identifies the tasks and paths in the process that take the longest time, have the highest cost, the highest failure rate, or have uneven resource utilization, and discovers bottlenecks, redundancies, or anomalies in the workflow. The quantitative analysis results of process mining are used as a prompt input to LLM to identify the root cause of the problem; LLM generates a series of multi-objective optimization strategies and predicts the impact on various performance indicators, and combines user feedback as the final optimization suggestions; The optimization suggestions are fed back to the PDA in natural language to trigger a new workflow design iteration.

Images