A scientific research task automatic execution method and system based on multi-agent collaborative scheduling

Abstract

The application discloses a kind of based on multi-agent collaborative scheduling scientific research task automation execution method and system.The method includes: receiving scientific research target and carrying out semantic analysis by task understanding Agent;According to task type, corresponding skill module is dynamically matched and loaded;Through task design Agent, structured experiment scheme is automatically generated;Through code generation Agent, executable simulation code is automatically generated;Through execution Agent, code is automatically executed and result is collected;Through result analysis Agent, output data is automatically analyzed;Through report generation Agent, structured scientific research report is automatically generated;The whole execution process is carried out under the control of collaborative scheduling kernel, which is responsible for the task distribution of multi-agent, execution state monitoring and result aggregation, and dynamically configures serial or parallel execution according to task dependency relationship.The application also provides corresponding optimization system.The application realizes the full-process automation closed loop from task design to report generation, is especially suitable for scientific research task in vertical field such as building industry, and has significant technical effect.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and scientific research automation technology, specifically relating to a method and system for automating scientific research tasks based on multi-agent collaborative scheduling. Background Technology

[0002] With the rapid development of large language model technology, scientific research automation has become an important research direction for leveraging artificial intelligence in scientific research. Traditional scientific research typically relies on researchers manually designing experimental schemes, writing simulation code, performing calculations, analyzing results, and writing reports—a time-consuming, labor-intensive, and error-prone process. In recent years, the combination of multi-agent systems and large language models has brought new possibilities to scientific research automation.

[0003] Currently, research institutions and enterprises both domestically and internationally have developed numerous scientific research agent systems. For example, the Denario project (see: M. Banon, et al. "Denario: A Multi-Agent System for Scientific Research Automation," arXiv preprint, 2024) has constructed a multi-agent research assistant system capable of performing various tasks such as generating ideas, reviewing literature, developing research plans, writing and executing code, creating diagrams, and drafting and reviewing scientific papers, featuring a modular architecture. NovelSeek (see: Y. Zhang, et al. "NovelSeek: A Closed-Loop Multi-Agent Framework for Autonomous Scientific Discovery," NeurIPS 2024) proposes a unified closed-loop multi-agent framework for autonomous scientific research. SciBORG (see: J. Liu, et al. "SciBORG: Modular Agent Framework for Domain-Specific Scientific Tasks," ACL2024) proposes a modular agent framework that allows agents based on large language models to autonomously plan, reason, and execute domain-specific tasks.

[0004] However, based on the analysis of the aforementioned existing technologies and in light of the actual needs of current scientific research, the existing technologies have the following main shortcomings: (1) The workflow is highly static and lacks adaptability. Most existing systems use pre-programmed fixed workflows, which cannot adaptively adjust the execution process according to changes in intermediate research results or task types. Existing research clearly points out that existing scientific research intelligent agent systems have two fundamental limitations: rigid workflows that cannot adapt to intermediate findings; and insufficient context management that hinders long-term research.

[0005] (2) The skill system is closed and lacks scalability. The toolchain of the existing system is mostly fixed and cannot dynamically load new skill modules according to the needs of new tasks. When the type of scientific research task changes or a new research field emerges, the system needs to be redeveloped or significantly modified, making it difficult to adapt to the diversity and rapid changes of scientific research tasks.

[0006] (3) Limited simulation code generation capabilities. Existing systems are mostly focused on literature research or data analysis tasks, and are weak in automatically generating professional simulation codes for vertical fields (especially numerical simulation and structural analysis in the construction industry). Some systems rely on calling external simulation software interfaces instead of generating executable professional simulation codes based on experimental schemes and executing them automatically, which limits their application value in the field of engineering simulation.

[0007] (4) The entire automated chain is incomplete. Existing systems mostly focus on a single link in the research process (such as literature retrieval or experiment execution), and a complete closed loop from task design to simulation execution, and from result analysis to report generation has not yet been formed. There is still a significant reliance on manual labor in the research results analysis and report generation stages.

[0008] In summary, there is an urgent need to develop a multi-agent collaborative scheduling method and system that can adaptively orchestrate workflows, support dynamic skill expansion, and cover the entire scientific research process. Summary of the Invention

[0009] This invention aims to provide a method and system for automating scientific research tasks based on multi-agent collaborative scheduling, in order to overcome the shortcomings of existing technologies such as static workflow, poor skill scalability, insufficient automatic generation capability of simulation code, and lack of closed-loop automation of the entire scientific research task process.

[0010] To achieve the above objectives, the present invention specifically provides the following technical solution: This invention provides a method for automating the execution of scientific research tasks based on multi-agent cooperative scheduling. This method utilizes a cooperative scheduling kernel to achieve dynamic task distribution, cooperative scheduling, and result aggregation among multiple agents. Combined with a dynamically loadable skill module system, it enables automated execution of the entire scientific research task process. The method includes the following steps: Step 1: Receive the research objectives input by the user in natural language form, and use the task understanding agent to perform semantic analysis and task type identification on the research objectives to identify the task type corresponding to the research objectives; Step 2: Based on the task type identified in Step 1, dynamically match the corresponding skill module from the skill system, and load the skill module at runtime through the skill loading mechanism to realize on-demand expansion of system capabilities; Step 3: Using the task design agent, a structured experimental plan is automatically generated based on the research objectives and the loaded skill modules. The experimental plan includes at least the experimental type, experimental parameter configuration, and experimental step planning. Step 4: Generate Agent through code. Automatically generate executable simulation code based on the structured experimental scheme. The simulation code is adapted to the numerical calculation, data analysis, or professional simulation requirements involved in the experimental scheme. Step 5: By executing the Agent, the simulation code generated in Step 4 is automatically executed, the code execution process is monitored, and the output data generated during the execution process is collected and organized. The output data includes numerical results, log information, and graphic files. Step Six: The result analysis agent automatically analyzes the output data collected in Step Five. The analysis includes data cleaning, statistical analysis, comparative analysis, and trend analysis to generate structured analysis conclusions. Step 7: Generate an Agent from the report. Based on the experimental design in Step 3, the output data in Step 5, and the analysis conclusions in Step 6, automatically generate a structured research report.

[0011] Step 8: During the execution of steps 1 to 7 above, the collaborative scheduling kernel is responsible for task distribution, execution status monitoring and result aggregation among agents. It configures the serial or parallel execution mode of each agent according to the task dependency relationship, and aggregates the output results of each agent after the entire process is completed to generate the final scientific research task output.

[0012] Furthermore, in this method, the skill system in step two includes multiple independently loadable skill modules. The skill modules include one or more of the following: literature retrieval skills, paper analysis skills, experimental design skills, simulation code generation skills, code execution skills, result analysis skills, and report generation skills. The dynamic matching and loading of skill modules supports the dynamic loading of new skill modules during system runtime without modifying the core code.

[0013] Furthermore, in this method, the simulation code in step four includes code written in Python, and the code calls at least one of a numerical computation library, a data analysis library, and a graphics rendering library; the code generation agent supports batch parameter adjustment and generates multiple sets of simulation code under different experimental parameters at once.

[0014] Furthermore, in this method, the output data in step five includes one or more of the following: numerical result data, log information output from the console, and graphical files.

[0015] Furthermore, in this method, the automated analysis in step six includes one or more of data cleaning, statistical analysis, comparative analysis, and trend analysis; the result analysis agent is also used to automatically generate data visualization charts.

[0016] Furthermore, in this method, the structured research report in step seven includes sections on research background, methods, results, and conclusions; the report generation agent automatically embeds the visualization charts generated during the analysis process into the corresponding positions of the report.

[0017] Furthermore, in this method, the collaborative scheduling kernel configures the serial execution mode or parallel execution mode of each Agent according to the task dependency relationship: for tasks with dependencies, they are executed serially according to the dependency relationship; for independent tasks without dependencies, multiple Agents are assigned to execute in parallel.

[0018] Based on the above method, the present invention also provides an automated scientific research task execution system based on multi-agent cooperative scheduling, the system comprising: The user interaction layer is used to receive research goals input by users and output execution results to users; The collaborative scheduling kernel, connected to the user interaction layer, is used for task distribution, collaborative scheduling, and result aggregation of multiple agents. It supports parallel and serial execution of multiple tasks and automatically allocates tasks based on task dependencies and agent capabilities. The skill system includes multiple independently loadable skill modules, which support dynamic loading and unloading at runtime and expand system capabilities as needed. Each skill module corresponds to an independent scientific research function. The skill system supports dynamic loading and unloading at runtime, and users can expand system capabilities as needed according to task requirements. The multi-agent layer includes multiple specialized agents, which at least include: a task understanding agent for parsing research objectives and identifying task types; a task design agent for generating structured experimental plans; a code generation agent for automatically generating simulation code; an execution agent for automatically executing code and collecting results; a result analysis agent for automatically analyzing experimental results; a report generation agent for automatically generating research reports; and an aggregation agent for summarizing the output results of each agent. All agents work collaboratively under the scheduling of a collaborative scheduling kernel. The research task workflow orchestration engine, connected to the collaborative scheduling kernel, is used for research task workflow definition, task scheduling, and execution control, and supports customized research workflows.

[0019] Furthermore, in this system scheme, the collaborative scheduling kernel includes a task decomposition unit, a dependency analysis unit, a task distribution unit, and a result aggregation unit; the task decomposition unit is used to break down the research objective into multiple sub-tasks; the dependency analysis unit is used to identify the sequential dependencies between sub-tasks; the task distribution unit is used to allocate sub-tasks according to the sub-task type and the capability characteristics of each Agent; and the result aggregation unit is used to collect and integrate the output results of each Agent after the completion of each stage of the task.

[0020] Furthermore, in this system solution, each skill module in the skill system is an independent plug-in unit that interacts with the corresponding Agent through a standardized interface, supporting the dynamic loading of new skill modules and the dynamic unloading of no longer needed skill modules during system operation.

[0021] Compared with the prior art, the solution provided by the present invention has the following beneficial effects: (1) This invention realizes dynamic task distribution and collaborative scheduling of multiple agents through a collaborative scheduling kernel (such as the OpenClaw scheduling kernel), which can adaptively adjust the execution process according to the task type and intermediate execution results, thus overcoming the defects of static workflow in the prior art.

[0022] (2) The present invention adopts a pluggable skill system design, which supports dynamic loading and unloading of skill modules at runtime. Users can expand the system capabilities as needed according to new task requirements without modifying the core code, thus solving the problem of fixed toolchain and poor scalability in existing technologies.

[0023] (3) This invention generates a professional simulation code (such as Python numerical calculation code) that is adapted to the experimental scheme by generating an agent through code generation. The agent automatically runs the code and collects the results, realizing an automated closed loop from experimental scheme to simulation results. It is especially suitable for scientific research tasks in professional engineering fields such as building structure analysis and fluid simulation.

[0024] (4) This invention realizes the automatic completion of the entire chain from task design to simulation code generation, code execution, result organization, result analysis and report generation, without human intervention, which significantly improves scientific research efficiency.

[0025] (5) This invention has optimized the simulation code generation capability for vertical fields such as the construction industry, and supports professional scenarios such as structural analysis, fluid simulation, and energy consumption analysis. It can realize a complete scientific research closed loop from scientific research task understanding and design, simulation code generation and execution, experimental result organization and analysis to report generation. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0027] Figure 1 This is a system architecture diagram of the present invention.

[0028] Figure 2 This is a flowchart of the multi-agent collaborative scheduling process in this invention.

[0029] Figure 3 This is a schematic diagram illustrating the dynamic loading of skills in this invention.

[0030] Figure 4 This is a schematic diagram illustrating the execution of the research task workflow orchestration engine of this invention. Detailed Implementation

[0031] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0032] This invention addresses the problems in existing scientific research automation technologies, such as static workflows, poor skill scalability, insufficient automatic simulation code generation capabilities, and the lack of closed-loop automation for the entire scientific research task process. It proposes an automated execution scheme for scientific research tasks based on multi-agent collaborative scheduling. This scheme achieves dynamic task distribution, collaborative scheduling, and result aggregation among multiple agents through a collaborative scheduling kernel, and, combined with a dynamically loadable skill module system, realizes automated execution of the entire scientific research task process. Specifically, this invention establishes a multi-agent collaborative scheduling architecture based on the OpenClaw scheduling kernel. The OpenClaw kernel serves as the core scheduling engine for the entire system, responsible for task distribution, collaborative scheduling, and result aggregation among the multiple agents. Unlike existing systems where agents execute sequentially and lack unified scheduling, the scheduling kernel in this invention intelligently allocates tasks based on task dependencies and the capabilities of each agent. It supports parallel execution of independent tasks by multiple agents and sequential execution of dependent tasks. This architecture, through the organic combination of centralized scheduling and distributed execution, ensures both unified control of the core logic and efficient collaboration among multiple agents.

[0033] Based on this, the present invention also incorporates the following innovative designs.

[0034] This invention constructs a dynamically expandable skill system. Employing a pluggable skill system design, it abstracts the professional capabilities involved in research tasks (such as literature retrieval, paper analysis, simulation code generation, and result visualization) into independent skill modules. Each skill module can be independently installed, uninstalled, and upgraded. During system runtime, the system dynamically matches and loads the corresponding skill modules based on the task type, achieving on-demand expansion of system capabilities. Unlike existing technologies where toolchains are fixed and difficult to adapt to new task requirements, this invention's skill system can continuously expand and optimize with the development of the research field and changes in user needs, adapting to diverse research scenarios without modifying the core code.

[0035] This invention establishes a closed loop for intelligent generation and automated execution of simulation code. Through the collaborative work of a code generation agent and an execution agent, it achieves a fully automated closed loop from experimental design to simulation results. The code generation agent automatically generates executable code (such as Python code, calling scientific computing libraries like NumPy, SciPy, Pandas, and Matplotlib) adapted to the needs of numerical computation, data analysis, or specialized simulations, based on the structured experimental design output by the task design agent. The execution agent automatically runs the generated code, monitors the execution process, handles anomalies, and collects and organizes the output data. This closed-loop mechanism is particularly suitable for research tasks in specialized engineering fields such as building structure analysis, fluid simulation, and energy consumption analysis. It can automatically generate and run multiple sets of simulation code based on different experimental parameters, forming a systematic simulation result dataset, significantly reducing the workload of researchers in writing and debugging simulation code.

[0036] This invention establishes a closed-loop automation for the entire scientific research task process. It achieves this by organically combining task understanding agents, task design agents, code generation agents, execution agents, result analysis agents, report generation agents, and aggregation agents, forming a complete automated chain from research objective input to final report output. Simultaneously, all agents work collaboratively under the unified scheduling of the OpenClaw scheduling kernel. The task understanding agent parses the research goals input by the user and identifies the task type; The task design agent automatically generates a structured experimental plan based on the identification results; The code generation agent transforms experimental protocols into executable code. The Agent executes the code and collects the results; The results analysis agent cleans, statistically analyzes, compares, and performs trend analysis on the raw data, generating analytical conclusions and visualization charts. The report generation agent automatically generates structured research reports based on experimental protocols, analytical conclusions, and charts. The aggregation agent summarizes the outputs from each stage to form the final research task results.

[0037] Based on the above-mentioned technologies, this invention specifically provides a method and system for automating the execution of scientific research tasks based on multi-agent cooperative scheduling. The specific implementation scheme of this invention will be described in detail below with reference to the accompanying drawings.

[0038] Figure 1 The system architecture diagram of the automated scientific research task execution system based on multi-agent collaborative scheduling in this invention is shown, clearly presenting the hierarchical relationship between the user interaction layer, the OpenClaw scheduling kernel, the skill system, the multi-agent layer, and the scientific research task process orchestration engine, as well as the data interaction relationship between each layer. Figure 2 The process of multi-agent collaborative scheduling, from receiving research objectives to outputting results, is presented in full in the form of a flowchart. Figure 3 The illustration shows the dynamic skill loading mechanism, including the complete process of skill registration, dynamic loading, and execution. Figure 4 The flowchart illustrates the connection between the four core stages of the scientific research task orchestration engine: task design, simulation execution, result analysis, and report generation.

[0039] Reference Figure 2 and Figure 4 The execution flow of the automated scientific research task execution method based on multi-agent cooperative scheduling provided by the present invention is as follows.

[0040] Step 1: Receiving and understanding research objectives.

[0041] This step is configured to first receive the research objective input by the user in natural language form through the user interaction layer, such as "analyze the impact of different shear wall arrangement schemes on the seismic performance of a certain building structure"; then, the task understanding agent performs semantic parsing and task type identification on the input research objective, extracting keywords and identifying the task type corresponding to the research objective. The task type mentioned here includes one or more of numerical simulation tasks, literature research tasks, and scheme generation tasks; finally, based on the identified task type, the task understanding agent also preliminarily determines the skill modules to be invoked and the subsequent sequence of steps to be executed to complete the task.

[0042] This step lays the foundation for subsequent task scheduling and skill matching by transforming research objectives described in natural language into structured task definitions.

[0043] Step 2: Dynamic matching and loading of skill modules.

[0044] This step configures the OpenClaw scheduling kernel to dynamically match the corresponding skill modules from the skill system based on the task type identified in step one.

[0045] See Figure 3 The skill system involved in this step includes multiple independently loadable skill modules, each corresponding to an independent scientific research function.

[0046] Specifically, the skills modules include, but are not limited to: literature retrieval skills, paper analysis skills, experimental design skills, simulation code generation skills, code execution skills, result analysis skills, and report generation skills.

[0047] Furthermore, in this step, for the skill modules in the skill system, a pluggable skill design is preferred, that is, each skill module is an independent functional unit that can be installed, uninstalled and upgraded separately without affecting other functions of the system.

[0048] This system supports dynamic loading at runtime, meaning that new skills can be dynamically loaded based on task requirements during system operation without restarting the system or modifying the core code. Users can select and load the necessary skill modules from the skill library according to the specific needs of their research tasks, enabling on-demand expansion of the system's capabilities.

[0049] Step 3: Automated design of experimental schemes.

[0050] This step is configured to be executed by the task design agent. Based on the research objectives identified in step one and the loaded skill modules, the task design agent automatically generates a structured experimental plan. This experimental plan includes at least the experiment type (e.g., numerical simulation, physical experiment design, or comparative experiment), experimental parameter configuration, and experimental procedure planning.

[0051] Specifically, the task design agent invokes the experiment design skill module and combines it with a domain knowledge base (such as a construction industry knowledge base) to optimize experimental parameters and schemes, ensuring the rationality and feasibility of the schemes. In engineering scenarios such as seismic optimization of building structures, the task design agent automatically designs various structural layout schemes (such as traditional uniform shear wall layout, centralized core tube layout, decentralized layout, and hybrid layout), determines the specific parameters of each scheme (such as floor height, wall thickness, material strength grade, etc.), and plans a complete sequence of experimental steps.

[0052] This step automates the traditionally manual design of experimental schemes, significantly reducing the workload of researchers in designing schemes.

[0053] Step 4: Automated generation of simulation code.

[0054] This step is configured to be executed by the code generation agent. Based on the structured experimental design generated in step three, the code generation agent automatically generates executable simulation code. The generated simulation code is adapted to the numerical computation, data analysis, or specialized simulation requirements involved in the experimental design.

[0055] In its implementation, the code generation agent internally includes a template matching unit and a code synthesis unit. The template matching unit is configured to select a corresponding code template from a pre-set code template library based on the type of experimental scheme (such as structural seismic analysis, fluid dynamics simulation, etc.). Each template contains a complete code framework, function signature, necessary import statements, and placeholders for parameters to be filled. The code synthesis unit is configured to fill the template placeholders with parameters from the experimental scheme (such as structural dimensions, material properties, load conditions, etc.) to form preliminary code.

[0056] For customized logic that the template cannot cover, a large language model hint containing an experimental scheme description and code context is constructed. The code synthesis unit calls the large language model to generate supplementary code segments and integrates them into the template.

[0057] The code generation agent performs syntax checks and dependency availability verification on the generated complete code to ensure it is executable.

[0058] Furthermore, the simulation code automatically generated by the code generation agent calls at least one of the following libraries: numerical computation library (such as NumPy, SciPy), data analysis library (such as Pandas), and graphics library (such as Matplotlib).

[0059] For specialized engineering scenarios such as building structure analysis, the code generation agent automatically generates simulation code containing the following functional modules: a finite element modeling module (using NumPy and SciPy to build structural models), a modal analysis module (calculating the natural frequencies and mode shapes of the structure), a seismic response analysis module (calculating the seismic response using time history analysis methods), and a results output module (organizing and outputting the analysis results). The code generation agent supports batch parameter adjustment and can generate multiple sets of simulation code under different experimental parameters at once. The technical advantage of this step is that it provides directly executable simulation programs without requiring manual coding, significantly reducing the technical barrier for researchers to write and debug simulation code.

[0060] Step 5: Automated code execution and result collection.

[0061] This step is configured to be executed by the execution agent. The execution agent automatically runs the simulation code generated in step four, monitors the code execution process, handles abnormal situations (including timeout interruptions, memory monitoring, error logging, etc.), and collects and organizes the output data generated during code execution. The output data includes numerical result data (such as displacement, stress, natural frequency, etc.), log information output from the console, and graphical files (such as mode shapes, displacement-time history curves, etc.).

[0062] In the scenario of batch parameter adjustment, the agent automatically runs the simulation code of multiple sets of experimental parameters in sequence, performs a complete structural analysis of each set of experiments, and organizes the output data of each set of experiments into a standardized result dataset.

[0063] Based on the steps, unmanned execution from code generation to result output can be achieved, solving the pain points of time-consuming and error-prone manual execution of simulation code.

[0064] Step Six: Automated Analysis of Experimental Results.

[0065] This step is configured to be performed by the Results Analysis Agent. The Results Analysis Agent performs automated analysis on the output data collected in step five, including: Data cleaning involves preprocessing raw data, including handling missing values, detecting outliers, and converting data formats. Statistical analysis, including descriptive statistics, analysis of variance, and correlation analysis; Comparative analysis involves comparing the results of multiple experimental groups to analyze the differences in the effects of different parameter configurations. Trend analysis involves analyzing the trends in time series or multiple sets of experimental data to reveal the relationships between variables. The results analysis agent automatically generates structured analysis conclusions and produces data visualization charts (including line charts, bar charts, heatmaps, etc.) to intuitively display the analysis results.

[0066] This step automatically transforms the raw simulation output data into readable analytical conclusions, eliminating the bottleneck of manual data processing and analysis.

[0067] Step 7: Automated generation of research reports.

[0068] This step is configured to be executed by the report generation agent. Based on the experimental design from step three, the output data from step five, and the analysis conclusions from step six, the report generation agent automatically generates a structured research report.

[0069] The research reports here automatically generate chapters such as research background, methods, results, and conclusions according to a standard format, and automatically embed the visualization charts generated during the analysis process into the corresponding positions in the report.

[0070] The report generation agent supports outputting both PDF and Word formats to meet the needs of different scenarios.

[0071] Step 8: Multi-agent collaborative scheduling and result aggregation.

[0072] During the execution of steps one through seven above, the OpenClaw scheduling kernel, as the core scheduling engine of the entire system, is responsible for task distribution, execution status monitoring, and result aggregation among the agents.

[0073] The scheduling kernel configures the serial or parallel execution mode of each agent based on task dependencies: For tasks without dependencies (such as generating simulation code for multiple independent experiments), the scheduling kernel allocates multiple agents to execute in parallel to improve efficiency. For tasks with dependencies (such as code execution that must be performed after code generation), the scheduling kernel executes them sequentially according to the dependencies.

[0074] After each stage of the task is completed, the scheduling kernel aggregates the output of that stage and passes it as input to the Agent for the next stage. After the entire process is completed, the scheduling kernel calls the aggregation Agent to summarize the output of each Agent and generate the final scientific research task output.

[0075] This step achieves efficient collaboration among multiple agents through a unified scheduling mechanism, ensuring both the orderly execution of tasks and maximizing the efficiency of parallel processing.

[0076] In addition to the existing method for automating scientific research tasks based on multi-agent cooperative scheduling, this invention further provides a system for automating scientific research tasks based on multi-agent cooperative scheduling, for automatically executing the above-mentioned method.

[0077] like Figure 1 As shown, the automated execution of scientific research tasks based on multi-agent collaborative scheduling provided by this invention consists of five core modules: user interaction layer, OpenClaw scheduling kernel, multi-agent layer, skill system, and scientific research task process orchestration engine, forming a complete automated processing system solution from user input to task execution.

[0078] The user interaction layer is located at the top layer of the system. It receives the research objectives input by the user through a graphical user interface or command line interface, and displays the status information during the execution process and the final execution result to the user.

[0079] The OpenClaw scheduling kernel, as the core scheduling engine of the entire system, connects the user interaction layer with the lower-level modules, and is responsible for task distribution, collaborative scheduling, and result aggregation among multiple agents. The scheduling kernel incorporates a task decomposition unit and a dependency analysis unit. The task decomposition unit is configured to break down complex research tasks into multiple executable subtasks, while the dependency analysis unit is configured to identify the dependencies between subtasks, thereby determining the execution order of the subtasks, such as sequential or parallel execution.

[0080] The multi-agent layer consists of multiple specialized agents, including task understanding agents, retrieval agents, task design agents, code generation agents, execution agents, result analysis agents, report generation agents, and aggregation agents. These agents collaborate under the unified scheduling of the OpenClaw scheduling kernel, with each agent responsible for one or more specialized stages in the research task process.

[0081] The skill system is horizontally integrated with the multi-agent layer, providing each agent with callable professional capability modules. The skill system includes literature retrieval skills, paper analysis skills, experimental design skills, simulation code generation skills, code execution skills, result analysis skills, and report generation skills. Each skill module is an independent functional unit that can be installed, uninstalled, and upgraded independently, and supports dynamic loading at runtime.

[0082] The research task workflow orchestration engine is located in the underlying support layer of the system. It is responsible for defining the workflow of research tasks, scheduling tasks, and controlling execution, and supports users to customize workflows according to specific research needs.

[0083] As further explanation, the five layers in this system communicate with each other through standardized data interfaces. The user interaction layer transmits the input research goals to the OpenClaw scheduling kernel; the scheduling kernel decomposes the tasks and distributes them to the corresponding agents in the multi-agent layer; each agent calls the corresponding skill module from the skill system when executing the task; the process orchestration engine provides the scheduling kernel with the ability to define and control the execution process.

[0084] The following details the specific implementation logic of each module in this system.

[0085] The user interaction layer in this system receives user input through a graphical user interface or command-line interface.

[0086] The input parameters include a natural language description of the research objectives, a task type indication (optional), and reference information required during execution (such as an existing list of references, initial parameter settings, etc.).

[0087] This layer is also responsible for displaying the intermediate states during execution, the output results of each agent, and the final execution report to the user. Users can manually intervene during execution through the interactive layer, such as adding information, adjusting parameters, or modifying the execution path.

[0088] The OpenClaw scheduling kernel, as the core scheduling engine of the system, contains a task decomposition unit, a dependency analysis unit, a task distribution unit, and a result aggregation unit.

[0089] Specifically, the task decomposition unit is configured such that after the scheduling kernel receives the research objectives input by the user, it first calls the task understanding agent to perform semantic parsing of the research objectives and decompose them into multiple sub-tasks with clear input and output boundaries.

[0090] The dependency analysis unit is configured to interact with the task decomposition unit to identify the dependencies between subtasks, construct an execution flow chart in the form of a directed acyclic graph (DAG), and transmit it to the task distribution unit.

[0091] The task distribution unit is configured to interact with the dependency analysis unit, and can assign each subtask to the most suitable agent according to the subtask type and the capabilities of each agent, and determine the execution order of each task according to the dependency relationship: for independent subtasks without dependency relationship, multiple agents are assigned to execute in parallel to improve efficiency; for subtasks with dependency relationship, they are executed serially according to the dependency relationship.

[0092] Specifically, the task distribution unit assigns subtasks with an in-degree (number of dependencies) of zero to idle agents for parallel execution based on the DAG, and executes dependent subtasks sequentially.

[0093] The result aggregation unit collects the output results of each Agent after the completion of each stage task, unifies the format and integrates the content to form standardized intermediate data for transmission to the next stage.

[0094] This OpenClaw scheduling kernel enables adaptive collaborative execution for multiple tasks, significantly improving the efficiency and flexibility of multi-task processing.

[0095] The skill system in this system is implemented using a pluggable design, with each skill module being an independent plug-in unit.

[0096] Specifically, each skill module in the skill system interacts with the Agent through a standardized application programming interface (API). Each skill module defines its own input parameter format, output data format, and invocation method. At startup, the system scans the skill registration directory, loads the installed skill modules, and builds a skill index. During runtime, when an Agent needs to invoke a specific skill, the scheduling kernel searches for the corresponding skill module in the skill index and loads it at runtime through a dynamic loading mechanism.

[0097] This allows the system to load newly installed skill modules during runtime without requiring a system restart. When a skill module is no longer needed, it can be dynamically uninstalled to free up system resources.

[0098] The resulting skill system enables plug-and-play functionality for skill modules, allowing the system to continuously expand its capabilities as the scientific research field develops without requiring modifications to the core code.

[0099] The multi-agent layer in this system is the main body responsible for carrying out scientific research tasks. The specific responsibilities and collaborative relationships of each agent in this multi-agent layer are configured as follows, combined with... Figure 4 As shown: The task understanding agent is configured to invoke a semantic parsing model (such as a large language model) to parse the research objectives input by the user, extract keywords, identify the task type (numerical simulation, literature research or scheme generation), identify the professional terms in the task domain, and output a standardized task definition object.

[0100] The retrieval agent is configured to understand the task definition output by the agent based on the task, invoke the literature retrieval skill module, and retrieve relevant academic papers and technical materials through academic search engines or paper databases (such as CNKI, Wanfang, arXiv, etc.). The retrieval agent supports multi-keyword combination retrieval, deduplication of search results, and relevance ranking.

[0101] The task design agent is configured to generate a structured experimental plan based on the technical information provided by the retrieval agent and the task definition output by the task understanding agent. The task design agent then invokes the experiment design skill module, combining it with professional rules from the domain knowledge base to optimize the experimental parameters and procedures.

[0102] The code generation agent is configured to automatically generate executable simulation code based on the experimental scheme output by the task design agent. The agent invokes the simulation code generation skill module to automatically generate suitable simulation code according to the experimental type and parameter configuration in the experimental scheme. For numerical simulation tasks, the agent generates simulation code written in Python, which includes modules for finite element modeling, modal analysis, seismic response analysis, and results output.

[0103] As further explanation, the simulation code generation skill module here includes a template matching unit and a code synthesis unit. The template matching unit is configured to select a corresponding code template from a pre-set code template library based on the type of experimental scheme (e.g., structural seismic analysis, fluid dynamics simulation, etc.). Each template contains a complete code framework, function signature, necessary import statements, and placeholders for parameters to be filled. The code synthesis unit is configured to fill the template placeholders with parameters from the experimental scheme (e.g., structural dimensions, material properties, load conditions, etc.) to form preliminary code.

[0104] For customized logic that the template cannot cover, a large language model hint containing an experimental scheme description and code context is constructed. The code synthesis unit calls the large language model to generate supplementary code segments and integrates them into the template.

[0105] The code generation agent performs syntax checks and dependency availability verification on the generated complete code to ensure it is executable.

[0106] Configure the execution agent to automatically run the simulation code generated by the code generation agent. The execution agent calls the code execution skill module to create an independent code execution environment (such as a virtual environment or container), executes the code in a controlled environment, and monitors resource consumption and execution status in real time. The execution agent automatically handles and records abnormal situations (such as timeouts, memory overflows, syntax errors, etc.) during code execution.

[0107] The Results Analysis Agent is configured to perform automated analysis on the output data collected by the Execution Agent. The Results Analysis Agent invokes the Results Analysis Skills Module to clean, statistically analyze, compare, and perform trend analysis on the data, generating structured analytical conclusions, and invokes the Data Visualization Skills Module to generate visual charts.

[0108] The report generation agent is configured to automatically generate structured research reports based on experimental protocols, output data, and analytical conclusions. The agent invokes the report generation skill module to generate a complete report in a standard academic report format, including sections on research background, methods, results, and conclusions, and automatically embeds visual charts and graphs in their corresponding positions within the report.

[0109] The aggregation agent is configured to summarize the output results of each agent after the entire scientific research task is completed, perform final quality checks, format standardization, and content integration, and generate the final scientific research task output.

[0110] Furthermore, the agents described here are configured to work collaboratively under the unified scheduling of the OpenClaw scheduling kernel, which maintains the single responsibility of each agent while achieving efficient collaboration through the scheduling kernel.

[0111] The research task orchestration engine in this system is responsible for defining research task workflows, scheduling tasks, and controlling execution. This engine supports two working modes: Predefined mode: Users predefine workflows according to their research needs, and the workflow orchestration engine executes tasks at each stage according to the predefined workflow. In adaptive mode, the workflow orchestration engine dynamically adjusts the subsequent execution flow based on the intermediate results during task execution.

[0112] This enables flexible configuration and adaptive arrangement of scientific research task processes.

[0113] The following section uses a specific example of a seismic optimization project for a building structure to illustrate in detail the deployment method of the system, the detailed execution process of the method, and the technical characteristics and effects of the entire solution in practical applications.

[0114] (I) System Deployment and Instance Parameter Settings The research task of this example is "to analyze the impact of different shear wall arrangement schemes on the seismic performance of a building structure". The system is deployed on an engineering workstation equipped with a multi-core processor and more than 32GB of memory. The software uses Python 3.11 combined with relevant numerical calculation libraries and machine learning frameworks to implement each agent module. The OpenClaw scheduling kernel uses distributed task queue technology to realize task distribution and state management. The skill system adopts a plug-in architecture to realize the dynamic loading of skill modules.

[0115] The example design parameters are as follows: The basic information of the building structure is a 30-story frame-shear wall structure with a seismic intensity of 8 degrees and a basic design seismic acceleration value of 0.20g; the experimental scheme includes four shear wall arrangement schemes: Scheme A is a traditional uniform arrangement, Scheme B is a core tube concentrated arrangement, Scheme C is a dispersed arrangement, and Scheme D is a mixed arrangement.

[0116] (II) System Execution Flow (in conjunction with) Figure 4 (As shown) Phase 1: Task understanding.

[0117] Users input their research objectives through the user interaction layer. The task understanding agent performs semantic parsing on the input research objectives, identifies the task type as "numerical simulation task," and extracts the keywords "shear wall layout scheme," "seismic performance," and "structural analysis." The task understanding agent outputs a standardized task definition object, including task type identifiers, domain classifications, and technical terminology tags.

[0118] Phase 2: Literature search and knowledge acquisition.

[0119] The OpenClaw scheduling kernel assigns the literature retrieval subtask to the retrieval agent. The retrieval agent invokes the literature retrieval skill module, using "structural analysis of seismic performance of shear wall layout" as the keyword to search for relevant academic papers and technical materials, obtaining approximately 20 relevant papers covering the latest domestic and international research results on seismic-resistant structures. The search results are then passed to the task design agent.

[0120] Phase 3: Experimental Design.

[0121] Based on the retrieved technical data and task definition, the Task Design Agent invokes the Experiment Design skill module to automatically design four shear wall layout schemes, determining parameters such as floor height (30 stories), shear wall thickness (200mm to 400mm), and concrete strength grade (C40) for each scheme. The Task Design Agent outputs a structured experimental scheme document.

[0122] Phase 4: Simulation code generation.

[0123] The code generation agent automatically generates four Python simulation code snippets based on the experimental design. Each code snippet includes a finite element modeling module (using NumPy and SciPy to build the structural model), a modal analysis module (calculating the natural frequencies and mode shapes of the structure), a seismic response analysis module (calculating the seismic response using time history analysis methods), and a results output module (organizing and outputting the analysis results). The code generation agent generates four different code versions with different parameter configurations for the four design schemes.

[0124] Phase 5: Code execution and result collection.

[0125] The agent automatically runs four simulation code segments to perform a complete structural analysis for each scheme. The agent monitors the execution process, automatically collecting output data, including numerical results such as displacement, stress, and natural frequencies, calculation process logs from the console, and graphical files such as mode shapes and displacement-time history curves. The agent then organizes the collected data into a standardized dataset.

[0126] Phase 6: Results Analysis and Visualization.

[0127] The results analysis agent compares and analyzes the results data of the four schemes: calculating the seismic performance indicators of each scheme, including natural frequency, maximum displacement, inter-story drift angle, and peak seismic response; performing analysis of variance to assess the significance of differences between the schemes; and automatically generating comparative charts, including a bar chart comparing natural frequencies, a comparison chart of maximum displacement, a comparison chart of inter-story drift angle, and a comparison curve of seismic response time history. The results analysis agent outputs structured analysis conclusions, identifying the scheme with the best seismic performance.

[0128] Phase 7: Report generation.

[0129] The report generation agent automatically generates a complete research report. The report structure includes: Chapter 1 Research Background and Significance, Chapter 2 Overview of Existing Technologies, Chapter 3 Experimental Design, Chapter 4 Simulation Results and Analysis, and Chapter 5 Conclusions and Recommendations. The report generation agent automatically embeds the comparison charts generated in Phase 6 into the corresponding positions in Chapter 4 of the report and outputs PDF and Word files.

[0130] Phase 8: Results Aggregation and Output. The aggregation agent summarizes the output results from each phase, including experimental protocol documents, simulation code files, raw result data, analysis conclusions, charts and graphs, and a complete research report, forming the final research task output, which is presented to the user through the user interaction layer. The entire execution process, from inputting research objectives to outputting a complete research report, is fully automated, requiring no manual intervention from experimental design to report generation.

[0131] As shown above, the OpenClaw scheduling kernel in this example solution achieves parallel scheduling of four sets of simulation code generation tasks and controls the serial dependency relationship between code generation and code execution through DAG dependency analysis, significantly improving the efficiency of multi-task processing. The dynamic loading mechanism of the skill system allows the system to load experimental design skills, simulation code generation skills, result analysis skills, and report generation skills on demand, without pre-installing all functional modules. The collaborative cooperation between the code generation agent and the execution agent achieves an automated closed loop from experimental design to simulation results, solving the pain point of time-consuming and labor-intensive manual simulation code writing. The end-to-end closed-loop architecture enables uninterrupted execution from literature retrieval to report generation, achieving end-to-end automation of scientific research tasks.

[0132] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0133] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for automating the execution of scientific research tasks based on multi-agent cooperative scheduling, characterized in that, The method executes the following steps under the control of the cooperative scheduling kernel: Step 1: Receive the research objectives input by the user in natural language form, and use the task understanding agent to perform semantic analysis and task type identification on the research objectives; Step 2: Based on the task type identified in Step 1, dynamically match and load the corresponding skill module from the skill system; Step 3: Using the task design agent, a structured experimental plan is automatically generated based on the research objectives and the loaded skill modules. The experimental plan includes at least the experimental type, experimental parameter configuration, and experimental step planning. Step 4: Generate Agent through code, automatically generating executable simulation code based on the structured experimental scheme; Step 5: By executing the Agent, the simulation code generated in Step 4 is executed automatically, the execution process is monitored, and the output data generated during the execution process is collected and organized. Step Six: The Result Analysis Agent automatically analyzes the output data collected in Step Five to generate structured analysis conclusions. Step 7: Generate an Agent from the report. Based on the experimental design in Step 3, the output data in Step 5, and the analysis conclusions in Step 6, automatically generate a structured research report. Step 8: The collaborative scheduling kernel is responsible for task distribution, execution status monitoring and result aggregation among agents, and after the entire process is completed, it aggregates the output results of each agent to generate the final scientific research task output.

2. The method according to claim 1, characterized in that, The skill system in step two includes multiple independently loadable skill modules, which include one or more of the following: literature retrieval skills, paper analysis skills, experimental design skills, simulation code generation skills, code execution skills, result analysis skills, and report generation skills. The dynamic matching and loading of skill modules supports the dynamic loading of new skill modules during system runtime without modifying the core code.

3. The method according to claim 1, characterized in that, The simulation code in step four includes code written in Python, and the code calls at least one of the numerical computation library, data analysis library, and graphics drawing library; the code generation agent supports batch parameter adjustment and generates multiple sets of simulation code under different experimental parameters at one time.

4. The method according to claim 1, characterized in that, The output data in step five includes one or more of the following: numerical result data, log information output from the console, and graphical files.

5. The method according to claim 1, characterized in that, The automated analysis in step six includes one or more of data cleaning, statistical analysis, comparative analysis, and trend analysis; the result analysis agent is also used to automatically generate data visualization charts.

6. The method according to claim 1, characterized in that, The structured research report mentioned in step seven includes sections on research background, methods, results, and conclusions; the report generation agent automatically embeds the visualization charts generated during the analysis process into the corresponding positions in the report.

7. The method according to claim 1, characterized in that, The collaborative scheduling kernel configures the serial or parallel execution mode of each Agent according to the task dependency relationship: for tasks with dependencies, they are executed serially according to the dependency relationship; for independent tasks without dependencies, multiple Agents are assigned to execute in parallel.

8. An automated scientific research task execution system based on multi-agent cooperative scheduling, characterized in that, The system includes: The user interaction layer is used to receive research goals input by users and output execution results to users; The collaborative scheduling kernel, connected to the user interaction layer, is used for task distribution, collaborative scheduling, and result aggregation among multiple agents. The skill system includes multiple independently loadable skill modules, which support dynamic loading and unloading at runtime, and expand system capabilities as needed; The multi-agent layer includes multiple specialized agents, including a task understanding agent, a task design agent, a code generation agent, an execution agent, a result analysis agent, a report generation agent, and an aggregation agent. Each agent works collaboratively under the scheduling of the collaborative scheduling kernel. The research task workflow orchestration engine, connected to the collaborative scheduling kernel, is used for research task workflow definition, task scheduling, and execution control.

9. The system according to claim 8, characterized in that, The collaborative scheduling kernel includes a task decomposition unit, a dependency analysis unit, a task distribution unit, and a result aggregation unit; the task decomposition unit is used to break down research objectives into multiple sub-tasks; the dependency analysis unit is used to identify the dependencies between sub-tasks; and the task distribution unit is used to allocate sub-tasks according to the sub-task type and the capabilities of each Agent. The result aggregation unit is used to collect and integrate the output results of each Agent after the completion of each stage of the task.

10. The system according to claim 8, characterized in that, Each skill module in the skill system is an independent plug-in unit that interacts with the corresponding Agent through a standardized interface, supporting the dynamic loading of new skill modules and the dynamic unloading of no longer needed skill modules during system operation.

Images