<scp>Meta-Agent-Workflow:</scp> Streamlining Tool Usage in LLMs through Workflow Construction, Retrieval, and Refinement

Large language models (LLMs) have recently shown significant advancements and are increasingly used as key components in automated agents for various web-based tasks. Typically, this agentization is achieved by carefully prompting LLMs to guide their behavior in using tools for specific tasks. However, this approach can be limited by the complexity of tasks and the inherent capabilities of LLMs. To enhance task-specific performance, a pre-defined workflow approach can be employed, reducing repetitive and error-prone planning for particular tasks. This workflow-driven process is especially well-suited for industrial applications, where task-specific agents can be easily configured using visual interfaces supported by various open-source platforms. In this paper, we introduce a novel framework called Meta-Agent-Workflowto create, retrieve, and refine agent workflows. Experiments on ToolBench demonstrate that our framework effectively transforms LLM tool-reasoning processes into task-specific workflows, retrieves workflows for different tasks based on various queries, and updates them based on execution feedback. We also open-source our code and follow the workflow architecture of an open-source agent platform (e.g., Dify) to facilitate further industrial and community use. The Meta-Agent-Workflow will be open-sourced in https://github.com/testlbin/meta_agent_workflows.