FlowScout Generates Reliable Tool-Using Agent Workflows from Execution Feedback.

Building reliable automation systems with Large Language Models (LLMs) often involves creating agentic workflows that orchestrate LLMs, tools, and control logic. However, designing these high-quality workflows typically requires extensive manual effort and specialized domain knowledge. While some research explores automatic workflow generation from historical data, these often produce LLM-centric workflows that simulate tool executions rather than integrating real tools, limiting their practical utility and stability. To overcome these limitations, researchers introduced FlowScout, an execution-guided framework designed to generate tool-integrated agentic workflows. FlowScout represents workflows as directed graphs comprising LLM nodes, tool-calling nodes, and dependencies. It begins by identifying a common tool coordination pattern from historical task-solving records to form an initial workflow. This initial workflow is then refined through a Monte Carlo tree search process, which is guided by real execution feedback. Experiments across four task domains demonstrate that FlowScout's generated workflows achieve at least a 92.69% improvement in tool invocation correctness and at least a 17.66% improvement in execution quality over baseline methods, while also reducing performance variability.