FlowScout builds reliable AI agent workflows automatically

Researchers from Fudan University's School of Computer Science have introduced FlowScout, a novel framework designed to automate the creation of reliable agentic workflows for LLM-based automation systems. Traditional agentic workflows rely heavily on manual construction and domain expertise, while existing automated approaches often produce simulated workflows that lack real tool execution stability.

FlowScout addresses these limitations by mining historical task-solving records to construct a common tool coordination skeleton as a directed graph composed of LLM nodes, tool-calling nodes, and dependency edges. The framework then refines this initial workflow using Monte Carlo tree search guided by real execution feedback. In evaluations across four representative task domains, FlowScout demonstrated significant improvements over three baselines—PM4Py, ReAct, and AFlow—achieving at least 92.69% higher tool invocation correctness and 17.66% better execution quality, while maintaining lower performance variation across repeated runs.