HierFlow: Training-Free Hierarchical Search for Agentic Workflow Synthesis

A recent study published on arXiv (2607.21609) presents HierFlow, an innovative architecture designed for hierarchical search at test time, aimed at automating the design of agentic workflows without the need for training. This method views workflow generation as a combined search process involving both topology and execution, where a higher-level topological framework sets the boundaries for subtasks, and execution results influence the topology. HierFlow integrates feedback-driven adjustments to topology with a rapid, MCTS-inspired tree search for optimizing sub-workflows, featuring a smart gating mechanism that initiates execution-level searches based on contextual requirements. This solution tackles the combinatorial search challenges that complicate automated workflow development for LLMs, eliminating the need for rigid and resource-intensive offline training.