Coupled Hierarchical Search over Topology and Execution for Agentic Workflow Synthesis
06:00 · July 27, 2026 · arXiv cs.AI RSS
Although structured workflows empower Large Language Models (LLMs) to tackle complex problems, automating their creation is severely hindered by a vast combinatorial search space, frequently resulting in inflexible and resource-heavy offline training dependencies. To address this, we conceptualize workflow generation as an intertwined topology-and-execution search paradigm, where the broader topological layer dictates subtask boundaries and lower-level execution outcomes actively reshape the topology itself. Building on this foundation, we introduce HierFlow, a training-free, test-time hierarchical search architecture that automates agentic workflow design by merging feedback-guided topology adjustments with a fast, MCTS-inspired tree search for sub-workflow optimization. HierFlow maximizes efficiency through an intelligent gating module that selectively triggers execution-level searches based on contextual necessity, a mechanism we further support with an in-depth analysis detailing how varying degrees of cross-task coupling impact the effectiveness of hierarchical splitting. Comprehensive testing across question answering, mathematical reasoning, and code generation benchmarks confirms that HierFlow consistently outperforms strong baselines, delivering an optimal balance of high-quality results and computational efficiency without any additional training overhead.
Summary
HierFlow addresses the challenge of automatically generating structured workflows that let large language models solve complex tasks through multi-step reasoning and tool use. Manual designs remain static and scale poorly, while prior automated approaches either require costly offline training or conduct exhaustive search within a single proxy representation such as code or topology. The authors therefore treat workflow synthesis as a coupled search over two interdependent spaces: an upper-level topology that defines task boundaries and a lower-level execution space that implements each subtask in executable code. Execution outcomes can in turn trigger revisions to the topology itself, creating a closed feedback loop rather than a fixed pipeline.
The resulting system, HierFlow, performs this search at test time without any training. It begins with an initial task decomposition in the topology space, then applies feedback-guided refinement at the upper level while selectively invoking a lightweight Monte Carlo Tree Search at the lower level to optimize individual sub-workflows. An adaptive gating module decides whether lower-level search is warranted for a given subtask, using a proxy signal that reflects both structural uncertainty and cross-task dependencies. This selective activation avoids unnecessary computation when sub-workflows are already adequate or when strong interdependencies make isolated optimization counterproductive.
The paper further supplies a coupling-aware analysis that characterizes the conditions under which hierarchical decomposition and gating remain effective. When cross-subtask dependencies are sparse, the separation of topology and execution search yields clear gains in both solution quality and search cost. As dependency density increases, the benefit diminishes and the gating mechanism becomes essential to prevent wasteful exploration. Empirical results across question answering, mathematical reasoning, and code generation benchmarks show that HierFlow consistently surpasses strong baselines while preserving the flexibility of a purely test-time, training-free method.
Why it matters
This research provides Dutch AI researchers and advanced practitioners with a highly novel, resource-efficient methodology for building autonomous LLM agents. Its training-free approach lowers computational overhead, aligning well with the Dutch and broader EU focus on sustainable, accessible AI solutions for SMEs and enterprise deployments.
