FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse

FlowBank introduces a three-stage framework for optimizing agentic workflows in LLM-based multi-agent systems, addressing the limitations of existing task-level and query-level methods. The framework includes DiverseFlow for generating a diverse set of workflows, CuraFlow for compressing these workflows into a compact portfolio, and a matching mechanism for efficiently routing queries to the most suitable workflow based on predicted utility. FlowBank demonstrates superior performance across five benchmarks, outperforming both automated and handcrafted baselines by 4.26% and 14.92% respectively, making it a significant advancement for practitioners seeking efficient workflow management in AI applications.