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

FlowBank presents a novel framework for optimizing LLM-based multi-agent systems by building a portfolio of complementary workflows rather than searching for a single universal solution or regenerating workflows per query. The approach balances computational efficiency with performance, achieving 4-14% improvements over existing methods while reducing inference costs.

FlowBank addresses a fundamental inefficiency in current LLM agentic systems: the false choice between expensive offline optimization that deploys a single workflow, or costly query-time synthesis. The research reveals that workflows discovered during offline search actually solve different query subsets, suggesting that reusability and diversity are underexploited resources in current systems.

The framework's three-stage pipeline reflects sophisticated system design thinking. DiverseFlow steers candidate generation toward underrepresented queries, avoiding redundancy in the workflow pool. CuraFlow then performs intelligent compression, recognizing that not all discovered workflows merit deployment. The matching stage treats workflow assignment as an edge-value prediction problem on a bipartite graph, enabling dynamic routing based on query-specific utility estimates rather than static assignment rules.

This work matters because it directly tackles the inference cost problem that constrains LLM agent deployment at scale. As organizations push multi-agent systems into production, reducing per-query computational overhead while maintaining performance becomes economically critical. The 4-14% performance gains combined with cost efficiency suggest meaningful practical value.

The research also signals a maturation in agentic systems thinking—moving from searching for universal solutions toward portfolio-based approaches that accept inherent query diversity. This pattern-matching approach could influence how AI infrastructure providers design agent orchestration systems. Future work likely involves scaling these techniques to larger workflow portfolios and investigating whether similar diversification principles apply to other multi-agent optimization problems.

• →FlowBank uses portfolio-based workflow optimization rather than searching for single universal solutions
• →The framework achieves 4-26% performance improvements while remaining cost-competitive compared to baselines
• →Three-stage approach diversifies candidates, compresses redundancy, and enables adaptive query-workflow matching
• →Research demonstrates that workflows solve different query subsets, making diversity a key optimization target
• →Approach reduces inference costs by reusing precomputed workflows instead of synthesizing new ones per query