arXiv 2606.11290

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

By Lingzhi Yuan, Chenghao Deng, et al.

Published 2026-06-09

Mindmap

Browse the paper's core ideas, clusters, and relationships in a structured outline.

Large Language Model (LLM)-based multi-agent systems are increasingly powerful, but current agentic workflow optimization paradigms make an unsatisfying trade-off. Task-level methods spend substantial offline compute yet deploy only a single workflow, leaving complementary candidates unused, while query-level methods synthesize a new workflow per query at substantial inference cost. Our motivating analysis shows the…

View the original paper on arXiv