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

2026-06-09 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

FlowBank is a novel three-stage framework designed to optimize Large Language Model (LLM)-based multi-agent systems by addressing the trade-off between task-level and query-level workflow optimization. Existing methods either deploy a single workflow after substantial offline compute or synthesize a new workflow per query at high inference cost. FlowBank proposes building a compact bank of reusable, complementary workflows and adaptively selecting among them at inference time. It tackles three coupled problems: generating complementary candidates, compressing them into a small deployable portfolio, and assigning queries to the optimal workflow. The framework includes DiverseFlow for high-coverage candidate generation, CuraFlow for compact portfolio compression, and a matching stage for query-workflow routing based on predicted utility. Across five benchmarks, FlowBank achieved the highest average score, improving over strong automated and handcrafted baselines by 4.26% and 14.92% relative, respectively, while remaining cost-competitive.

Key takeaway

For AI Engineers designing multi-agent LLM systems, you should reconsider traditional single-workflow or per-query generation paradigms. FlowBank demonstrates that building a compact, diverse portfolio of precomputed workflows and adaptively matching queries to them significantly improves performance and cost-efficiency. Implement a three-stage approach focusing on generating complementary candidates, compressing them, and dynamically routing queries to optimize your agentic workflows. This strategy can yield substantial gains over current automated and handcrafted baselines.

Key insights

Optimizing LLM multi-agent systems requires a compact bank of reusable, complementary workflows selected adaptively at inference time.

Principles

• Complementary workflows outperform single-workflow or per-query generation.
• Precomputed workflows can solve many queries handled by expensive generation.
• Workflow optimization involves generation, compression, and adaptive matching.

Method

FlowBank uses DiverseFlow for candidate generation, CuraFlow for portfolio compression, and edge-value prediction on a bipartite graph for query-workflow matching.

In practice

• Explore precomputing diverse workflows for common query subsets.
• Implement a system to dynamically route queries to optimal precomputed workflows.
• Evaluate workflow portfolios based on coverage, compactness, and utility.

Topics

• Multi-agent Systems
• Large Language Models
• Workflow Optimization
• Query-Adaptive Systems
• Portfolio Optimization
• Inference Cost Reduction

Best for: Research Scientist, AI Architect, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• HPC Coding Agent - Part 2: An MCP Tool for Code Optimization with OpenEvolve — Integrating evolutionary code optimization tools like OpenEvolve with AI agents enhances code performance beyond single-pass LLM generation.
• GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving — GraphFlow uses a unified graph (wGraph) to dynamically manage LLM-agent workflows, improving performance and memory efficiency.
• Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI — AI agents are fundamentally transforming software development and research by enabling extensive task delegation and automation.
• Common workflow patterns for AI agents—and when to use them — Structuring AI agent autonomy with sequential, parallel, or evaluator-optimizer workflows optimizes performance and reliability.
• From Workflows to Multi-Agent Systems: How to Choose — Choosing between workflows, single agents, and multi-agent systems depends on task complexity, dynamism, and cost-benefit analysis.
• Gemini 3.1 Pro, Claude Sonnet 4.6 & The OpenClaw Hire That Killed the Chatbot Era - EP99.35 — Smaller, cost-effective models in agentic loops are increasingly outperforming frontier models for many practical tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.