# FlowBank: Query-Adaptive Agent Workflow Optimization via Precomputation and Reuse

> FlowBank proposes a new paradigm for agent workflow optimization. By building a reusable workflow library and adaptively selecting workflows during inference, it addresses the dilemma between offline computation and inference costs faced by traditional methods.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-06-09T17:58:21.000Z
- 最近活动: 2026-06-11T01:48:08.712Z
- 热度: 117.2
- 关键词: 智能体工作流, 大语言模型, 多智能体系统, 查询自适应, 工作流优化, 预计算, FlowBank
- 页面链接: https://www.zingnex.cn/en/forum/thread/flowbank
- Canonical: https://www.zingnex.cn/forum/thread/flowbank
- Markdown 来源: floors_fallback

---

## FlowBank: A New Paradigm for Agent Workflow Optimization

FlowBank proposes a new paradigm for agent workflow optimization. By building a reusable workflow library and adaptively selecting workflows during inference, it resolves the dilemma between offline computation and inference costs in traditional methods. Its core idea is to treat task-level and query-level optimization paradigms as complementary, constructing a compact complementary workflow library to balance performance and cost.

## Background: The Dilemma of Agent Workflow Optimization

Multi-agent systems based on large language models are powerful, but current workflow optimization faces a trade-off dilemma: task-level methods search for optimal workflows offline but waste complementary candidates, while query-level methods synthesize workflows from scratch for each query, leading to extremely high inference costs. The essence is the need to choose between the high cost of offline precomputation and the inference cost of per-query generation.

## Core Insight: Complementary Rather Than Competitive Optimization Approach

The FlowBank team found that task-level and query-level paradigms are complementary: workflows from offline searches cover different query subsets, and a single optimal workflow cannot meet all needs; many queries can be solved by precomputed workflows instead of expensive query-level generation. Thus, they propose a new goal: building a compact, reusable complementary workflow library and adaptively selecting workflows during inference.

## FlowBank Framework: A Three-Stage Solution

FlowBank solves the problem through three stages: 1. The diversification stage uses the DiverseFlow algorithm to generate diverse candidate workflows covering a wide range of query types; 2. The curation stage uses the CuraFlow algorithm to compress candidates into a compact combination, minimizing redundancy while retaining coverage capability; 3. The matching stage models query-workflow assignment as edge value prediction, routing queries to the workflow with the highest expected utility.

## Experimental Validation: Performance of FlowBank

In five benchmark tests, FlowBank achieved the highest average score while maintaining competitive cost efficiency. It improved by 4.26% compared to the strongest automated baseline and 14.92% compared to the strongest manually designed baseline, breaking through the traditional performance-cost trade-off dilemma.

## Practical Significance and Application Prospects

For system architects: It provides an efficient and flexible agent system framework, with precomputed libraries reducing inference costs; For researchers: It opens up a new direction for workflow combination optimization; For practical applications: It reduces operational costs and improves service consistency in scenarios such as customer service and code generation.

## Summary and Outlook

FlowBank rethinks the optimization paradigm, breaking through traditional limitations with precomputation + adaptive selection. The core innovation is building a complementary workflow library and making intelligent selections; this idea may inspire other AI scenarios that balance precomputation and flexibility. In the future, more advanced diversification strategies, compression algorithms, and routing models can be explored.