Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

AI 导读

该研究提出Growing Harness方法，通过任务反馈将重复控制转化为可重用代码，减少LLM调用并提升效率，实验显示其在多个基准测试中表现最优。

This paper introduces Growing Harness, a method that transforms recurring control into reusable code via task feedback, reducing LLM calls and improving efficiency across benchmarks.

重点速览

• 通过任务反馈将控制决策转化为可执行代码，降低LLM调用频率 Converts control decisions into executable code via task feedback to reduce LLM calls
• 实验显示成功率在多数基准测试中领先，部署成本降低超74% Achieves highest success rates in most benchmarks with 74%+ cost reduction
• 代码重用机制使小模型也能保持44.7%的高成功率 Code reuse maintains 44.7% success across small models
• 失败引导训练框架能联合修复多个错误并回滚有害修改 Failure-guided training jointly fixes errors and rolls back harmful changes
• 相比传统工具调用方式，LLM调用减少76-91.8% Reduces LLM calls by 76-91.8% vs traditional tool calling

一句话：Growing Harness通过代码重用显著提升LLM代理效率，为小模型应用提供新范式。 / Growing Harness enables efficient LLM agents through code reuse, offering new paradigms for small model applications.

推荐理由 该方法创新性地将控制逻辑转化为代码，为LLM代理优化提供可落地的解决方案。