Anthropic outlines evaluation-driven design for AI agent tools
Anthropic’s engineering guidance recommends designing tools around specific agent workflows rather than directly wrapping existing APIs. Its approach combines prototyping, scenario-based evaluation, transcript analysis, and iterative refinement, with emphasis on clear descriptions, namespaced tools, high-signal responses, and token-efficient outputs. Anthropic says these recommendations largely draw on its experience optimizing internal tools with Claude Code.
The guidance gives AI agent developers a concrete framework for improving tool selection, parameter accuracy, context usage, and response quality through evaluation-driven iteration.
一个小改动,让 Claude 在 SWE-bench 上刷新纪录:工具描述有多重要
原文:mp.weixin.qq.com/s/4VPb8r-n6…(2026年10月19日)
很多人给 Agent 接工具,做法很朴素:把现有的 API 一个个包一层,丢给模型。
结果是,Agent 要么调错工具,要么参数填错,要么被一大坨返回数据淹没。
Anthropic 工程团队发了一篇博客,讲的就是怎么写出 Agent 真正好用的工具。更有意思的是,文中大部分建议,来自他们用 Claude Code 反复优化自家内部工具的经验。