编码智能体主要读指令文件而非API文档

Rohan Paul · @rohanpaul_ai · X·2026-08-30 23:09·13小时前
AI 导读

一项针对557次真实编码会话和33,097个智能体PR的实证研究发现,编码智能体主要阅读AGENTS.md、CLAUDE.md等指令文件(占35.4%)及自身计划笔记(25.1%),API参考仅占1.3%。文档阅读直接触发代码编辑的情况极少(1328次中仅3次),且未发现智能体对照文档检查工作的案例。研究建议开发者更重视指令文件而非API文档的编写。

Rohan Paul@rohanpaul_ai
60AI 编辑部评分,满分 100

编码智能体主要读指令文件而非API文档

2026-08-30 23:09· 13小时前
AI 导读

一项针对557次真实编码会话和33,097个智能体PR的实证研究发现,编码智能体主要阅读AGENTS.md、CLAUDE.md等指令文件(占35.4%)及自身计划笔记(25.1%),API参考仅占1.3%。文档阅读直接触发代码编辑的情况极少(1328次中仅3次),且未发现智能体对照文档检查工作的案例。研究建议开发者更重视指令文件而非API文档的编写。

Coding agents mostly read the instruction files and notes they were given or wrote themselves, so your AGENTS.md deserves more care than your API reference.

because that is what agents actually read, not your API reference.

The authors tracked 3,033 documentation events across 557 real agent coding sessions, plus file changes in 33,097 agent-written pull requests.

Instruction files like AGENTS.md and CLAUDE.md were 35.4% of that activity, and the agent's own plans and notes another 25.1%. API references got 1.3%, troubleshooting guides 0.4%. Agents also wrote these files nearly as often as they read them.

More importantly, the assumed read-the-docs-then-write-the-code loop barely showed up. Only 3 of 1,328 documentation reads led straight to a code edit, and the authors found no case of an agent checking its work against documentation.

– arxiv. org/abs/2608.20195

Title: "From Agent Behaviour to Agent-Friendly Documentation: An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation"