Harness Engineering：为什么2026年最好的AI工程师停止编写代码

A researcher tested the same AI model on the same coding benchmark twice. The first time, it scored 42%. The second time, it scored 78%. Same model. Same test. Same everything.

一位研究人员在同一个编码基准测试上测试了同一个 AI 模型两次。第一次，得分 42%。第二次，得分 78%。同一个模型。同一个测试。一切都相同。

The only thing that changed was the harness.

唯一改变的是 harness。

Not the prompts. Not the model. Not the temperature setting. The harness — the system of rules, tools, skills, memory files, and feedback loops that wraps around the AI and tells it how to behave.

不是 prompts。不是模型。不是 temperature 设置。harness — 围绕 AI 并告诉它如何行为的规则、工具、技能、内存文件和反馈循环的系统。

This is the most important finding in AI-assisted development right now. And almost nobody outside of a few engineering teams is talking about it.

这是当前 AI-assisted development 中最重要的发现。而且除了少数工程团队之外，几乎没有人谈论它。

LangChain proved the same thing independently. Their coding agent went from outside the top 30 to the top 5 on Terminal Bench 2.0 by changing nothing about the model. They only changed the harness. OpenAI proved it too. Their Codex team built a production application with over one million lines of code where zero lines were written by human hands. The engineers did not write code. They designed the harness.

LangChain 独立证明了同样的事情。他们的 coding agent 在 Terminal Bench 2.0 上从前 30 名之外跃升到前 5 名，而模型本身没有任何变化。他们只改变了 harness。OpenAI 也证明了这一点。他们的 Codex 团队构建了一个生产应用，代码超过一百万行，没有一行是由人类手写的。工程师没有写代码。他们设计了 harness。

The skill of designing these harnesses has a name now: harness engineering. It was coined weeks ago. It is already getting traction on YouTube and in engineering blogs. But there is almost zero mainstream coverage. Most developers have never heard the term.

设计这些 harnesses 的技能现在有了一个名字：harness engineering。这个名字几周前被创造出来。它已经在 YouTube 和工程博客上获得关注。但是主流报道几乎为零。大多数开发者从未听说过这个术语。

That is about to change. And the developers who learn it first will have a career advantage that is extremely hard to replicate.

这一切即将改变。最早学会它的开发者将拥有极难复制的职业优势。

This is the complete guide. What harness engineering is. Why it matters more than the model you use. How to start doing it toda...