The engine
is not
the car.引擎
不等于
整台车

Yes. With one caveat that changes how you build.是的。但有一个会改变你造法的前提

A model converts context into tokens the way an engine converts fuel into torque. Nothing in it steers, brakes, remembers where you were going, or refuses to drive off a cliff. 模型把上下文变成 token，就像引擎把燃料变成扭矩。它本身不会转向、不会刹车、不记得你要去哪，也不会拒绝开下悬崖。

All of that is the harness: the code around the weights that decides what the model sees, what it can touch, when it stops, and how you know whether it worked. In a shipped AI product, the harness is where most of the engineering hours and almost all of the failure modes live. 这些全都归 Harness 管：包在权重外面的那一层代码，决定模型看到什么、能碰什么、何时停下、以及你如何知道它有没有做对。在真正上线的 AI 产品里，绝大多数工时和几乎全部故障，都发生在 Harness 里。

But a chassis doesn't change an engine's torque curve — and a harness absolutely changes a model's measured ability. That single asymmetry is where the metaphor starts to leak. 但底盘不会改变引擎的扭矩曲线 —— 而 Harness 一定会改变模型被测出来的能力。这个不对称，就是比喻开始漏水的地方。

Buy the engine. Build the vehicle. Expect to rebuild it when the engine changes. 引擎靠买，车靠造；引擎一换，车就得重造。

Five questions that make it visible五个问题，让它现形

A harness is invisible because it's the part nobody demos. Open any AI product — a coding agent, a support bot, a research assistant — and ask these five. The answers are the harness. Harness 之所以看不见，是因为它是没人拿来演示的那部分。打开任何 AI 产品 —— 编程 agent、客服机器人、研究助手 —— 问下面五个问题。答案本身就是 Harness。

Harness is a borrowed word. Twice.Harness 是借来的词，借了两次

It didn't start in AI. In software engineering, a test harness is the code that sets up, feeds, runs and checks a unit under test. AI borrowed it for evaluation harnesses — the machinery that pipes a benchmark into a model and scores the output. Then agent builders borrowed it again for the loop, tools and context that wrap a model in production. Research papers more often call this last one scaffolding. 它不是从 AI 开始的。在软件工程里，test harness（测试挂载）是负责搭建、喂数据、运行并检查被测单元的那段代码。AI 先把它借去做评测 harness——把基准题送进模型、再给输出打分的那套机器。随后 agent 开发者又借了一次，用来指生产环境中包住模型的循环、工具与上下文。论文里更常把最后这一种称作 scaffolding（脚手架）。

Keep the three senses separate when you read, or you'll think two papers disagree when they're talking about different objects. 阅读时把这三层含义分清楚，否则你会以为两篇论文互相矛盾，其实它们说的根本不是同一个东西。

The parts list零件清单

Eight components. If your system is missing three of them, it's not an agent — it's a demo. 八个部件。如果你的系统少了其中三个，那它不是 agent，只是个 demo。

Fifty lines, three tools, one scoreboard五十行代码，三个工具，一块记分板

Don't start with a framework. Start with a loop you can hold in your head, on a task where "correct" is machine-checkable — tests pass, the number matches, the file compiles. 别从框架开始。从一个你脑子里装得下的循环开始，选一个"对不对"能被机器判定的任务 —— 测试通过、数字对上、文件能编译。

# the whole idea, minus the yak-shaving
state = {goal, files, notes}

for step in range(MAX_STEPS):          # stop condition = brakes
    ctx  = assemble(state)             # fuel injection — most wins here
    out  = model(ctx, tools)           # the ENGINE
    if out.is_final: break
    if violates(policy, out): halt()   # guardrail, not a prompt
    obs   = run_tool(out.call)         # the drivetrain
    state = update(state, out, obs)    # memory
    log(step, ctx, out, obs, cost)     # dashboard

score(result, rubric)                  # the test track

Three caveats三个前提

And one loop the metaphor can't hold at all: vehicles don't build engines. But agent harnesses generate the traces, environments and reward signals used to train the next model. The vehicle is quietly reshaping the engine — which is why "just a wrapper" has been wrong for a while now. 还有一个比喻完全装不下的回路：车不会造引擎。但 agent 的 Harness 会产出用于训练下一代模型的轨迹、环境和奖励信号。车正在悄悄重塑引擎 —— 这也是"不过是个套壳"这句话早就不成立的原因。

Go read these six先读这六个

A four-week training plan一份四周训练计划

Reading gives you the vocabulary. Only a measured ablation gives you the answer.读文献给你词汇，只有做过一次有数据的拆件实验，才能给你答案。

Sources出处

Everything asserted above traces back to one of these. Primary sources first; read the papers, not the summaries of the papers.上面每一条判断都能追溯到下面某一项。优先读一手资料 —— 读论文本身，而不是论文的转述。