What is an agent harness?什么是智能体系统（agent harness）？
You've seen the demos. A chat window drafts an email, summarizes a contract, writes a plan. Impressive — and almost useless on its own. A model by itself is a very good new hire with no desk, no phone, no files, no permissions, and no supervisor. It can talk. It can't work.你见过那些演示：一个聊天窗口起草邮件、总结合同、写计划。很惊艳——但单独拿出来几乎没用。一个模型本身，就像一位能力很强的新员工：没有工位、没有电话、没有档案、没有权限，也没有主管。它能说，但没法干活。
The thing that turns a model into a worker is the . The model is the engine; the harness is the rest of the car — the steering, the brakes, the fuel gauge, the seatbelts. Nobody commutes on a bare engine.把模型变成劳动力的东西，叫 。模型是发动机，harness 是车的其余部分——方向盘、刹车、油表、安全带。没有人骑着一台裸发动机去上班。
What's actually in it它里面到底有什么
Five things, roughly. Tools: connections to your email, documents, spreadsheets, and databases, increasingly through open standards like so each connection doesn't have to be invented from scratch. Memory: what the agent knows about your business and remembers between tasks. Permissions and : what it's allowed to touch and how much it's allowed to spend. : the places where it must stop and ask a human. and : the logs that show what it did, and the tests that show whether it's still doing it well.大致是五样东西。工具：连上你的邮件、文档、表格、数据库——越来越多地通过 这样的开放标准，让每一条连接不必从零发明。记忆：它对你业务的了解，以及任务之间记得住的东西。权限与：它能碰什么、最多能花多少。：它必须停下来问人的那些位置。与：日志证明它做了什么，测试证明它还做得好不好。
What breaks without one没有它会坏在哪
The same thing, every time. The demo works. The pilot mostly works. Then, quietly, quality drifts — a changed supplier format, a new edge case, a model update — and nobody notices, because nothing is measuring it. The first sign of trouble is a customer pointing it out. Without logs, you can't reconstruct what happened. Without gates, it happened in your name.坏法每次都一样。演示能跑，试点也基本能跑。然后，质量悄悄漂移——供应商换了格式、冒出一个新边角情况、模型更新了——没有人发现，因为没有任何东西在测量它。出问题的第一个信号，是客户指出来。没有日志，你无法还原发生了什么；没有审批节点，这一切还是以你的名义发生的。
What to ask anyone selling you an agent有人向你推销智能体时，该问什么
- Can I see the log of everything it did last week?我能看到它上周做过的每件事的日志吗？
- What happens when it's wrong — who catches it, and how fast?它出错时会怎样——谁来发现，多快发现？
- What does one task cost, and what stops that cost from surprising me?一个任务花多少钱？什么机制保证这个数字不会吓到我？
- If we part ways, what do I keep?如果我们不再合作，我能留下什么？
If the answers are vague, you're being sold an engine. The model gets the headlines. The harness decides whether any of this works.如果这些问题的回答很含糊，那你买到的只是一台发动机。上头条的是模型，但这套东西能不能用，由 harness 决定。