Build your own

Because the four parts are separate, they can be swapped. The same model can sit in a chat app, a coding agent or an assistant inside your email. Some harnesses let you choose among models from several companies,¹² and the same harness can be given different tools. So what you experience is the combination: how good the answers are, where your words go and who approves actions. It is not the model alone.³

Here are three combinations, described by their parts. The verdicts are general rules of thumb written for this book, not a review of any product, and every real product is set up in its own way.

• A mid-size model in a chat app, with web search. Good for questions, drafts and explanations, including ones that need fresh information. Not good for acting on your files or other systems: it mostly writes text back to you. Your words go to the company that runs the app, under the terms of your account. There is little to approve, but every page it reads is text that could try to give it orders.
• A top model in a coding agent, with your files and a terminal. Good for changing code, running tests and other long, many-step jobs on a project; coding agents of this kind edit files and run commands.⁴⁵ Your files are read on your computer, and what the model needs to see is sent to the model’s provider. You approve actions, command by command or through settings you choose.
• A small model on your own computer, with the text you give it. Good for narrow, private jobs, such as tagging or summarizing documents you supply; once downloaded, such models can run without an internet connection.⁶⁷ Not good for broad factual questions without a source: one maker says its small model lacks the capacity to store much factual knowledge.⁸ Your words can stay on your machine. You approve whatever you set up.

Run the lawyer’s job through the same questions. He needed real decisions, so the combination that fit was a harness with a tool connected to a real legal database, plus the check his profession required. Any capable model in a setup like that reads real decisions. Even a brilliant one in a setup without it reads none.

ChatGPT launched in 2022 as a chat app.⁹ As of October 2026, the Claude app offered web search as one of its tools.¹⁰ Claude Code and OpenAI’s Codex were coding agents of the second kind.⁴⁵ Hermes Agent, an open-source AI agent, could run on your own server with a model you choose.² Spot it today: name the brain, the body and the hands of the AI tool you use most. The names change; the four parts don’t.

Try it: choose a brain, a body and hands, and see what the combination is good for.

Your turn

Try it. Pick a question from your own work that needs current information, such as a rule or schedule that changed recently in your field, and ask any AI tool you already use. Then look for the loop. Did the app show that it searched, or list its sources? Did it ask you before doing anything? If it showed no search, ask it to find sources and see whether the answer changes. Then test this chapter’s main claim by naming each of the four parts in what you just used. On paper instead: choose a task you would like help with, write out the six steps of Figure 9.1 for it, and mark the step where you would want to be asked first.

Think about it. Think of one AI tool you use at work, or might. If its answers kept going wrong, which of the four parts would you check first? And who in your organization controls that part?

So between your question and its answer, a loop runs. The harness briefs the model, the model predicts an answer or asks for a tool, the harness checks and acts, and the result goes back in until the answer comes out. Every turn of that loop carries your words somewhere: into a briefing, onto a company’s servers, into a tool. Where they go, and who may keep them, is the subject of the next chapter, which opens Part IV.