Skip to content
Hi, Bot
Back to blog
explainerbuild-with-aifor-kidsgrown-ups

A brain in a jar needs a body: what an AI harness is

An AI model can think, but it can't do anything on its own. The harness is the program around it that gives it tools, a notepad, a checklist and rules. Here's how the four parts work.

November 20, 2026 · Hi, Bot

Picture a very smart brain in a jar.

Ask it anything and it answers. But it can't open an email. It can't save a file. It can't remember what you said five minutes ago unless you say it again.

That's an AI model, the part of a chatbot that does the thinking. It can think. It can't act.

The body around the brain

A harness is a program built around the model. It lets the model do things, not just talk.

A model plus a harness is called an agent. An agent is an AI that can finish a job, not just answer a question.

Agent = Model (the brain) + Harness (the body and the rules)

The harness is not the AI. People write it. It runs next to the model and handles everything the model can't.

Here is what it does.

1. Hands: tools

The model can say, "I should search the web." It can't search the web. It only makes words.

The harness reads those words and does the search. Then it gives the results back to the model. The same goes for opening a file or running code.

The model picks the move. The harness makes it.

2. A notepad: memory

Models forget. In a long chat, the early parts slip away. Start a new chat and everything is gone.

The harness keeps notes. It saves things like what you're building and what you already tried. Then it shows those notes to the model again later.

If a chatbot ever said "I remember you like space stories," that wasn't the model remembering. The harness saved a note and handed it back.

3. A manager: the loop

A plain model gives one answer and stops.

A harness keeps it going in a loop:

  1. Try something.
  2. Look at what happened.
  3. If it didn't work, try another way.
  4. Repeat until it's done.

Here's a small example. You say, "Will it rain Saturday? Put it on my calendar." The model asks for a weather search. The harness runs it. The results come back. The model asks to add a calendar event. The harness does that too. If the calendar says "error," the model sees it and tries again. You only see the last step.

4. A safety net: guardrails

Some rules are not up to the model.

A harness might let the AI read a list of names but never delete one without asking a person first. The model can't argue its way past that rule. The rule isn't in the model. It's in the harness, and the harness holds the keys.

Does the body matter more than the brain?

Often it matters a lot. An okay model with good tools and a good loop can beat a smarter model with no tools. A genius with no hands can't fix your bike. A decent helper with a toolbox can.

But a harness doesn't make a model wiser. If the model misunderstands the job, a good harness just helps it do the wrong thing faster. Harnesses have bugs too. So someone still has to check the work. Sometimes that someone is you.

Try it on any AI tool

Ask four questions:

  • Hands: What can it actually do? Search? Open files? Run code? Or only talk?
  • Notepad: Will it remember me tomorrow? Where is that saved?
  • Manager: Does it give one answer, or keep trying until the job works?
  • Safety net: What is it not allowed to do? Who decided that?

If you can answer all four, you understand that tool better than most people who use it.

› Keep reading

Get the next dispatch.

Occasional, never spammy.

By submitting, you agree to our Terms and Privacy Policy.