Writing

Most people don't know this, but the AI model inside an agent can't do anything except write text.

By ·

So when a vendor shows you an "AI agent", the thing that decides what it's allowed to touch is a different part, and that part rarely shows up in the demo.

It's called the harness. Theo from t3.gg, a developer who builds tools for AI coding, wrote a long post about it on X last week because he was tired of explaining it. It's the clearest explanation I've read.

A model can write "open this file", but writing it doesn't open anything. The harness is the code around the model that turns that text into action, and it has five parts:

1. Tools that do things, like reading a file. 2. A list of those tools, so the model knows what it has. 3. A fixed format the model uses to ask for a tool. 4. Code that spots that format and runs the right tool. 5. A way to hand the result back to the model, so it can pick its next step and keep going until the task is done.

Put a model and a harness together and you get what everyone now calls an agent. That's also why two agents on the same model can act so differently, since their tools and instructions differ.

When I build Klarefi, this is where most of my decisions go. Its agent gets tools to read what an applicant sends, ask for what's missing and look the company up in the KVK. But it gets no tool for the decision, so that always stays with the credit team.

So before you buy an agent, ask to see its list of tools, because that list is everything it can do.

I'm thinking of writing explainers on this. Which would you want first: how a harness works inside, how to set one up, or why so many companies are building on them right now?