---
id: 1
title: "what is an ai harness, and how is it different from the model and the app?"
description: "think of a restaurant. the model is the chef, the harness is the kitchen, the app is the front of the house. you can use the inside without the outside, never the outside without the inside."
date: 2026-11-17
updated: 2026-11-17
tags: [agents, harness, ai]
url: https://attenzioneplease.com/notes/what-is-an-ai-harness
---

# what is an ai harness, and how is it different from the model and the app?

think of a restaurant. the model is the chef, the harness is the kitchen, and the app you type into is the front of the house. you can use the inner layers without the outer ones, but not the outer ones without the inner ones.

## the chef is the model

the model is raw intelligence. talent. a chef who studied every cookbook and is incredibly good at cooking.

but without the fire he doesn't cook. and without the recipes coming in, which is the context, he doesn't know what to make. a few years ago the chef was already there, there just wasn't a kitchen powerful enough. so he did his own small things.

## the kitchen is the harness

model, tools, prompts and max steps. that's a harness. take the output of a model and pass it back to a model again, that's already a harness. it's a loop.

when i say harness i include its configuration. the configured harness has the recipes. which tools the model gets offered is up to the harness.

and a kitchen does nothing on its own if there's no place to go eat.

## the front of the house is the app

you walk into the restaurant, sit down, the forks are there. you don't see any machines. that's chatgpt.com.

take images. the chatgpt app doesn't make the image. the harness passes the request to the model, the image comes back, and the app shows it to you nicely. same as google. the google ui isn't google. there's a front end, a back end and a lot of stuff around it.

claude code is partly a harness. it's an app plus anthropic's harness underneath. codex is the same with openai's. and one harness can sit behind several apps, a web app and a terminal app can both use the same one.

## then where is the agent?

when you resume a session in the terminal you're resuming an agent. it's your harness with the mission you gave it, already in motion.

same in chatgpt. you write "hi, send an email" and it starts doing things. that's an agent you created by invoking the harness through the interface, which then uses the model.

## and the waiter

there's also a proxy. it's where the orders pass. the waiter takes your order to the place that handles orders, and if you ask for something bad he tells you no, we can't make that. but forget about it for now, it's not the point.

## why everyone mixes them up

the harness is invisible. nobody sees it. if you never built something with a model in code, you don't get what it is. to most people chatgpt just does more stuff now.

i didn't invent the kitchen thing btw. but it works really well, i think
