A raw model is a brain in a jar. It can think. But on its own, it cannot open Slack, send an email on your behalf, or even know when to quit.

A harness is the software around the model that lets it do those things. It gives the model access to tools and information, manages permissions, and controls when it should retry, stop, or ask you for help.

A simple way to describe it is:

Agent = Model + Harness

For a coding agent, the harness provides tools to read files, edit code, and run tests. It also manages the instructions and feedback the model receives.

Two apps can use the same model and still perform differently because their harnesses give it different tools, instructions, and ways to check its work.

That’s why the software around a model matters so much. OpenAI’s article on harness engineering describes how its team built an environment where Codex could write code, run checks, and respond to feedback.

When you compare AI tools, look at how they help the model complete your task, and which model they use.