Jargon:Harness
1 min read
Say this at the party
“Everything wrapped around an AI model that lets it actually do things: its tools, its instructions and the loop it runs in.”
In plain English
The model is the brain. On its own, it can only write words back to you.
The harness is the body and toolbox around it: the ability to search the web, run code, and read and edit files, plus the instructions and checks that keep it on track. Claude Code is a harness. So is NotebookLM.
That's why the same model can feel brilliant in one app and useless in another. It's often the harness, not the brain.
Where it came from
Engineers have talked about a "test harness" for decades, meaning the kit that runs and checks a piece of software. In AI it came to mean the setup around a model, and Ethan Mollick brought it to a wider audience in his 2026 guide that splits choosing an AI into the model, the app and the harness.
See it in action
In a plain chat box, a model can tell you how to fix a spreadsheet. In a harness with access to your files, it opens the spreadsheet, fixes it and saves a copy.
What people get wrong
- "Which AI is best?" is often the wrong question. A good model in a weak harness can lose to a decent model in a great one.
- A harness isn't a safety strap. It's what gives the AI its powers, which is exactly why it needs limits.
Go deeper
Where it comes up
Related jargon
Drafted with AI, checked by the AI Council (Claude, ChatGPT and Gemini) and signed off by Phin.