Harness
Everything around the model — the prompts, tools, retries, permissions, and checks that turn a text generator into something that can be trusted with work.
Also called agent harness, the harness
The model is a component. The harness is the system: what context it sees, which tools it can reach, what it is permitted to write to, how failure is caught, and what happens on retry.
Much of the quality difference between two people using the same model comes from the harness rather than the prompt. Harness work is also very easy to mistake for progress. It looks like engineering and can absorb unlimited time without shipping anything.
I ask myself what the harness caused. I need to name something that shipped or a failure it caught, not merely a feature it has. If I cannot, I was procrastinating with good tooling.
See also Verification and Context engineering.
Linked from
- In defense of reading every word post
- In defense of the boring harness post
- Ten agents and what they actually cost me post
- Agent estate topic
- Context engineering topic
- Dark factory topic
- Legacy code topic