Agent = model + harness: why the model is only half the tool
People choose AI coding tools by asking which model is best. It is the wrong question, and it leads to disappointment.
There is a formula going around that is worth remembering: agent = model + harness.
What the harness is
The model supplies raw intelligence. The harness is everything wrapped around it that turns intelligence into reliable work:
- How the tool decides what parts of your project to show the model
- Which tools it can call, and how it recovers when one fails
- How it plans a multi-step task and keeps track of where it is
- What it does when a test fails: retry, rethink, or stop and ask
- Where it asks permission, and where it acts
Two products using the identical model will produce different results, because those decisions are different.
Why this matters commercially
If capability came only from the model, every tool would be equal the day a new model shipped. That is not what happens. Tools built on the same model are consistently better or worse than each other, and the gap comes from context handling and error recovery.
It also means a cheaper model in a good harness often beats an expensive model in a poor one. That is a real cost decision, not a theoretical one.
How to judge a harness
Give the same non-trivial task to two tools and watch what happens when something goes wrong. Failure is where harnesses separate.
- Does it notice the test failed, or carry on regardless?
- Does it read the actual error, or guess at a fix?
- Does it stop after two failures, or loop expensively forever?
- Does it tell you honestly that it could not do the thing?
The last one matters most and is the rarest. A tool that says it could not do something is more useful than one that produces something plausible and wrong.
The same idea outside coding
This applies to any AI agent, not just coding ones. When we built Serra, our local AI operator, almost all the work was harness: what she is allowed to touch, what happens when a request does not match anything, how every action gets written down.
The model was the easy part. It usually is.
