A good AI harness lets you use much cheaper models

· Published · 2 min read

A good AI harness is a great way to… save money.

We often talk about the quality gain in generated code thanks to a good harness. But we talk a lot less about the other side of it: at equal quality, a harness also lets you use much cheaper models.

The test

I ran a very simple test:

  1. I set up a codebase following a layered architecture.
  2. I deliberately broke the import rules in a few files.
  3. I stated the rules in a CLAUDE.md (level-0 harness).

Without a sensor, Haiku falls for it

Result on a feature request:

  • Opus saw that the existing code violated the rules, and managed to produce correct code.
  • Haiku fell straight into the trap: it reused the bad code and introduced new errors.

With a sensor, Haiku fixes it on its own

I then reinforced the harness with ESLint import rules (with exceptions to keep the existing errors in place).

Result: Haiku makes the mistake, sees the (prompt-like) message from my custom rule… and fixes it directly, perfectly.

The economics: a model 4 to 5 times cheaper becomes viable

We are talking about a model that becomes viable on certain well-scoped tasks, at a cost 4 to 5 times lower. What CLAUDE.md could only describe as an intention, an ESLint rule turns into a constraint: the cheaper model no longer even needs to fully understand the rule, just to know how to fix it once the sensor flags it.

It’s the kind of sensor I set up when I join a team.

Today, we are still fed on cheap tokens, paying only a fraction of their real cost… but within a few years, the party will probably be over. And those who got ahead on their harness will be rubbing their hands.

So, you… where is your harness at?

Sources