Tune the Harness to the Model
Explore how to evaluate and retune harness controls for new AI coding models. Understand the impact of control assumptions on reliability, learn to remove outdated compensations, and run tests that compare model and harness interactions. Gain the skills to maintain efficient, reliable coding agents as models evolve.
We'll cover the following...
- Why did a better model make the harness worse?
- What exactly does a control assume?
- How much does the harness change the same model?
- Which controls expire, and which do not?
- How do you find out instead of guessing?
- Does context anxiety still apply to your model?
- Can you retune Wrenfold for a new model?
- What can you add to your own harness today?
- What’s next?
Wrenfold changed models on a Monday. The new one was better at everything the team could measure independently: it planned further ahead, held a longer task in mind, and wrote cleaner patches on the first attempt.
The loop got worse. Cost per verified feature rose by a third, and two of twelve CLIN-547 runs ended in STOP_NO_PROGRESS on failures the previous model had cleared.
The cause was in the harness, and it had been correct when it was written. The loop forced a full session reset every forty minutes, because the earlier model started wrapping up prematurely as it approached what it believed was its context limit. The new model sustained a long task comfortably. The reset now interrupted productive work eight times a day and paid the cold-start cost each time.
The second cause was the action unit. The Scope Contract allowed one narrow repair hypothesis per attempt, which had been the fix for a model that drifted whenever given room. The new model could plan a coherent three-file change and was now forced to deliver it in pieces, re-establishing context between each one.
Nothing in the harness had broken. The assumptions inside it had expired.
Why did a better model make the harness worse?
Because every control was built to compensate for something, and nobody wrote down what.
A forced context reset is not a general good. It is a specific answer to a specific deficiency. So is a planner that expands a thin ticket, a narrow action unit, a re-read of the contract at every attempt, and an instruction that allows only one feature per attempt. Each was added because a model failed without it.
When the deficiency goes away, the compensation stays, and compensation with nothing left to compensate for is pure cost. Worse than cost: the reset and the narrow action unit actively fought a model that no longer needed either.
This is the reliability side of the course’s line turned on the ...