Paper argues AI coding harnesses need contracts, not autonomy

A paper posted July 10 rejects the idea that agent harnesses should rewrite themselves, arguing for a fixed contract instead, per arXiv 2607.08028.

The failure mode it targets is common: a prototype whose entire behavior lives in prompts and retrieval context. For teams shipping these systems, that's the layer most likely to break when nothing else in the code changed.

The fix, per the paper, is to move deterministic behavior out of the prompt. It goes into code, manifests, schemas, and validation artifacts, arranged around a model you can swap out. The model changes. The contract doesn't.

TTHE, published the same day, argues close to the opposite: the harness should rewrite itself as it runs. The two papers disagree over which layer of the system is allowed to move.

My read: they're both half right, and most builders split it the wrong way. Verification, whether the output is actually correct, should be the static, hard-coded part. Strategy, what to try next, should be the part that flexes. Wire it backwards, rigid strategy with soft verification, and you're back to the fragile prototype this paper is arguing against.