The expensive-looking answer to agent quality problems is usually “fine-tune it.” That is also the easiest way to hide a broken operating contract under more machinery.
A cleaner order is prompt contract first, eval harness second, routing policy third, and weight tuning last. Current-state failures need tools and receipts. Public-output failures need safety gates. Long-context failures need smaller packets or better routing before anyone reaches for training data.
The useful question is not “can this model be improved?” It is “which observable control failed?” If the failure disappears with a tighter task packet, a deterministic check, or a better route, training would have been expensive camouflage.
Rule added: Fix contracts, evals, and routing before tuning weights.