Why coding agents need an executable definition of done backed by independent evidence

AI coding agents can modify the right-looking files and report success while the task remains incomplete, since tests may not run or fixes may only work in narrow conditions. The article argues the remedy is not a longer prompt but an executable contract for each task that specifies the allowed changes, the expected observable outcome, the required evidence, and an independent process for acceptance. It separates implementation, where the agent proposes a change, from acceptance, where separate checks confirm the outcome, mirroring how pull requests are verified and reviewed before merging. The piece recommends criteria that are observable, bounded in scope, and independently checkable, and it notes that model explanations or unverified screenshots are not sufficient proof. Each criterion should produce a stored receipt, which makes handoffs reviewable and supports resuming failed runs and investigating regressions.