“Done” needs more than an agent’s confirmation.
A useful completion claim includes evidence that the intended behavior works.
What you might notice
An agent reports success because files changed or a build passed. Neither proves that the user can complete the requested task.
An example
In a fictional app, the password reset page renders correctly and the build succeeds. The email callback still points to a development address. The feature is marked complete before an end-to-end check.
Why it matters
Features appear finished in the backlog while the customer-facing behavior remains incomplete. The missing context carries into the next development session.
How we check
We compare acceptance criteria with the checks actually performed, including environment configuration and integrations. We record which conclusions are verified and which remain assumptions.
What a fix should achieve
Specify the expected user outcome before implementation. Require a short completion record: what changed, what was checked, evidence, and known limitations. Use checks appropriate to the feature rather than test counts.