Writing a Definition of Done for AI Agents
Turn a subjective request into observable acceptance criteria.
Name the artifact and its properties
For a website feature: the new interaction exists, the supplied acceptance criteria are met, existing affected flows work, and no misleading controls remain. For research: the report cites material claims, labels assumptions, and reaches a decision with falsifiers.
Choose observable checks
“The result is excellent” is subjective. “The form accepts valid input, rejects invalid input with a clear message, and preserves the saved draft after reload” can be checked. Use checks that reflect user needs.
Ask for evidence
The final response should connect validation to the acceptance criteria. A successful build alone does not show that every interaction works. If validation was unavailable, the agent should say what remains unverified.
Include the final handoff
Specify the useful report: what changed, evidence of validation, remaining limitations, and the next decision. Completion can mean “ready for approval” when publishing or another external action is gated.