AI agent glossary
Definition of done
A concrete, checkable statement of what must be true for a task to count as finished, which tells an agent when to stop and you how to verify.
Agents stop when they believe they are finished. Without a definition of done, that belief is a guess, and you get work that is half complete or wanders far past what you asked for.
A good definition of done is something you or the agent can check: the tests in a given file pass, the type checker reports no errors, the page renders at phone width, the report cites a source for every number. It sits alongside the goal, the context, and the constraints as one of the parts of a task an agent can finish.
The stronger the check, the more autonomy you can safely give. Tests make an excellent definition of done because the agent can run them itself and you can rerun them independently.
Instead of 'fix the date bug', the task ends with: done when the new test for leap-year invoices passes, all existing tests still pass, and no files outside src/billing changed.
Learn more in the course
Related terms
Go from definitions to doing
Learn to use and build AI agents in 130 practical lessons.