Agent completion verification compares the intended outcome with evidence from the affected environment. Evidence may include tests, exact file contents, measured outputs, screenshots or service responses.
Verification should be part of the build process and independent of the agent's final statement. High-impact tasks require stronger evidence and human judgment than low-risk, easily reversible work.
ELI5
Agent completion verification checks real evidence that an AI agent finished the requested job instead of accepting the agent's own claim. The evidence can include tests, records, required files, constraints, or a review rubric tied to the requested outcome.
For example, a coding task is complete only when the expected behavior and tests pass without changing the tests to hide a defect. Less objective work may need a defined checklist, sampling, and human judgment.
Why is an agent's final message not enough?
The agent can misunderstand the goal, miss an error or confuse an intermediate step with the required outcome.
What counts as completion evidence?
Relevant tests, inspected files, database receipts, service responses and direct checks of the affected surface can all count.
