Your AI Agent Says “Done.” What Did It Actually Verify?

Описание и применение

Проверка завершения работы AI‑агента: как убедиться, что исправление действительно исправило баг и не нарушило другие функции.

Задача

When an AI agent reports completion, it is unclear whether the intended behavior was truly achieved and verified.

Результат

The workflow ensures that an AI agent’s claim of completion is backed by concrete, verifiable evidence, reducing the risk of merging incomplete or incorrect changes.

Рабочий процесс

• Define observable "done" criteria before the agent starts.
• Ask the agent to include evidence in its completion report.
• Ensure the test catches the original bug by verifying expected outcomes.
• Inspect any changes made to the tests themselves.
• Match verification methods to the type of change (calculation, API, UI, etc.).
• Check for unintended side effects or scope creep in the diff.
• Treat blocked verification as unfinished and document missing checks.
• Use a reusable prompt to standardize reporting and verification.
• Ask for evidence that the change meets the requirement before merging.

Доказательства

"Fixed the bug. All tests pass."

Источник

"Added a regression test for excessive discounts. The checkout test suite passed. Browser checkout was not tested because payment credentials were unavailable."

Источник

When a discount exceeds the order subtotal, the total should be zero. Add a regression test. Preserve valid discount behavior.

Источник