AI Engineering
An agent’s final message is not proof of completion
Verify the state an agent leaves behind before treating its final message as evidence that the requested work is complete.
Note1 min read
An agent saying it saved a change is a claim to verify. For a settings task, I’d read the saved value back and reload the screen. A successful tool call and a convincing summary can still leave the user’s problem unresolved. Make the resulting application state part of the acceptance check.