Thanks, this is an important point. An agent saying “Done” isn’t proof that anything happened.

I agree that defining postconditions per action is the right safeguard. A few ways I’d connect it to the platform:

In the instructions: make the completion condition explicit, and tell the agent to re-read the record after a write (state, assignment, approval) before it confirms success to the user. In testing: agentic evaluations score task completeness and tool calling, not just the final message. That helps catch runs that “sound” complete but aren’t. After go-live: check the execution plan and its tasks (sn_aia_execution_plan / sn_aia_execution_task) against the actual record changes. Any gap between what the agent reported and what the records show is your failure signal.

Supervised mode on write actions adds another check, because a human sees the proposed change before it runs. I may cover postcondition checks as their own post. Appreciate you adding this!