The AI agent says done, but the deliverable is missing: verify outcomes

An agent reports that files were pushed, a ticket was closed or a customer update was sent. The destination contains nothing. A recent practitioner account describes teams manually checking and closing agent work after a “done” message. Its numbers are one team's account; the transferable lesson is to verify the result where it should exist.

Define the receipt before the agent starts

For every important task, specify an observable end state: repository commit at the expected remote ref, document ID and status, sent-message ID, test result, or customer-visible page. The receipt should identify the target and time, not just say that an API request returned success. Where a workflow has several steps, name which ones are required for completion and which are merely attempted.

Separate process from outcome

Microsoft Foundry's agent evaluators distinguish tool-call accuracy from end-to-end task completion. Use both lenses. A correct tool call may still leave work unfinished because the external system rejected a later step, the target changed, or the deliverable was not readable. Conversely, an answer can look polished while no action happened.

Build a small acceptance check

After an agent claims completion, read the destination system and compare its state with the user's request. For publishing, check the public page; for a report, open the artifact; for a data update, query the affected record. Return “needs attention” with the missing receipt instead of silently re-routing or making the agent repeat everything. Sample human review can then focus on real decisions and exceptions.

Adovo’s view

Completion is a property of the world, not of the agent's sentence. Put the proof at the destination and let the agent report it. That turns confident narration into accountable work.