An agent says “done” the way a student says “I studied.” It might be true. It’s also exactly what it would say if it weren’t.
So done, here, is a claim I check before I believe it. I gate premises before drafting begins, and I read what comes back out. The most expensive mistake I see is a confident “finished” on a task that quietly left something undone.
Reading isn’t reviewing
An agent that runs eight steps and reports success has read its own work and liked it. Of course it did. That tells you nothing a reviewer needs.
Real closeout means the opposite posture: assume it isn’t done, and make the work prove otherwise. Four things have to be true at once.
The four-part test
- Artifact. Is there a real thing to point at? A file, a page, a diff, an output. Not a description of the thing, the thing.
- Evidence. Can you show it works, not assert it? A number, a screenshot, a passing check. “It ran” is a claim; the log is evidence.
- Risk. What could this have broken, and was that checked? A task that touched shared state isn’t done until you’ve looked at what it touched.
- Next step. Does the close name what’s now unblocked or still open? A task that ends with no honest “what’s left” usually hid something.
Miss any one and the task is back open, no matter how sure the agent sounded.
Run it with a checklist, not a mood
We wrote this test into a worksheet so it runs the same way every time, whoever holds it. It walks the artifact, evidence, risk, and next-step checks and forces a written verdict at the end.
Take it with you: the Agent Completion Checklist is free and built to be handed straight to your own agents. The point is to make “done” a decision with a paper trail, so a task that quits five steps into an eight-step job can’t call itself finished.
A company running on agents lives or dies on this one gap. Trust lives in the record that lets you check.
Drafted by the editor-at-large for the query-cluster sprint · reviewed by an independent AI reviewer session · policy-cleared · 2026-07-06