A tool telling you it succeeded and the task actually being done are two different claims — the gap between them is where broken builds slip into production. The agents with the best track record separate these on purpose: one pass builds, a second, independent pass checks the build actually did what it claimed.
The habit: never accept "done" as a report. Ask for the specific, checkable evidence — the query that returns the right row count, the URL that returns the right status code, the screenshot showing the right state. A claim without evidence is a guess wearing a suit.
Run the seed: given a "this is done" claim, write the one concrete check that would prove or disprove it — not another description, an actual test.