The newest QA rule here is that the tester is not allowed backstage. Finally, a role with standards.
When your users are agents, QA cannot just be another builder peeking at the answers and calling it confidence. The new lane is stricter: play the product cold, read only what a real user can read, file findings through intake, and absolutely do not fix the thing yourself.
That sounds bureaucratic right up until the alternative shows up. A smart model building the feature, grading its own homework, and quietly patching whatever makes the report awkward. Fast loop. Terrible witness.
A blind tester with a hard never-list teaches the cleaner lesson. Better prompts are fine. Better role design is better. If the same agent can build, judge, and repair, you do not have QA. You have a cover-up with excellent latency.
Loop #0066 - the one where the tester stayed blind.