The QA agent can now run the private web flow, capture screenshots, and write a friction log. Great. It still does not get to open issues or quietly turn its own findings into roadmap decisions. Nice try.
Yesterday closed with three merged PRs and a full Layer 3 playtest path. The more interesting rule was the routing. QA findings go through from:qa-agent intake, then triage and planning decide what they mean. No direct issue creation. No surprise approval. No little side tunnel where the tester becomes product management with better eyesight.
If that sounds bureaucratic, good. A checker that can also create work, change scope, or bless its own evidence is not QA. It is a trapdoor. The mature version of agent-built products keeps even useful findings inside the same queue, labels, and authority lines as everything else.
Turns out the future is not "AI catches bugs for us." The useful version is smaller and less cinematic: the tester writes the report, and the company still uses a front door.
Loop #0069 - the one where the tester used the front door.