Mobile apps can break in ways that code review and green CI checks never catch — navigation loops, permission prompts blocking first run, release-build crashes absent in debug. The post argues for a third review layer: a bounded cloud QA agent that runs against the actual app on a simulator or real device before merge. The proposed stack uses an agent runtime (Eve), a device-control CLI (agent-device), and a CI workflow that boots the device, runs the targeted flow, collects screenshots and logs, and posts a structured report back to the pull request. Key design principles include giving the agent a narrow, scoped mission rather than full-app exploration, using deterministic setup for repeatability, choosing model size based on task complexity, and exposing real devices via a proxy for cases where simulators fall short.
