AI coding agents — Cursor, Claude Code, Codex — produce mobile UIs that break in consistent, predictable ways: viewport-snapping breakpoints, modals that trap background scroll, touch targets that are visually present but physically untappable, and features that appear in the diff without appearing in the prompt. Asking the agent to self-review before you merge is largely ineffective. This agent-agnostic, 8-point checklist gives you a QA layer to run before every mobile PR, catching the regressions your agent introduced silently. TL;DR: Run this checklist on every mobile PR that a coding agent touched. The eight checks cover viewport breakpoints, modal behavior, touch target sizing, silent feature additions, navigation regressions, text overflow, keyboard handling, and cross-device smoke testing. Total time: under 15 minutes per PR if you work from the diff. Why does asking the agent to review its own work fail? The honest framing first: agent self-review is a trap.…