The Problem
Standard release engineering assumes one PR per developer per day. Code review is the safety gate; humans read each diff carefully; deploy happens after the gate.
That model breaks at AI velocity. A vibe coder with Claude Code or Codex can author 30 PRs a week. A senior reviewer can read maybe three a day with care. The math doesn't work; either reviews become rubber-stamps, or velocity collapses, or something has to change about where safety lives.
Safety has to move from review-as-gate to release-engineering-as-gate. Feature flags become default. Rollback becomes a one-button operation. Environment promotion happens continuously, not on a release schedule. The reviewer's job shifts from "is this code correct" to "is this safe to expose to users."
The Core Insight
If a change can't be flag-gated, it shouldn't be merged.
That single rule is the foundation of release engineering at AI velocity. It means: any user-visible behavioral change is behind a flag; any new endpoint is behind a flag; any schema change is migrated in a way that's safe with the flag both on and off. The flag is the kill switch; the kill switch is what lets you merge fast.