How the Claude Code team reviews agent-written PRs
At a glance
Anthropic’s Claude Code team documents a layered review stack—pre-PR verify, AI review + security block, risk-tiered human approvals, merge queue/flags, and fast rollback—so most agent PRs merge with ≤1 human while revert rate fell >40% as volume doubled. This workflow is listed under code-review, coding-agents, human-in-the-loop, devops, learnable-paradigm.
Agents / tools
Claude Code, Claude Code Review, REVIEW.md, CLAUDE.md, GitHub
AWS Labs open-source (5k+★): one harness-neutral core turns Claude Code/Codex/Cursor/Copilot/Kiro/opencode into structured, verifiable delivery workflows via `aidlc config --harness` + doctor—install script, not a new agent.
Empirical study: 34% of popular active repos ship agent instruction files (AGENTS.md dominant); half exceed ~100 lines; ~3/5 rules could be script-checked but <1/20 name the check—actionable lint: short map, name enforcers, one rulebook.
MIT Go daemon + git hook reviews every commit with the CLIs you already pay for—Codex, Claude Code, Cursor, OpenCode, and first-class Grok Build (grok / grok-build). Findings stay in a ledger until closed; agent hooks push open review IDs back into the session for /roborev-fix before context goes cold. Reviews stay on your machine.
Native macOS app orchestrates official Claude Code, Codex and Grok CLIs as a role-based crew (lead/builder/reviewer/docs/auditor/on-call) on your Mac or VPS with schedules, one-click risky-command approvals (Telegram too), and server screen/browser/terminals—no Bandito cloud for your code.