Reading up on code-review

Developers increasingly ignore AI-generated pull requests because the effort to review sloppy code doesn't feel reciprocated.

Doug Slater argues developers should match code-review rigor to risk, shipping unreviewed LLM code is acceptable only for limited-distribution, low-consequence software.

Human reviewers should focus on offering out-of-distribution knowledge (context, design decisions) rather than line-by-line checks, which AI can handle.

Software development remains a multiplayer game where code review, not design docs, is the primary artifact for team judgment as AI reduces code production costs.

The piece defines a structured pre-commit code review skill for AI agents, specifying contracts, scope governance, and engine isolation across multiple review engines including Codex, Claude, and others.

Using multiple LLMs for rigorous code review and bug finding can produce higher-quality code than the common fast-output approach, though it slows development velocity.

Stacked commits, not monolithic PRs, are the correct unit of code review—a truth Gerrit knew in 2008 that AI's speed and tooling like Mergify Stack are now forcing GitHub to finally admit with its native stacked PRs feature.

Separating mechanical from non-mechanical changes across intentionally broken commits, then squash-merging them, makes code review easier without cluttering history.

Excessive reliance on AI coding tools creates comprehension debt—a growing gap between code volume and human understanding—that erodes system knowledge and eventually must be paid with interest.

Vercel Labs released OpenReview, an open-source, self-hosted AI code review bot that deploys to Vercel and uses Claude to provide on-demand PR reviews with inline suggestions and automated fixes.

Anthropic's Code Review service uses multi-agent analysis to catch logic errors, security flaws, and regressions in GitHub pull requests, posting findings as inline comments without blocking merges.

Human code review is becoming obsolete at AI-generated code scale; teams should shift to upstream spec review, adversarial agents, and deterministic verification.

Stacked pull requests, managed with tools like Graphite, keep code review fast and team velocity high by breaking large work into small, independent, mergeable units.

The author replaced manual code review with 13 parallel AI agents, catching a bug in 15 minutes that a traditional review would have missed.

Kevin Murphy shares a systematic PR-reading method using the five Ws, including test-first review, to improve code review quality.

Cursor improved Bugbot's bug-finding rate by defining a custom AI-driven resolution rate metric that enabled systematic hill-climbing experiments, raising resolution from 52% to over 70% and doubling resolved bugs per PR to roughly 0.5.

Coding agents make implementation cheap but shift engineering difficulty to defining context, constraints, and verification, making PR descriptions the new center of work.

Code review catches mostly style issues, not production bugs, costing a typical 80-engineer firm $3.6M annually while preventing few real incidents.

Engineers must shift their role from generating code to proving code works, especially with LLMs, by providing tests and verification as part of contributions.

Software engineers must prove code works through manual and automated testing, shifting accountability away from AI agents to the human submitting the PR.

Review.Fast uses AI to generate concise “review stories” for GitHub PRs, aiming to reduce human review time by 3x.