Code review (as we know it) needs to die
a day ago
- AI-driven development is breaking traditional code review: faster code generation caps velocity at slow human reviews.
- Asynch diff-centered code review, ~20 years old, fails in an era where writing code is cheap but reviewing isn't.
- Diffs lack context, requiring reviewers to reconstruct intent and design, adding expensive overhead.
- All changes get equal scrutiny regardless of risk, wasting attention on low-risk changes and under-scrutinizing risky ones.
- Code review's core jobs—defect detection, maintainability, knowledge sharing, control, and norm enforcement—remain vital but must be unbundled.
- Defect detection shifts to automated checks, independent AI review, and human verification of product, not just code.
- Knowledge sharing moves up: review focuses on intent and design, while code comprehension becomes pull-based via AI.
- Norm enforcement moves into specs and automated tools, not comments on diffs.
- Control stays as merge gates: humans approve changes based on higher-level review, trusting agents and automated verification.
- New process: align design before code, surface changes with context, run adversarial AI review, show evidence, route by risk, and absorb mistakes through system safeguards.
- Examples like Linear Diffs, Meta RADAR, and Uber's Code Inbox show components of this new approach in practice.
- Costs include less vetted individual changes and shipping some bugs, but winners accept this and build resilient systems.
- Ultimately, redesigning review around AI accelerates velocity and lets engineers focus on creative, high-leverage work.