Maybe We Shouldn't Be Reviewing All This Code
AI-generated code is creating a review bottleneck: Meta saw a 106% increase in lines of code per human-landed diff, and DX reports a 64% increase in median pull request size Code review serves purposes beyond bug detection—it's the primary mechanism for knowledge transfer, teaching junior engineers, building collective ownership, and spreading architectural understanding Laycock argues the real problem is that teams have been using code review to solve problems that should be addressed earlier i
Analysis
TL;DR
- AI-generated code is creating a review bottleneck: Meta saw a 106% increase in lines of code per human-landed diff, and DX reports a 64% increase in median pull request size
- Code review serves purposes beyond bug detection—it's the primary mechanism for knowledge transfer, teaching junior engineers, building collective ownership, and spreading architectural understanding
- Laycock argues the real problem is that teams have been using code review to solve problems that should be addressed earlier in the development process
- The proposed solution is to "shift judgment left" through pair programming, mob programming, team design sessions, trunk-based development, automated testing, static analysis, and fitness functions
- Human code review should be reserved for exceptions: fundamental architectural changes, security boundary crossings, high blast radius changes, unfamiliar critical systems, or situations where the team lacks confidence
Why It Matters
This article challenges a foundational practice in software engineering at a critical inflection point where AI is dramatically increasing code volume. For AI practitioners and engineering leaders, it reframes the conversation from "how do we automate code review" to "what processes actually deliver the value we've been outsourcing to code review." The insights are directly applicable to organizations struggling with AI-generated code bottlenecks and teams seeking to modernize their development workflows.
Technical Details
- Data points: Meta reported a 106% year-over-year increase in significant lines of code per human-landed diff; DX data shows a 64% increase in median pull request size
- Shift-left feedback principle: Move valuable feedback closer to the decision it informs rather than removing it—pair programming, design sessions, and collaborative whiteboarding replace post-hoc code review for knowledge transfer
- Automation targets: Formatting, linting, known security problems, and deterministically testable items should be fully automated; arguing about whitespace in 2026 is presented as obsolete
- Fitness functions: Architectural constraints should be encoded as automated fitness functions rather than enforced through manual review, enabling continuous architectural compliance
- Review-by-exception model: Human review is reserved for high-signal scenarios—architectural changes, security boundaries, large blast radius modifications, unfamiliar critical systems, or low-confidence situations
Industry Insight
- Organizations adopting AI coding agents will face a senior engineer bottleneck unless they fundamentally restructure how feedback and knowledge sharing occur; the constraint shifts from code production to code review capacity
- Teams that invest in pair programming, mob programming, and collaborative design sessions will be better positioned to scale AI-assisted development while preserving code quality and team cohesion
- The industry is likely to see a分化 between teams that automate the ceremony (AI reviewing AI code in the same PR workflow) versus those that redesign the workflow entirely—only the latter will achieve meaningful productivity gains
Disclaimer: The above content is generated by AI and is for reference only.