Modern code review bundles five distinct functions into a single process: enforcing formalizablequality rules, detecting structural design problems, educating developers, evaluating designjudgments, and organizational gatekeeping. Drawing on fifty years of empirical research—fromFagan’s 1976 inspections through large-scale studies at Google and Microsoft—this articledemonstrates that each function has a more effective alternative: automated linting and AST-levelrewriting, periodic codebase-wide analysis, synchronous pair programming, architecture evaluationmethods, and lightweight CI-based governance. AI-powered code review tools accelerate thisundifferentiated process without decomposing it, concealing rather than resolving its structuraldysfunction. As AI coding agents enable parallel code generation across multiple worktrees,pull-request-based human review becomes not merely inefficient but physically impossible, makingthis decomposition structurally necessary. The five-layer framework proposed here offerspractitioners a concrete model for reassigning each function to its optimal mechanism.
Franny Philos Sophia (Fri,) studied this question.