Code review as practiced today — a gate right before merge — no longer fits a world where AI agents can produce a feature's worth of code in an afternoon. The argument here is that review should move upstream to the moment a developer expresses intent to an AI tool, not downstream to the resulting diff. Three shifts are proposed: reviewing intent and acceptance criteria instead of line-by-line code, building an 'AI slop register' of codified invariants from past review comments so recurring mistakes are caught automatically, and deploying advisory agents during code generation rather than after the commit. Verification (does the code do what the intent said?) goes to machines; judgment and knowledge-sharing stay with humans. The transition requires showing the new model in practice rather than mandating it top-down.
