Why code review must move from diffs to specifications in the age of AI agents

Short summary

Code review was always a compromise: reviewing diffs because intent lived in unwritten specifications. When humans wrote code, skilled reviewers could reconstruct intent from implementation choices. AI-generated diffs break this model — the diff is a rendering of constraints, not reasoning, so there is nothing to infer. Review must move upstream to specifications, scenarios, and domain models where intent actually resides.

• •Code review on diffs was a compromise for verifying intent that lived upstream in specs
• •Human-authored diffs carry reasoning fingerprints; agent-authored diffs only render constraints
• •Review must shift from line-level diff inspection to specification and scenario verification

Generated with AI, which can make mistakes.