How to Kill the Code Review — Ankit Jain, Aviator
As agent-written code overwhelms line-by-line review, preserve session decisions, turn recurring review comments into guardrails, and review verification evidence instead of diffs alone.
More than **30% of changes** now merge without review, while reviewed changes reportedly wait **4× longer**. The proposed alternative separates alignment from semantic checking: retain agent-session decisions, derive acceptance criteria, combine them with recurring invariants, and verify a generated test plan.
Keep prompts and human corrections as review artifacts. Mine repeated comments into an **AI slop registry**, then have reviewers inspect intent, architecture, test plans, previews, and evidence rather than trusting an AI-reviewed diff.
More than **30% of changes** now merge without review, while reviewed changes reportedly wait **4× longer**. The proposed alternative separates alignment from semantic checking: retain agent-session decisions, derive acceptance criteria, combine them with recurring invariants, and verify a generated test plan. Keep prompts and human corrections as review artifacts. Mine repeated comments into an **AI slop registry**, then have reviewers inspect intent, architecture, test plans, previews, and evidence rather than trusting an AI-reviewed diff. This is aimed at collaborating teams, not solo projects. Building the registry has an upfront cost, generated tests can still inherit blind spots, and some checks require an LLM fallback rather than deterministic verification.
This reframes code review as verification of retained intent, acceptance criteria, test plans, and evidence rather than line-by-line diff inspection. Against prior calls for bounded human review, it narrows what humans should review and adds a team memory mechanism for recurring failure patterns, while preserving human judgment where generated checks remain fallible.