What it was for
Review was the quality gate human-written code passed on its way to trusted: catching defects, enforcing standards, and — its quiet second function — teaching, as senior judgment left comments on junior work.
The verdict
TRANSFORMS. The volume produced by the inner loop makes line-by-line human inspection arithmetically impossible; a team that insists on it becomes a human gate bolted to a machine shaft. But waving everything through ships confidently machined defects. Review survives by moving up a level: from reading the code to verifying the intent.
What changes
The reviewer's questions change from "is this line right?" to "is this change what we meant, does it fit what exists, did it touch what it must not?" Mechanical concerns — style, convention, obvious defects — are delegated to automated checks and the machines themselves. Sampling becomes a legitimate discipline: deep-reading in proportion to consequence, and being honest about the difference between verified and glanced at. The teaching function migrates too — into evals, architectural guardrails, and the written intent that machines and juniors both learn from.
The strongest objection
Reading code is how engineers stay sharp; review-as-verification atrophies the muscle. Partially conceded — which is why deliberate deep reads of consequential changes stay in the practice. The skill does not dissolve. It gets rationed toward where it matters.
Corroboration (Jul 2026). The Thoughtworks retreat described review being "unbundled" into its four functions — mentorship, consistency, correctness, trust — each needing a new home, which is this entry's migration of the teaching function into evals and guardrails. It also framed the move as risk tiering: verification proportional to blast radius. That sharpens "deep-reading in proportion to consequence" into a rule — match review investment to what breaks if the code is wrong, not to line count. (Thoughtworks retreat)