Under agent PR volume, teams are redesigning review — not abandoning it
Gergely Orosz surveys how labs and startups actually handle the review flood — from reviewing the AI review, to blast-radius triage, to reviewing plans and schemas instead of diffs.
Source: What is happening with code reviews? — The Pragmatic Engineer
GitHub’s own chart is the setup: PRs up roughly fivefold over three years, with another near-doubling since late 2025. Orosz’s survey of CTOs and eng leads is useful because it refuses a single slogan. Most shops are not dropping human review; they are relocating scarce attention.
The patterns that keep showing up: humans review the AI review (sometimes after multi-vendor or multi-agent passes, with noise still the failure mode — Uber’s uReview grades and filters comments before a person sees them); triage by blast radius so low-risk paths can clear on automated gates while auth, schema, and public API stay human; review the plan, tests, or database schema harder and the implementation less; and force smaller PRs so review stays tractable. Duckbill’s risk-label experiment is the receipt for the triage move — roughly double merge throughput once most PRs skipped a human pass behind stronger tests and lint.
Why it matters for code review
The durable question is not “AI review or human review.” It is which gates earn a person, which earn a heterogeneous agent stack, and how you keep silence and risk labels honest when generation got cheap and understanding did not.