Yegge Predicts Code Review's End, HN Pushes Back

Steve Yegge argues code review is nearly dead in an essay that hit Hacker News' front page on August 3.

His piece, "The Shape of Things to Come," describes agents landing more than 175 commits a day. They drain a merge queue that reached 166 commits deep, a pattern he calls the Land Rush. He predicts CI/CD collapses under that rate, and human review survives only as vestigial SOC 2 compliance theater. Its replacement is a "Wish Factory," where agents implement issues straight from user reports.

Boris Cherny, who created Claude Code, backed the volume claim with a number. 85% of his engineers run dozens or hundreds of agents at once, using what he calls graph engineering, per a post on X.

The pushback

Three other posts from the same week argue the opposite, and they outdrew Yegge on Hacker News.

Niklas Gruhn's "Don't Be a Meat Proxy" topped the site at roughly 1,750 points. He tells engineers to read and validate AI output instead of relaying it untouched.

Ankur Sethi, at 503 points, goes further. He never lets an assistant write to his repo and retypes every generated line by hand, because typing it himself is how he catches hallucinations.

Sean Goedecke's "LLMs Reward Expertise" pulled 1,042 points arguing the deciding skill in prompting is domain knowledge. His example is Terence Tao steering ChatGPT through a Jacobian Conjecture counterexample with precise questions.

Yegge routes around the review bottleneck with automation. Gruhn defends it with discipline. Neither one addresses that discipline scales with the number of humans, while automation scales with the number of agents.

I run agents on my own projects every day, and my bet is Yegge's right about the volume and wrong that review goes vestigial. Judging the output became my job. Writing it didn't.

Claude Code's Focus view, added in version 2.1.221, is a quiet admission of the same problem. It collapses tool activity into a summary instead of streaming every Read, Bash, and Edit call. The decision worth catching gets buried under hundreds of tool calls otherwise.