Harvard AI Coding Agents Review Bottleneck

AI coding agents made engineering teams write 30% more code. They did not ship more software.

A Harvard study with Jellyfish analyzed 300 million work events from 700,000 employees at more than 700 software companies. After adopting AI coding agents, lines of code rose 30%, commits rose 20%, and pull requests rose 23%.

Delivered software output stayed flat. The researchers ruled out gains above 12%.

The reason is review. Review time per pull request rose 49%, revisions nearly doubled, and comments per pull request climbed 35%. The time saved writing code was absorbed downstream.

My read: AI moved the bottleneck, it did not remove it. Generating code is now cheap. Verifying it is not, and verification is still done by the same humans, in the same hours.

For any leader funding AI coding tools, the metric that matters is no longer lines written. It is features delivered per quarter, and the cost of the review queue that sits between the two.

Where is the bottleneck in your engineering organization today: writing code, or reviewing it?