AI coding agents boost output 30% but slow code review by 49%

a computer screen with a bunch of code on it

More code does not mean more shipped software. A Harvard working paper by Ph.D. candidates Fiona Chen and James Stratton makes that case with hard numbers across a large sample of real organizations.

What the Research Found

The study covered 725,938 workers at 718 firms. Engineers at firms that adopted AI coding agents added roughly 30 percent more code. Finished work showed no measurable increase. Getting code through review took 49 percent longer.

The bottleneck did not disappear. It moved. AI agents generated more raw output than reviewers could absorb, and the checkpoint that had always existed became the constraint.

The Review Problem in Numbers

By March 2026, nearly 80 percent of the firms in the dataset had used an AI coding agent. Yet AI generated only 23.3 percent of review comments, and only 10.8 percent of pull requests received even one AI comment. The review function remained almost entirely human even as the volume of code requiring review grew sharply.

The Operator Takeaway

If you are deploying AI coding agents on a small team, the constraint you need to plan for is not generation speed. It is review capacity. More agent output with the same review bandwidth will slow your actual delivery rate, not accelerate it.

The paper also draws a parallel to legal AI tools. Harvey’s motion-to-dismiss workflow and a court-filing redaction agent Cooley is developing with Google Cloud both place a lawyer at a comparable review checkpoint. Neither has reported a measure of whether that step has slowed, according to the article.

Stay on top of AI & Automation with BizStack Newsletter
BizStack  —  Entrepreneur’s Business Stack
Logo