The Harvard Study Says AI Wrote More Code and Shipped Nothing. Read It Twice.
More lines, more pull requests, same shipped software. The bottleneck moved to review.
Researchers at Harvard looked at data from over 700,000 workers across 718 firms and found that fully agentic coding tools raised lines of code by 30 percent and pull requests by 23 percent. Both numbers were statistically significant.
The uncomfortable part: they found no measurable change in resolved issues or epics, their stand-in for software that actually ships. The gain went into volume, not outcomes.
They traced it to review. Time to review a pull request rose 49 percent after teams adopted agents, and comments per request rose 35 percent. The agents wrote faster than humans could check.
The lesson isn't 'stop using agents.' It's 'hire reviewers, or the code piles up unread.' We optimized the writer and forgot the editor.