Harvard study: AI coding agents add 30% code, flat shipped output

A Harvard working paper covering roughly 300 million work events from more than 700 software firms finds that AI coding agents lift per-worker code volume by about 30%, while the amount of finished software shipped at the firm level does not move, according to Ars Technica. Economists Fiona Chen and James Stratton paired engineering telemetry from Jellyfish with GitHub activity signals across more than 700,000 employees between 2021 and March 2026.

After firms adopted tools in the Claude Code, Cursor and Devin class, each worker added nearly 1,500 extra lines of code per month, 20% more commits, and 23% more pull requests. Resolution of Jira issues and epics, the paper's proxy for finished software, showed no statistically significant change.

The constraint sits downstream. The paper reports that "the code review process significantly increases in length, pull requests are more likely to require revisions, and reviewers leave more comments." Average time between a pull request being submitted and merged rose 49%, comments per pull request climbed 35%, and the share of pull requests with changes requested nearly doubled.

AI reviewers have not closed the gap. By March 2026, 80% of measured firms ran some form of AI code review, yet AI-generated comments accounted for only 23.3% of all review comments and appeared on just 10.8% of pull requests. Chen and Stratton conclude that "Realizing the benefits of AI adoption requires complementary investments in downstream review capacity, rather than simply increasing code output." Two of the analysts we follow surfaced the paper in their feeds.