Harvard Study: AI Coding Agents Lift Code 30%, Shipped Output Flat

AI coding agents lift per-worker code volume by about 30%, yet shipped software output at the firm level does not move. The finding comes from a working paper by Harvard economists Fiona Chen and James Stratton, who paired engineering telemetry from Jellyfish with GitHub activity signals across more than 700 software companies, reported by Ars Technica.

Chen and Stratton drew on roughly 300 million work events from 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. The resolution rate for Jira issues and completed epics, the units the authors use as a proxy for shipped features, did not move in a statistically significant way, and firm-level employment did not fall either.

Instead, the gains moved downstream. Average time between a pull request being submitted and merged rose 49%, revision requests nearly doubled, and comments per pull request climbed 35%. The share of workers assigned to code review rose 14%. By March 2026, 80% of firms were running AI code review tools, but only 23% of review comments were AI-generated.

"Both technologies increase coding productivity," Chen and Stratton write. "However, productivity gains do not fully pass through to changes in software output or employment." Their prescription is just as blunt: "Realizing the benefits of AI adoption requires complementary investments in downstream review capacity, rather than simply increasing code output."

Engineers in the paper's qualitative interviews said the same thing in plainer words. "People could make a huge number of commits quickly, but code review was still the bottleneck," one junior engineer told the researchers. The result lands into a run of adoption stories we've been tracking, including Epic Games' AI-first mandate earlier this week, with more agent seats going out across the industry and, so far, no firm-level shipped-feature dividend showing up in the data.