AI coding agents: 30% more code, no more shipped work

AI coding agents made engineers write about 30% more code. The firms did not ship measurably more. That is the core finding of a Harvard working paper by Fiona Chen and James Stratton, built on Jellyfish engineering data from 718 companies. The extra code piled up at code review. If you pay for agent seats and expect features to arrive faster, this is the study to read before you buy more.

What actually happened

The paper, Artificial Intelligence in the Firm: Bottlenecks in Software Production, covers about 300 million work events from 700,000 workers, January 2021 to March 2026. It uses GitHub activity, Jira issues, calendar data and usage logs for GitHub Copilot, Cursor and Claude Code. The authors compare firms that adopted AI tools earlier and later (a staggered difference-in-differences design). It is observational, not a randomized trial.

What the authors found after firms adopted coding agents:

• Code output rose. Lines of code went up 30%, commits 20% and pull requests 23%.
• Shipped work did not rise measurably. Resolved Jira issues and epics showed small positive effects that were not statistically significant. For agents, the estimates rule out an output gain larger than 12% of the baseline.
• Review became the bottleneck. Average time to review a pull request rose 49%. Comments per pull request rose 35%. The share of pull requests with changes requested nearly doubled. The share of workers doing reviews rose 14%.
• AI review did not fix it. The bottleneck stayed after firms adopted AI code review tools. Humans still did most of the reviewing.
• No measurable job cuts. The authors found no significant employment effect from agents.

A note on the headlines: "zero extra issues resolved" overstates it. The estimate was positive but too noisy to separate from zero.

Why it matters for your business

Writing code is roughly 30–40% of an engineer's time, per the paper. Speed up only that part and the work waits at the next step. For a small team, the next step is usually one senior person who reviews everything. Give three developers agents and that person becomes the release schedule.

What we do about it:

• Measure merged and deployed work, not lines or pull requests. Agent dashboards that count lines measure the wrong thing.
• Make agent pull requests smaller. One change per pull request, with tests attached, so a review takes minutes.
• Budget review time when you add agent seats. Put a person's hours on it, not a hope.
• Set who pays for AI review. GitHub added organization billing controls for Copilot code review on Oct. 8: bill reviews to the org with a budget, and limit who can trigger them.

Sources: Chen & Stratton, Harvard (working paper), GitHub Changelog.