AI coding agents generate more code, but not more software
A firm-level study finds that coding agents raised code volume while leaving measured software delivery unchanged, with longer review cycles absorbing much of the gain.
TL;DR
- Agents raised code lines 30%, commits 20%, and pull requests 23%.
- Completed issues and epics did not significantly increase after adoption.
- Longer review and more revisions may absorb generation gains.
Researchers Fiona Chen and James Stratton analyzed engineering activity across more than 700,000 employees at over 700 software firms through March 2026. Their study links AI coding agents to more code production, but not to a statistically significant increase in completed software issues or epics. [1] [2]
The analysis reports average increases of 30% in lines of code, 20% in commits, and 23% in pull requests after agents were introduced. The authors also observed longer review processes and more revision requests, suggesting that speed at code generation can move work downstream rather than remove it. [1] [2]
Ars Technica reports that the study uses aggregated Jellyfish engineering data and a difference-in-differences design comparing changes before and after adoption. This is observational firm-level evidence; it does not establish that every tool or team experiences the same result. [1] [2]
The open question is whether firms can improve review capacity and integration practices enough to turn additional code into shipped functionality. The study covers activity only through March 2026, so later changes in agent quality and workplace practices are outside its observation window. [1] [2]
Why it matters
The study finds that AI agents can increase code output without raising measured delivery, shifting productivity evaluation toward finished work and review costs.
Editor's note
Ars Technica headline reproduced verbatim; findings are observational, not a universal causal claim.