Harvard study finds AI coding agents generate more code but not more software
Ars Technica AI · rss · 2026-10-10
Harvard researchers Fiona Chen and James Stratton analyzed Jellyfish data spanning 300M+ work events, 700+ firms and 700,000+ employees (2021–March 2026):
- AI coding assistants generate more code, but human code review becomes a significant bottleneck that absorbs the efficiency gains from the coding phase.
- With AI tools, review processes lengthen significantly, pull requests more often require revisions, and reviewers leave more comments.
- The study finds little evidence firms increase software output or reduce employment—more code, but not more software.
More from AGI Musings
- a16z charts: call center jobs flipped from +4% annual growth to -4% decline — a16z · 2026-10-10
- Mathematician littmath: theorem-proving is just an easy-to-measure proxy, and AI solving problems isn't the real threat — littmath · 2026-10-10
- Economist: most published math papers are right, econ papers are wrong — and that's why AI lags — paulnovosad · 2026-10-10
- Freeswarms by 2027: AI bot swarms that earn money, buy compute and code themselves — max_paperclips · 2026-10-10
- Mike Frank: AI welfare talk is regulatory capture, not genuine care for models — MikePFrank · 2026-10-10
- The Most Viral Videos on Social Are Now AI-Made — And Parents Aren't the Ones Sharing — bennash · 2026-10-10