AI Coding Agents Boost Code 30%, but Software Output Stalls
Harvard study finds AI coding agents lift code volume 30% but not software output, as review time rises 49% and employment remains unchanged across firms.
Summary
Harvard University researchers Fiona Chen and James Stratton analyzed 300 million Jellyfish work events, including commits, pull requests and issue management data, from more than 700,000 employees at over 700 software firms between 2021 and March 2026. Using measured AI adoption, GitHub activity and a difference in differences model, they compared workflows before and after firms introduced AI assistants, which complete human-written code, and agents, which autonomously write and submit code.
AI coding agents increased total lines of code 30%, commits 20% and pull requests 23%. However, resolution rates for Jira Issues and Epics showed no statistically significant change, nor did their size or complexity. Average time from pull request submission to merger rose 49%, the share requiring changes nearly doubled, comments per request increased 35%, and the share of workers reviewing code climbed 14%. Jellyfish and LinkedIn data showed no significant employment change attributable to AI.
By the March 2026 cutoff, 95% of studied firms had implemented coding agents and 80% used some AI code review. Yet AI generated only 23.3% of review comments and 10.8% of pull requests, leaving humans responsible for most review work. Agent capabilities have improved since the cutoff, and firms are still learning where to deploy them, but the measured productivity gain was absorbed downstream rather than converted into more completed software.
Positives
- AI coding agents increased total lines of code by 30% after adoption.
- Commits rose 20% and pull requests increased 23%, demonstrating faster code production.
- 95% of studied firms had implemented AI coding agents by the March 2026 cutoff.
- 80% of measured firms used some form of AI code review by March 2026.
- Newer agent upgrades and growing deployment experience could improve the balance between coding and review time.
Risks & concerns
- Jira Issue and Epic resolution rates showed no statistically significant improvement after AI adoption.
- Average pull request review time increased 49%, absorbing gains from faster code generation.
- Pull requests requiring changes nearly doubled, while comments per request rose 35%.
- The share of workers performing code reviews increased 14% after agents were introduced.
- AI produced only 23.3% of review comments, leaving humans responsible for most review work.
- Jellyfish and LinkedIn data showed no significant employment reduction attributable to AI.