A study by Harvard researchers Fiona Chen and James Stratton found that deploying AI coding tools increases code output without boosting total software delivery. Analyzing Jellyfish data covering 300 million work events across 700 software firms from 2021 through March 2026, the researchers evaluated the operational impact of AI assistants and autonomous coding agents.

The research revealed that introducing AI coding agents drove a 30 percent increase in generated lines of code, a 20 percent increase in commits, and a 23 percent increase in pull requests. However, resolution rates for tracking issues like Jira Epics showed no statistically significant improvement, as human review processes absorbed efficiency gains.

Code review times ballooned by 49 percent following the adoption of AI agents. Additionally, the share of pull requests requiring revisions nearly doubled, and comments per request rose by 35 percent, forcing firms to assign 14 percent more engineering staff to review duties.

Why it matters

  • Proves generated code volume does not equal enterprise productivity without redesigning downstream code review workflows.

  • Highlights human code review as the primary operational bottleneck for teams adopting autonomous coding agents.

  • Provides engineering leaders with empirical data to manage headcount, quality control, and pull request policies.

Source: arstechnica.com