Technology
AI coding agents increase code production without improving software output
A study of engineering activity across more than 700,000 employees at over 700 software firms found that coding agents sharply increased lines of code, commits, and pull requests. Yet firms saw no statistically significant improvement in resolving tracked software issues or features. The researchers argue that longer reviews, revisions, and reviewer comments absorb much of the initial speedup. The analysis uses aggregated company data and statistical comparisons around AI adoption, so its measures are production proxies rather than a complete measure of software quality.
Editorial summary · Publisher article
What matters
- Coding agents were associated with 30 percent more generated lines of code, 20 percent more commits, and 23 percent more pull requests on average.
- Issue and Epic resolution rates did not change significantly after adoption.
- The dataset covered 300 million work events from more than 700 software firms between 2021 and March 2026.
- Human review became a downstream constraint, with longer reviews and more requested revisions.
Context & caveats
- The findings are based on aggregated engineering analytics and proxies for software output, not a direct audit of product quality.
Why this made the edition
Detailed research-based reporting provides a large empirical sample, methodology, measured effects, and an important distinction between more generated code and unchanged software output.
Source material