Independent Research (open-source repositories)
MSR 2026 study: coding agents give open-source repos a velocity bump only as their first AI tool, but lasting technical debt either way
A longitudinal staggered difference-in-differences study of open-source repositories in the AIDev dataset, with matched controls, defining adoption as a repo's first agent-generated pull request. When agents were a project's first AI tool, commits rose about 36% and lines added about 77%; repos that already used AI IDEs saw roughly +3% commits and −6% lines. Static-analysis warnings and cognitive complexity rose persistently in both groups.
The numbers (as reported)
| Commits (agents as first AI tool) | +36.3% |
|---|---|
| Lines added (agents as first AI tool) | +76.6% |
| Commits / lines (prior AI IDE users) | +3.1% / −6.3% |
| Static-analysis warnings | ~+18% (persistent) |
| Cognitive complexity | ~+39% (persistent) |
Facts
- Company / subject
- Shyam Agarwal, Hao He, Bogdan Vasilescu (MSR 2026)
- Tools
- —
- Codebase
- brownfield
- Work type
- Research (open-source repositories)
- Stack
- —
- Published
- 2026-04-13
- Last checked
- 2026-10-19
Caveats
Observational, open-source repos only, and a Mining Challenge paper on one dataset. Adoption is inferred from the first agent-authored PR, so earlier unlabelled AI use is missed. Replication package: github.com/shyamagarwal13/agentic-coding-impact.