ESEM 2026 study of 669 developers: after adopting coding agents they write longer PR descriptions and test more, with no drop in effort
A longitudinal study of 669 developers across 103 repositories in the AIDev dataset compared 11 behaviour metrics before and after they started working with coding agents, against 228 control developers in repositories without agent activity. The one large agent-linked effect was a 76.8% increase in PR description length, with smaller improvements in all three testing measures. Coding effort stayed stable. Reviews became faster, less detailed and more permissive, but that also happened in the control repositories, so the authors put it down to wider trends.
The numbers (as reported)
| Developers studied | 669 across 103 repositories, plus 228 controls |
|---|---|
| PR description length | +76.8% after agent adoption |
| Coding effort | Stable |
Facts
- Company / subject
- Muhammad Hamza, Dominik Siemon and Wardah Naeem Awan (ESEM 2026)
- Tools
- Unstated
- Codebase
- brownfield
- Work type
- Research / benchmark
- Stack
- —
- Published
- 2026-10-05
- Last checked
- 2026-10-21
What the source claims Independent
Reported by an independent source, such as a study or the press.
- PR descriptions got 76.8% longer after developers started using agents, with smaller but consistent gains in testing discipline +76.8%
“We find one large agent-associated effect: a 76.8% increase in PR description length”
Caveats
Drafted from the abstract only. Uses the same AIDev dataset as two existing MSR 2026 stories. Measures developer behaviour, not delivered software.
More stories by tool: Unstated
Related topics: Research and benchmarksBrownfield codebasesIndependent sources