Preprint: Agent PR revert rates range from 6.1% to 14.5% by vendor
A Texas Tech preprint analyzed 37,623 provenance-labeled PRs across 2,807 GitHub repos from December 2024 to July 2025. 90-day revert rates: Codex 6.1%, Devin 14.5%, Copilot 12.5%, Cursor 11.4%, Claude Code 10.5%, versus 11.5% for a matched human baseline. The spread across agents exceeds the gap between the best agent and humans.
- Codex: 6.1% reverts (n=17,756); Devin: 14.5% (n=2,185)
- Copilot 12.5%, Cursor 11.4%, Claude Code 10.5% vs 11.5% human baseline
- Agent code had fewer security smells (OR 0.63), effect only in largest PRs
- Claude Code PRs waited longest for first human review: median 12.6 hours
Read next
AI