37,623 provenance-labeled agent PRs: Codex code was reverted half as often as human code, Devin's 31% more.

arXiv 2609.17598 studies PRs from OpenAI Codex, Devin, GitHub Copilot, Cursor and Claude Code across 2,807 repositories (Dec 2024 to Jul 2025), combining AIDev with 58,792 cached GitHub API responses. Codex PRs were reverted 6.1% of the time against a human baseline of 11.5% (OR 0.50); Devin hit 14.5% (OR 1.31). Agent code pooled across vendors was less likely than human code to carry a security smell (OR 0.63), driven by fewer hardcoded credentials and eval-style constructs. Review effort concentrated unevenly: Copilot PRs drew the most human reviews and change requests, Claude Code PRs waited longest for a first human review at a 12.6 hour median. Quality here is vendor-specific, not a property of "AI code," which kills most of the arguments people have about this. Pick by measured revert rate on your repo, and budget review latency separately from generation quality.