Agent-merged PRs draw fixes at 1.62x the odds of human PRs

Agent-written pull requests draw follow-up fixes at 1.62 times the odds of human ones, a comparison of 6,774 merged agent PRs against 5,044 human PRs in the same repos found.

The gap matters for anyone reviewing agent code as if it were done once merged. Repos in the sample all carry at least 500 GitHub stars. These are production codebases accepting PRs from Codex, Copilot, Devin, Cursor and Claude Code, and fixing them more often than they fix human work.

69.6% of the follow-up fixes on an agent PR come from the same agent that wrote it. 76.4% of fix PRs are agent-authored across every commit. Agents return to patch their own mistakes more than humans return to patch a stranger's, according to the paper on arXiv.

A merged PR isn't the finish line for agent output the way it is for human output. Repos that treat an agent's merge as equally final to a human's are signing up for a second review pass more often than they'd expect. The odds say that pass should go back to the same agent that wrote the code, not a cold human reviewer.