Stats of GitHub PRs opened/merged by LLM agents (simonwillison.net)

Albert Avetisian’s PRarena repo uses the GitHub Search API and a collect_data.py job (run every three hours via GitHub Actions) to track pull requests attributed to different coding agents, updating a public leaderboard and time-series chart of adoption and merge success. The snapshot highlights stark differences: OpenAI’s Codex Cloud dominates by volume (≈1.9M PRs, ≈1.6M merged, ~84% merge rate), GitHub Copilot shows high usage but a lower merge rate (≈247k PRs, ≈152k merged, ~61.5%), and Claude Code sits between them in merge rate (≈146k PRs, ≈123k merged, ~84.2%). The repo relies on agent-specific search terms in PR bodies or authors to attribute contributions; the “earliest” example links are heuristically filtered and may include false positives. For the AI/ML community this is a compact, data-driven lens on real-world agent adoption and practical impact: raw PR counts show footprint and integration scale, while merge rates offer a coarse proxy for usefulness or reviewer acceptance. Important technical caveats apply — attribution depends on searchable markers and repo conventions, merged ≠ correct code, and different agents serve different workflows (automation vs. reviewer-facing suggestions), which can skew rates. Still, PRarena is a valuable monitoring tool for researchers, tool builders, and engineering managers to quantify agent usage trends, compare real-world outcomes across models, and spot areas where agent output needs improvement or policy work.