We are researchers from ETH Zurich interested in the real-world adoption and impact of Code Agents.
To measure this, we built a dashboard, scraping all public PRs on GitHub, analyzing which are created by different code agents (Codex, Jules, Copilot, Devin, etc.), and measuring their merge rates, sliced by various repository and PR characteristics.
Since mid-May, we've analyzed over 10 million PRs and already found some interesting trends:
Usage is high, but shallow. Agents submit ~7% of all PRs overall, but only ~1–2% on popular repos. Most activity is in low-star or experimental projects.
Merge rates vary drastically. On low-traffic repos, some agents get 90%+ of their PRs merged. On popular projects, that can drop to <25%.
Pre-review helps. Agents that require human-in-the-loop review (e.g., Jules, Codex) have 30–50% higher merge rates than Copilot-style fire-and-forget PRs.
Bias toward new code. Agent PRs mostly add code. Refactorings and deletions are rare.
If you have ideas for what other characteristics we should look at let us know or play with the code yourself
scottgg•7mo ago