EngiSignal › Research

AI Pull Request Scoreboard, Edition 1: How AI Code Gets Merged

Research brief 02November 4, 20265 minute read

AI coding agents now open pull requests on their own. We looked at 361,296 of them across popular open-source projects to see what happens next. Most get merged. Many get merged in minutes. And more than half of the merged ones have no approving review on record.

68%

of AI agent pull requests were merged

55%+

of merged ones had no approving review on record

1 in 4

merged within 10 minutes of being opened

46%

of reviews on these pull requests were written by bots

Data: AIDev, an open research dataset of AI agent pull requests on GitHub, February 2025 to March 2026. Method and limits are at the end.

Speed does not follow risk

If review matched risk, a bug fix would wait longer than a docs edit. It mostly does not. Share of merged AI pull requests that were merged within 10 minutes of being opened:

Docs32%
Bug fixes26%
Features25%
Refactors22%
Tests19%

Based on a random sample of 22,000 pull requests. Type is read from the pull request title.

A quarter of AI bug fixes and new features went in faster than most people finish a coffee. Some of those are fine: the person who asked the agent merged it after a quick look. But nothing in the record tells a team which ones got real scrutiny and which ones did not.

The scoreboard

How pull requests from each agent were handled. This measures review behavior, not code quality.

AgentMergedMedian time to mergeMerged within 10 min
OpenAI Codex80%4 minutes60%
Claude Code79%1 hour26%
Google Jules*69%1 hour 10 min25%
GitHub Copilot66%2 hours 15 min8%
Devin63%1 hour 50 min20%
Cursor60%1 hour 25 min21%

Merged is the share of closed pull requests that were merged. Sample of 22,000 pull requests across 6,157 repositories. *Google Jules: 145 pull requests, read with care.

What happens after merge

Independent studies of the same kind of data give a first look at the outcome side.

What this means for your team

  1. Know which AI changes were actually reviewed. An approval on record is the minimum. Track the share of AI pull requests merged without one.
  2. Let risk set the pace. Docs can move fast. Authorization, data, payments and shared code should not, whoever wrote them.
  3. Treat bot reviews as input, not sign-off. Nearly half of reviews here came from bots. A human owner still decides on risky changes.

Coming in edition 2

The outcome scoreboard: which merged AI pull requests were reverted or needed a follow-up fix within 30 days, by agent and by type of change. Same open data, same open method.

Get edition 2 first.

Waitlist members get each scoreboard before it goes public. Tell us what you want measured next.

Join the waitlist
  1. Method: counts of pull requests, merges and reviews cover the full curated AIDev set (361,296 pull requests from popular repositories, February 28, 2025 to March 31, 2026). Timing, type and per-agent figures come from a random sample of 22,000 of them. "No approving review on record" is a lower bound: there were 110,863 approvals in total against 246,777 merges. Some projects let maintainers merge their own work without a formal approval, so read it as missing sign-off, not proof of no review. Analysis by EngiSignal, November 4, 2026.
  2. AIDev dataset, Hao Li et al., version 5, CC BY 4.0
  3. AIDev: Studying AI Coding Agents on GitHub
  4. Will It Survive? Deciphering the Fate of AI-Generated Code in Open Source, January 2026
  5. Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild, March 2026
  6. Security in the Age of AI Teammates: An Empirical Study of Agentic Pull Requests on GitHub, January 2026