427,042 pull requests in 24 hours. That is how many were opened on public GitHub repositories by AI coding agents between Nov 11 and Nov 12, 2026, counted by the mark each agent leaves on its work. It is 31.9% of the 1,339,893 pull requests opened there in the same 24 hours.
I wanted one number that is not a survey and not a press release, so I asked GitHub's own search.
Four widely used coding agents mark their pull requests by default. One adds a "Generated with…" line to the description, and all four work on branches with a fixed prefix. A pull request counts when it carries one of those marks.
The query runs against GitHub's public search API with is:pr is:public and a 24-hour window. The number is the total_count GitHub returns. Nothing is sampled or extrapolated.
| Count | Share | |
|---|---|---|
| Pull requests by agents, last 24 hours | 427,042 | 31.9% of all pull requests |
| All pull requests, last 24 hours | 1,339,893 | |
| Pull requests by agents, last 30 days | 10,862,952 | |
| of those, merged | 9,887,585 | 91.0% |
| of those, closed without a merge | 376,440 | 3.5% |
| of those, still open | 598,927 | 5.5% |
The 30-day rows add up the four agents, so a pull request that carries two marks is counted twice there. In the 24-hour count it is counted once. By agent, the 30-day merge rate runs from 68.5% to 92.6%.
It does not mean agents are that good.
Of the latest 115 agent pull requests on the board, 63% were opened by the repository's own owner, 13% by a member or collaborator, and 24% by someone outside the team. Most of this is people using an agent in their own repository and merging their own work minutes later.
The picture changes when you look at one project that takes pull requests from strangers. One large open source repository, openclaw/openclaw, got 806 agent pull requests in 30 days, which is 4% of the 22,310 opened there. 492 were merged and 125 were closed without a merge. Of the latest 30, 29 came from outside the team.
It undercounts. Private repositories are invisible. So is agent work that carries no mark: a tool used inside an editor, or anyone who deletes the footer. Pull requests that an agent opens under a bot's own account can't be tied to a person. And a human can name a branch with one of the prefixes by hand, which would be counted.
So the total is a floor for signed agent work on public GitHub. It is not a census of AI-written code, and it says nothing about quality: a merge in your own repository is not a review.
The count runs live, with a board that pages through real pull requests as they are opened, merged or closed.
Type a GitHub username and it shows what that person's agents opened in 2026. Type owner/repo and it shows how many agent pull requests a repository received in 30 days, and how many were turned away. Those lookups run in your browser against the public API. The site stores a count of lookups, not who was looked up.
There is also a small generator for the pull request rules that agents read in a repository's AGENTS.md.
Disclosure: the counter is ours (MyClawn). It and this post were written by an AI agent working for us, and the numbers are what GitHub's search returned at the time stated. The code is open source under MIT, and the exact searches are on the method page.
If you maintain a project: what share of your incoming pull requests would you guess is agent-written, and how far off was the report? The live board: https://agents.myclawn.com/?utm_source=devto&utm_medium=social&utm_campaign=myclawnlabs&utm_content=explain-agent-arrivals