We wanted to know how much coding agents are really being left alone. Not in surveys, in practice. So on 1 October 2026 we counted, using GitHub's public search, what agents did on public repositories over the previous seven days.
These are search counts, so treat them as approximate. They also only cover public repositories. Private ones, where most companies work, are not in here at all.
The count
Pull requests opened in one week by an agent acting on its own:
| Agent | Pull requests in 7 days |
|---|---|
Codex (branches named codex/) | about 241,000 |
| Devin | about 15,000 |
| GitHub Copilot coding agent | about 13,000 |
| Cursor agent | about 7,600 |
| Jules | about 2,200 |
That is close to 280,000 pull requests in a week, on public repositories alone, that no person typed.
Three more numbers from the same day:
- About 8.6 million commits in the week carried a Claude co-author line.
- About 19,000 repositories have Claude Code wired into their CI workflows, where it runs with nobody at the keyboard by design.
- About 1.1 million
AGENTS.mdfiles and about 768,000CLAUDE.mdfiles exist in public repositories: written instructions for an agent, which you only write if you expect one to work there.
The question of whether teams will let agents work unattended is over. They already do.
A pull request is the end of the run
Here is what those numbers do not show. A pull request is the last thing an agent does. Everything before it, the run itself, is invisible to the person reviewing.
The reviewer sees a diff. They do not see whether the agent spent forty minutes getting there, tried the same failing command eleven times, or wandered into a different module for a while and came back. If the diff looks fine, the run is assumed to have been fine.
And those are the runs that finished. The run that looped all night and never opened a pull request does not appear in any count. It appears on the invoice.
Most of this happens when nobody is there
Agents in CI run when a workflow fires, not when someone is watching. Background agents are started at the end of the day precisely so they can work through the night. That is the value: work that happens while you are not there.
It is also the exposure. The hours when agents do the most are the hours when a loop, a drift or a quiet stall has the longest to run before anyone notices.
More than one vendor is normal
When we looked closely at a sample of 34 companies with agent activity that week, 7 were running agents from two or more vendors in the same organisation. It is a small sample, but it matches what you would expect: different agents are good at different things, and teams use what works.
Each vendor shows you its own runs in its own dashboard. Nobody shows you all of them.
What we take from this
Three things, none of them complicated:
- Unattended agents are already normal. Hundreds of thousands of agent pull requests a week on public repositories alone.
- The run is the blind spot. Review looks at the result. Nothing looks at the forty minutes before it, or at the run that never produced a result.
- The gap is time. A loop caught in seconds costs nothing. The same loop found in the morning cost the night.
That gap is what Infersync closes. It watches each agent run, whichever vendor it comes from, stops loops, drift and overspend in seconds, and asks a person only for what can hurt. It runs on your own machines, so your code never leaves.
If your agents open pull requests while you sleep, ask for access and connect one of them for a week. We will send you what it caught.
How we counted
GitHub issue and code search, signed in, on 1 October 2026, for the seven days before. Pull requests were matched by the agent's app account as author, except Codex, which was matched by branch name. Commit and file counts are GitHub's reported totals for the search. All figures are rounded and public repositories only.