AI agents pull requests on GitHub faster than ever: more than 40% of all pull requests opened on the platform now start with an agent-generated draft, according to GitHub’s Octoverse 2026 report – a figure that would have sounded like science fiction eighteen months ago.
The trend line moves even faster than the headline. In March 2026, agent-started pull requests sat at 17% of the total. By September, GitHub counted roughly 110 million of them in a single month – about six times the March volume. The agents did not just join the workforce; they opened their own branches.
From 17% to 40% in six months
GitHub published the 2026 Octoverse report on October 28, framing it as a record year for commits and pushes across every industry and field – with agents doing an increasing share of both. The 40% figure counts pull requests that were initiated by an AI agent writing the draft changes, not merely pull requests that happen to contain AI-generated lines somewhere in the diff.
That distinction matters. Starting a pull request means an agent picked the task, wrote the code, ran whatever checks it could, and opened the change for human review. The human’s job begins where the agent’s ends.
Merging is the easy part
The follow-through numbers are the ones maintainers should chew on. Figures attributed to GitHub’s research put the merge rate for agent-assisted pull requests at roughly 84%, with about 55% of merged AI-generated changes landing without any human modification at all. If those margins hold, the bottleneck in open source is no longer typing code – it is reviewing it.
Review has always been the unglamorous half of software: reading carefully, asking what happens when the network flakes out, and politely rejecting the pull request that renames every variable for fun. As agent volume climbs, that tasteful judgment call becomes the scarcest resource on the platform.
What developers should take from Octoverse 2026
Three practical read-outs:
- Review capacity is the new scaling limit. Teams that measured developer output in features shipped should start measuring merge latency and review load instead.
- Agent drafts are becoming the default input. Linters, templates and CI checks tuned for human-shaped diffs will increasingly encounter machine-shaped ones.
- The 17%-to-40% slope does not flatten on its own. Six months took the technology from a minority practice to the plurality default; the next six months will tell whether platforms hit a quality ceiling or blow through it.
GitHub’s full report is published on the GitHub Blog. The agent era of software did not arrive with a keynote – it arrived as a growing pile of pull requests waiting for a human to click approve.
Statistics are drawn from GitHub’s Octoverse 2026 report and secondary reporting on it, current as of October 30, 2026.
