The Pulse
GitHub Says Agents Started More Than 40% of September Pull Requests
GitHub says detectable coding agents produced the first draft for more than 40% of pull requests in September 2026, up from 17% in March. The company counted 110 million agent-started pull requests that month and says 90% of agent-involved
AI.info Team · Written with AI · fact-checked ·
“By September 2026, more than 40% of pull requests on GitHub started with an agent’s first draft, up from 17% in March.”
GitHub Staff, GitHub
What GitHub counts as an agent-started pull request
GitHub published the figures on October 28 in its 2026 Octoverse report. The company says agents started 110 million pull requests in September, roughly six times March’s volume. Its measure describes a pull request whose first draft came from a coding agent that GitHub can detect—not every pull request that received some help from AI.
That distinction matters. GitHub calls detected agent activity a floor because its telemetry may miss some agent use. The headline share therefore tracks identifiable agent-written first drafts, not a complete tally of AI involvement across all software changes.
September brought 110 million agent starts
The rise coincided with a broader jump in activity on GitHub. Across the 12 months ending September 30, developers merged 1.3 billion pull requests, more than three times the total two years earlier. GitHub says its overall activity accelerated around June, as coding agents took on longer-running and more remote tasks.
Agent-started changes more often touched several files: about 87% did so, compared with 79% of pull requests without detected agent involvement. The figures suggest that agents are being used for work spanning multiple parts of a project, not just isolated edits. They do not, by themselves, show whether those broader changes required less work from human developers.
More agent-involved pull requests merged within two weeks
GitHub also reported a shift in how quickly agent-involved pull requests reached a merge. In February, 83% merged within two weeks, slightly below the roughly 86% rate for pull requests without detected agent involvement. By September, the agent-involved share had risen to 90%, while the comparison group remained near 86%.
Those figures describe a change in observed merge timing, not proof that agents caused faster reviews. The report does not establish why the rates moved or whether the pull requests were comparable in complexity. Still, the change challenges the idea that agent-written work necessarily waits longer before it lands.
Copilot users are handing over larger tasks
GitHub’s Copilot figures point to wider adoption of agent-style work inside its own product. Among Copilot monthly active users, 55% used agentic features, up from 35% a year earlier. The company says use of Copilot cloud agent—which lets a developer assign a task for an agent to work on independently—grew 230% year over year; agent chat grew 85%.
The comparison suggests developers are moving beyond asking for help in a conversation and assigning agents work to carry out. GitHub’s data does not say how much human prompting, editing, testing or review those tasks still require. A pull request remains a proposal for a code change, not a measure of finished work or software quality.
The review workload remains visible in the numbers
GitHub’s own account places review alongside the increase in agent output. Use of AI-assisted code review and GitHub Advanced Security each grew more than 90% year over year, while its push-protection system ran 5.8 billion scans during the Octoverse year. The report also says more than 264,000 developers had used stacked pull requests by September, a workflow that breaks related work into smaller changes for sequential review.
The clearest finding is a change in where many pull requests begin: detectable agents now provide the first draft for more than two in five, according to GitHub. The company’s September count—110 million starts—puts a concrete scale on that shift, while the two-week merge comparison shows that human review and acceptance remain part of the process.