AI agents started more than 40% of GitHub pull requests in September, up from 17% in March, according to GitHub’s 2026 Octoverse findings. That is a striking shift in how code changes begin, and it is already bringing questions of capacity, cost and review along with it.
Microsoft AI Watch analysis
What happened
Quartz’s account of the Octoverse findings says GitHub’s report also found that 61% of Copilot users worked with multiple models in September, with no single model accounting for more than a quarter of use. The figures point to developers using agents and switching between models, rather than settling on one tool for every job.
The same account describes pressure on GitHub’s service as agent use grew. It cites more than a dozen outages lasting over an hour since March and reports that GitHub temporarily stopped accepting new individual Copilot plans in April, with usage-based charging due to start in June. Those details come through Quartz’s reporting, which also draws on CNBC and comments from Microsoft and GitHub representatives.
Key findings
- Agents are starting more pull requests
Their share rose from 17% in March to more than 40% in September, according to GitHub. - Copilot users are mixing models
GitHub says 61% used multiple models in September, and none topped 25% of use. - Service capacity is part of the story
Quartz cites extended outages and GitHub’s response to rising demand for agent work. - AI is shaping language trends too
GitHub says TypeScript, Python and Rust grew quickly, with AI involved in roughly 40% to 44% of merged pull requests in those languages.
Why it matters
This is more than a new way to draft code. When an agent proposes the first version, developers’ work shifts towards checking, testing and deciding what to accept. The reported service strain also shows that the supply of agent help depends on platform capacity and pricing, not just a developer’s appetite for automation.
The findings are from GitHub’s own report, as presented by Quartz. They describe activity on GitHub, not every software team or coding platform. Still, the scale of the reported change makes this a useful early measure of where agent-assisted development is heading.
Our read
The headline number is impressive; the less glamorous details may matter just as much. Agents can propose code at speed, but reliable services, clear costs and good review practices determine whether that speed is useful. “More pull requests” is a workload statistic, not a productivity verdict.
What to watch
- Whether GitHub shares updated figures on agent-started pull requests and Copilot usage.
- How usage-based charging affects developers’ access to agent features.
- Whether teams report measurable gains in delivery without sacrificing code quality.
Discussion spark: If agents start most code changes, should developers be judged more on how much they write, or on how well they review and test what agents propose?
Sources and evidence
- The biggest findings on AI agents, top languages and fastest-growing developer communities – qz.com (10 November 2026, 12:15 UTC)
not affiliated with or endorsed by Microsoft