GitHub Octoverse 2026: AI Agents Now Start 40% of Pull Requests

Four years ago, an AI tool opening a pull request on GitHub was a novelty worth a blog post. Today, according to GitHub's own numbers, it's just Tuesday — agents started more than 40% of all pull requests in September 2026, up from roughly 17% in March.

That figure comes from GitHub's annual Octoverse report, unveiled at GitHub Universe on October 28. It's the company's yearly data dump on how people (and increasingly, software) actually use the platform, and this year's edition reads less like a trend piece and more like a snapshot of a workflow that's already changed. A pull request, for anyone outside software teams, is simply a formal proposal to merge a chunk of code into a project — the unit of work that gets reviewed, debated, and eventually accepted or rejected. Increasingly, the first draft of that proposal isn't typed by a person.

The numbers behind the shift

GitHub says agents opened around 110 million pull requests in September alone. Zoom out to the full year — October 2025 through September 2026 — and developers collectively pushed 17.9 billion commits-worth of changes and merged 1.3 billion pull requests across public and private repositories. These aren't niche experiments confined to a few AI labs; GitHub's framing is that agentic coding has become the baseline, not the exception.

A few details matter more than the headline percentage:

• Agent-started pull requests tend to touch more files than human-started ones — around 87% involve multiple files, compared to 79% for the rest.
• They're also merging faster and more reliably over time: about 83% of agent-involved pull requests were merged within two weeks back in February, and that rose to roughly 90% by September.
• GitHub is careful to note the 40% figure counts pull requests an agent started, not the share of code actually written by one — a human still reviews, edits, and signs off before anything ships.

That last caveat is worth sitting with, because it's where the real story is. Code24 covered the run-up to this report when GitHub Universe's schedule was first revealed to be stacked with AI-agent sessions — the hype was expected. What's notable is that the usage data now backs it up.

A pull request that starts with an agent isn't a finished product. It's a first draft waiting for a human editor — and GitHub's own numbers suggest that editor is getting faster at the job, not redundant.

Why this matters for India specifically

Buried in the same report is a number that should matter more to Indian readers than the agent statistic: India now has around 32 million developers on GitHub, putting it second globally behind the United States' 38.3 million — and growing faster. GitHub's data shows roughly one in every seven new developers joining the platform worldwide this year is Indian, and the company's own projection puts India ahead of the US by 2030, at an estimated 57.5 million developers against America's 54.7 million.

Put those two data points together and the implication for Indian software teams is direct: a workforce that's already scaling faster than anywhere else is also the one most likely to be reshaped early by agent-assisted development. For engineering teams in Bengaluru, Hyderabad, and Pune hiring junior developers, the job increasingly looks less like writing code line-by-line and more like reviewing, steering, and catching mistakes in what an agent already drafted. That's a different skill to train for, and most computer science curricula in India haven't caught up to it yet.

What happens to code review

The practical fallout lands hardest on code review, the process where a second set of eyes checks a change before it ships. When a small fraction of pull requests come from agents, human reviewers can treat each one as an exception worth extra scrutiny. When four in ten do, review has to become a filter built for volume — which is exactly why GitHub used Universe to push its own AI-assisted review and model-routing tools alongside the Octoverse numbers, a pattern security researchers have flagged as something teams should budget review time for rather than treat as a free efficiency gain.

None of this means human programmers are becoming optional. It means the valuable skill is shifting from writing the first version of something to deciding whether the first version an agent wrote is actually safe to ship. Companies that treat that shift as a training problem will likely end up ahead of the ones that just buy the tooling and hope the habits take care of themselves.