Here’s a stat that made us put our coffee down. In March, AI agents started about 17% of pull requests on GitHub. By September, that number had jumped past 40%. That’s the headline from the GitHub Octoverse 2026 report, and it says a lot about how software actually gets written now.
If you write code, manage developers or just wonder where all those “AI will take over coding” headlines are heading, this one’s for you. Let’s walk through it.
So what actually changed this year?
GitHub publishes Octoverse every year. It’s a big data dump on how the world’s developers use the platform. This year’s edition landed on 28 October at GitHub Universe, the company’s annual conference, and it’s all about agents.
An AI agent, in plain terms, is software that can take a task and work through several steps on its own. Think “fix this bug and open a pull request,” rather than “suggest the next line of code.”
Wait, what’s a pull request?
A pull request is a proposed change to a codebase. Someone writes the change, then teammates review it before it goes into the main project. So when GitHub says agents started over 40% of pull requests, it means an AI wrote the first draft of four in ten proposed changes.
In September alone, that came to about 110 million agent-started pull requests. That’s roughly six times the March volume, according to GitHub’s own write-up.
The GitHub Octoverse 2026 numbers are kind of wild
The agent stat isn’t the only jaw-dropper. Developers pushed code to GitHub 17.9 billion times in the 12 months to 30 September. That’s more than 2024 and 2025 combined.
Commits hit 34.3 billion, up from 12.9 billion the year before. Merged pull requests reached 1.3 billion, compared with 538 million in 2025. Developer and blogger Simon Willison summed it up neatly: “This is a pretty stark representation of the impact of agentic engineering on GitHub.”
GitHub also says activity bent sharply upward around June. Since then, about 55% of commits have come from brand-new repositories. In other words, people aren’t only tweaking old projects. They’re spinning up new ones at a wild pace.
One more fun detail: new 2026 repositories with an agent configuration file averaged eight times the commits of those without one. That’s a small file that tells an agent how a project works. Sure, busy projects are more likely to set one up in the first place. Still, it’s a hint.
Do agent-written changes actually get merged?
Fair question. A flood of AI-written code means nothing if humans throw most of it away. But the data here looks surprisingly good for the bots.
By September, 90% of agent-started pull requests merged within two weeks. That’s up from 83% in February. Pull requests without a detected agent stayed around 86%. Agent-started changes also touched multiple files more often, 87% of the time versus 79%.
Now, a couple of honest caveats. GitHub can only count agents it can detect, so it calls these numbers a floor. Also, a quick merge isn’t the same as good code. It only tells you a human said yes. And since we’ve written before about how AI agent hacking just got cheap, careful review still matters a lot.
Okay, but should you care?
If you’re a developer, yes, though not in a scary way. Copilot data in the report shows 55% of monthly active users now use agentic features, up from 35% a year earlier. Your teammates are probably already doing this.
Nobody’s married to one model, either. In September, 61% of Copilot users who picked a model by hand used more than one. Our guess? The skill that matters most is shifting toward reviewing. You’ll spend less time typing boilerplate and more time checking, testing and deciding what ships. Tools like background AI agents push that even further, because they keep working while you’re away.
There’s a global angle too. GitHub now counts more than 225 million developers. The US has 38.3 million and India has 32.0 million, and each added over 10 million this year. The report also points to a fast-growing developer community in Africa, naming Kenya, Nigeria and South Africa.
For smaller tech hubs like Armenia, that’s both a warning and a chance. We think agents can shrink the gap between a five-person team and a fifty-person one. So a small, sharp team that learns to work with agents can punch well above its size.
What should you learn next?
If you’re picking a language, the report has clues there too. TypeScript is now the top language by contributors, with Python second and JavaScript third. Rust doubled in the past year. In TypeScript, 44.1% of merged pull requests had detected AI involvement, compared with 21.2% in Java.
Our take on the GitHub Octoverse 2026 report
Honestly, the 40% number is less surprising than how fast it happened. In six months, agents went from a side experiment to a regular teammate.
Still, we’d watch one thing closely. Code is getting cheaper to write, yet it isn’t getting cheaper to trust. The teams that do best next year will probably be the ones that get really good at saying “no” to an agent’s pull request.
So here’s the question to take back to your own team: who’s reviewing the robots?