GitHub says agents now start over 40% of pull requests, and its own numbers show how late the surge came

GitHub's annual Octoverse report, released on October 28 at the opening of GitHub Universe, has one number it clearly wants people to repeat: by September, more than 40% of pull requests on GitHub started with a first draft from a coding agent. In March the share was 17%. In absolute terms that is 110 million agent-started pull requests in September alone, about six times March's volume.

The rest of the report is mostly about scale. Over the Octoverse year (October 1, 2025 to September 30, 2026, public and private repositories), developers pushed 17.9 billion times, more than in 2024 and 2025 combined. Commits reached 34.3 billion, more than three times the figure from two years earlier, and merged pull requests more than tripled to 1.3 billion, an average of 3.5 million a day.

The year was mostly its last few months

Annual totals hide how late the growth arrived. GitHub says activity climbed steadily from January and then surged around June, which it ties to better models, better harnesses and agents that can keep working across days and run in the cloud. Two figures from the keynote make the shape visible. According to Simon Willison, who was in the room, the slides put merged pull requests at 220 million a month and commits at 7.38 billion a month, with most of the increase in the second half of the year.

Divide the report's own annual totals by twelve and you get roughly 108 million merged pull requests and 2.9 billion commits in an average month. So by the end of the year GitHub was running at about twice its average pull request rate and two and a half times its average commit rate. That is our arithmetic, not GitHub's, but it means the annual comparison with 2024 understates where things now stand.

GitHub's agent numbers can be pushed a little further the same way. If 110 million agent-started pull requests in September was about six times March, March had roughly 18 million. At 17% of the total, that implies about 108 million pull requests opened in March. In September, 110 million at "more than 40%" means fewer than 275 million in total. Agent-started pull requests went up about sixfold while everything else, by these rough numbers, did not even double. Most of the new volume is agent volume.

What "agent" means here

The definitions matter, and GitHub is upfront about them. An "agent-started pull request" is one whose first draft came from a coding agent GitHub can detect. The report calls its agent signal "a floor, since not all agent use may be detected". The comparison group is described as pull requests "without detected agent involvement", not as human pull requests. A developer who has Claude Code or Codex write a change locally and then pushes it under their own name is, as far as this telemetry knows, not using an agent. The true share is probably higher than 40%, and the "non-agent" side is not a clean human baseline.

That caveat applies to the report's quality-flavored comparison too. In February, 83% of agent-involved pull requests merged within two weeks, slightly behind about 86% for the rest. By September the agent figure was 90%, while the other group stayed near 86%. A merge is not a review verdict, though. Many agent pull requests are presumably opened at the request of the person who will merge them, in their own repository, which makes a quick merge more likely regardless of quality. The number says agent pull requests are no longer stalling. It doesn't say they are better.

The 8x figure

The statistic GitHub puts first is that among repositories created this year, the average repo with an agent configuration file (a file that tells a coding agent how to work in the project) logged eight times as many commits as the average new repo without one. GitHub reads this as a sign that teams "building with agents from day one are moving at a markedly different pace".

It may be that, but the comparison is between a repository someone bothered to configure for an agent and the average new repository, which includes every abandoned experiment, homework assignment and one-commit test. Writing an agent config file is itself a sign that a project is being actively worked on. GitHub's own glossary adds that "repositories without one may still use AI". The 8x is a real difference between two groups; how much of it the agents cause is not something this data can show.

Do agents pick TypeScript?

The report also suggests a new influence on language choice: the language an agent picks when the developer doesn't specify one. Among repositories created in 2026, TypeScript was the primary language in 35% of those with an observed agent signal and 24% of those without. GitHub concludes that agents "appear to favor a narrower set of languages, TypeScript chief among them".

The next sentence in the report complicates that. For repositories created before 2026, mostly before agents were writing whole projects, the gap is similar: 27% with an agent signal versus 17% without. In those repositories the language was chosen long before an agent arrived, so the pattern there runs the other way: TypeScript projects are more likely to bring in agents. Both effects can be true at once, but the data shown doesn't separate them. TypeScript did stay the most used language on GitHub by contributors, adding 3.62 million monthly contributors, ahead of Python's 3.23 million; Rust grew fastest in percentage terms, doubling.

Smaller details

• No dominant model. Developers on paid GitHub Copilot plans used 29 distinct models in May and 47 in September, and no single model took more than 25% of requests in September. Of developer-selected requests, 38% went to general-purpose models, 37% to deep reasoning and planning models and 25% to fast, lightweight ones. 61% of users who picked models themselves used more than one. These figures exclude auto model selection, which 63% of paying Copilot developers used at least once in September.
• Hands-off work grows fastest. Copilot cloud agent, where a task is assigned and the agent works on it independently, grew 230% year over year, against 85% for agent chat.
• Review tooling is catching up. Stacked pull requests, in preview since July 30, had more than 264,000 users by September. Push protection, the secret scanner, ran 5.8 billion scans, nearly four times 2025's pace.
• Reverse engineering is up. New public repositories mentioning interoperability or reverse engineering rose from about 30,000 in 2025 to more than 130,000 in 2026, which GitHub reads as developers delegating more ambitious tasks.
• The inference stack draws contributors. llama.cpp, SGLang and vLLM appear on both GitHub's fastest-growing and most-engaged lists of community projects, alongside PyTorch and CPython. On the vendor side, the fastest-growing repositories by contributors are coding agents and editors: Claude Code, Codex, opencode, VS Code and GitHub's Copilot CLI.

A few of the report's classifications lean on models themselves. Repositories were sorted into industries by a fine-tuned embedding model reading names, descriptions, topics and READMEs, and country attribution for pushes was adjusted using LLM classification of profile locations. GitHub notes that the industry and AI-mention shares "can be related because both read the same text". Those sections are estimates layered on estimates, and the report says so.

The core numbers, pushes, commits and pull requests, are GitHub's own telemetry, and nobody outside GitHub can check them. Taken as given, they describe a platform that ended its year running at about twice its average monthly pace, with the extra volume coming mostly from agents and mostly in the last few months. GitHub's answer to that is more review tooling and more models to route between. Whether review keeps up with a sixfold increase in machine-drafted changes is the question the report doesn't try to answer.