GitHub has published its 2026 Octoverse report. The reporting window runs from October 1, 2025 to September 30, 2026, a year in which developers pushed code to GitHub 17.9 billion times, more than 2024 and 2025 combined. The figure that has drawn the most attention: as of September, more than 40% of pull requests were started by an agent.
"As of September, developers start more than 40% of pull requests with an agent."
More than doubled in half a year
The share of agent-initiated PRs has climbed quickly. It stood at 17% in March and passed 40% in September. In absolute terms, agents started about 110 million PRs in September, roughly six times the March figure.
The report also compares merge rates. In September, 90% of agent-initiated PRs were merged within two weeks, against about 86% for the control group with no agent involvement; in February the figure for agent PRs was 83%. Among PRs that touch multiple files, agents started 87% and humans started 79%.
GitHub's own Copilot data points the same way. The share of monthly active users who have used agent features rose from 35% a year ago to 55%, and Copilot's cloud agent grew 230% year over year. In September, 61% of Copilot users used more than one model, up from 50% four months earlier; the number of models actually invoked by paying users grew from 29 in May to 47 in September, and no single model accounted for more than 25% of requests.
Volume is up across the board
The basic metrics have all multiplied. There were 34.3 billion commits over the year, against roughly 12.9 billion in 2025. Merged PRs reached 1.3 billion, about 3.5 million a day, compared with 538 million in 2025. GitHub Actions ran 45.4 billion jobs, or 1,440 per second. Dependabot auto-merged more than 30 million PRs.
One caveat: this edition changes the methodology. Earlier reports covered September through the following August and counted public activity only. Starting this year the numbers include private repositories, and the 2024 and 2025 data have been recalculated under the new definition. The figures therefore do not match the original numbers in past reports, and year-over-year comparisons should use the recalculated values in this report.
By programming language, TypeScript remains first, followed by Python and JavaScript. TypeScript added 3.62 million monthly contributors over the past year and Python added 3.23 million. Rust contributors doubled year over year. The share of merged PRs involving AI was 44.1% for TypeScript, 41.1% for Python and 40.4% for Rust, with Java the lowest at 21.2%.
China ranks third among developers
GitHub now has more than 225 million developers. The United States (38.3 million) and India (32 million) remain the two largest communities, each adding more than 10 million over the year. China ranks third with 12.4 million, followed by Brazil at 10.2 million and Japan at 7.15 million. Ranked by public repository pushes, the top three are likewise the United States, India and China.
Country attribution is based on the location developers enter themselves, and GitHub applies a publication threshold of 100 developers. Many teams in China keep their main codebases on Gitee, self-hosted GitLab instances or internal company networks, and that activity does not appear in GitHub's statistics. The 12.4 million figure is therefore closer to the number of Chinese developers active on GitHub than to the total number of developers in China.
For teams in China, the share of agent PRs translates into direct pressure on code review. The merge rates in the report suggest most of these PRs are ultimately accepted, but GitHub does not disclose revert rates or post-release defect rates for agent PRs, so the merge rate alone does not settle the question of quality. Mainstream coding assistants in China, including Tongyi Lingma, Trae and CodeBuddy, have not published PR data on the same basis, so a like-for-like comparison is not yet possible.
What the report does not answer
The 40% figure is defined as PRs "started by an agent". The report does not break down how much humans changed afterward or how many lines of code ultimately came from a model. The industry classification section relies on a language model classifying a random sample, and GitHub itself labels the results as estimates.
Sources: GitHub official blog, Octoverse 2026 report; CocoLoop cross-checked totals for pushes, commits, merged PRs and Actions jobs, the share of agent PRs, and developer counts by country.