AI Agents Started 40% of GitHub Pull Requests in September, Up From 17% in March. And They Merge More Often.
More than 40% of GitHub pull requests in September started as an agent’s first draft. That’s 110 million in one month, 6x March, and those PRs now merge within two weeks more often than the rest: 90% to 86%.
TL;DR
- 🤖 17% to 40%+ in 6 months: the share of GitHub pull requests that started with an agent’s first draft, March to September 2026. GitHub calls it a floor.
- ✅ Agent PRs pulled ahead: 90% merged within two weeks in September, to about 86% for PRs without detected agent involvement. In February they trailed, at 83%.
- 📈 1.3 billion merged PRs in the Octoverse year, up from 538 million. Commits hit 34.3 billion, more than 3x 2024.
- 🔀 No model wins: Copilot’s paid users ran 47 models in September, and none carried more than 25% of requests.
What GitHub counted
GitHub published Octoverse 2026 on October 28, the opening day of GitHub Universe. The Octoverse year now runs from October 1, 2025 to September 30, 2026, and unless GitHub says otherwise the numbers include private repositories as well as public ones.
That’s a change from earlier reports, which used a September-to-August window and public activity only. GitHub recalculated its 2024 and 2025 figures on the new window, so they won’t match what you read last year. The comparisons below use GitHub’s recalculated numbers.
The headline sentence is short: “By September 2026, more than 40% of pull requests on GitHub started with an agent’s first draft, up from 17% in March.” GitHub adds that this is “a floor, since not all agent use may be detected.”
The agent numbers
| Metric | Earlier | September 2026 |
|---|---|---|
| Share of PRs started by an agent | 17% (March 2026) | More than 40% |
| Agent-started PRs in the month | ~1/6 of September (March) | 110 million |
| Agent-involved PRs merged within 2 weeks | 83% (February 2026) | 90% |
| PRs without detected agent involvement, merged within 2 weeks | ~86% (February 2026) | ~86% |
| Agent-started PRs that change multiple files | ~87% (vs 79% for the rest) |
Source: GitHub Octoverse 2026. GitHub gives March agent volume only as “about 6x” less than September’s 110 million.
Merged within two weeks, agent-involved PRs vs the rest
Source: GitHub Octoverse 2026. Bar length is the merge rate itself.
Two things happened at once. Agents went from roughly one PR in 6 to more than 2 in 5, and the PRs they draft stopped losing on the one quality signal GitHub reports: whether a human merged them within two weeks.
They also aren’t one-line typo fixes. About 87% of agent-started PRs touch more than one file, against 79% for PRs with no detected agent.
Read the merge rate carefully
“Merged within two weeks” is a speed and acceptance metric, not a bug count. A PR can merge fast and still ship a regression. GitHub doesn’t publish revert rates or post-merge defects for either group.
The comparison group is also fuzzy. It’s “PRs without detected agent involvement,” and GitHub says its detection misses some agents. Some of that 86% is agent work nobody labeled.
Selection matters too. GitHub doesn’t break down what kind of work agents get. A team that hands an agent the easy tickets will see a high merge rate for reasons that have little to do with the model.
The fair reading: agent PRs are no longer the ones reviewers send back. It doesn’t prove they are better code. It proves teams now trust them enough to merge them on the same clock as everything else.
The platform under the agents
The agent share grew inside a platform that itself exploded. Every row below is GitHub’s own count, public and private repositories included.
| 12 months to September 30 | 2024 | 2025 | 2026 |
|---|---|---|---|
| Pushes | 5.2B | 6.7B | 17.9B |
| Commits | 10.0B | 12.9B | 34.3B |
| Merged pull requests | 402M | 538M | 1.3B |
| GitHub Actions jobs | 8.0B | 14.0B | 45.4B |
Source: GitHub Octoverse 2026, recalculated on the new window.
Merged pull requests per Octoverse year
Source: GitHub Octoverse 2026. Bar length relative to 2026’s 1.3 billion.
Pushes in the last 12 months beat 2024 and 2025 combined. Merged PRs went from 538 million to 1.3 billion, about 2.4x by our math, or 3.5 million a day. At GitHub Universe, EVP Jay Parikh put the monthly rate at 220 million merged PRs, as Simon Willison reported from the room.
GitHub also says activity climbed from January and surged around June, and from June on about 55% of all commits came from brand-new repositories. New repos created this year with an agent configuration file logged 8x as many commits, on average, as new repos without one.
Where AI shows up, by language
| Language | Merged PRs with detected AI involvement | Community growth, YoY |
|---|---|---|
| TypeScript | 44.1% | 134% |
| Python | 41.1% | 118% |
| Rust | 40.4% | 162% |
| Java | 21.2% | 30% |
Source: GitHub Octoverse 2026, past 12 months.
The fastest-growing languages are also the most agent-heavy. GitHub says AI shows up in “more than one in five merged pull requests in every language shown,” with Java lowest. TypeScript stays the top language on GitHub, ahead of Python and JavaScript, and among repos created in 2026 it was the primary language in 35% of those with an agent signal versus 24% without one.
Nobody picks one model
The Copilot numbers answer a question every coding-tool buyer asks: which model should I use? Developers’ answer is “several.”
| GitHub Copilot, developer-selected requests | Value |
|---|---|
| Users who picked a model and used more than one (September) | 61% (50% in May) |
| Distinct models used on paid plans | 29 in May, 47 in September |
| Largest single model’s share of requests (September) | Under 25% |
| Requests per 100: general-purpose / deep reasoning / fast | 38 / 37 / 25 |
| Paying users who used Auto model selection at least once | 63% |
Source: GitHub Octoverse 2026. Excludes Auto selection, staff traffic, background models and bring-your-own-key requests; each model needed at least 1,000 weekly active users. GitHub notes these figures describe use, not model quality.
The split is the interesting part: 37 of every 100 requests go to deep reasoning and planning models, 25 to fast and light ones. Developers already route by task by hand. GitHub’s answer is automation: its Project HydraFusion router, which handled almost 60,000 sessions in preview, is now generally available as GitHub Copilot Fusion.
Agent mode is now the mainstream Copilot experience, too. 55% of monthly active Copilot users use agentic features, up from 35% a year earlier. The Copilot cloud agent grew 230% year over year; agent chat grew 85%.
Fun facts
- 45.4 billion GitHub Actions jobs works out to 1,440 jobs per second, or 124 million a day.
- GitHub passed 225 million developers in September. The US has 38.3 million and India 32.0 million, and both added more than 10 million this year.
- Construction has the highest share of repos with agent-opened PRs of any industry, about 9%. In mining, health care and agriculture, more than 4 in 10 repos mention AI, but agent-opened PRs stay under 3%.
- Stacked pull requests launched in preview on July 30. More than 264,000 developers had used them by September.
- Repos about interoperability and reverse engineering grew more than 300%, from about 30,000 to more than 130,000.
What this means if you code with AI
The pull request is the agent’s output format now. With 110 million agent-started PRs a month, the useful question about a coding agent isn’t “does it autocomplete well?” It’s “does it open a PR my reviewers will merge?” Judge tools on that.
Review is the bottleneck you’re buying for. If 4 in 10 PRs start as an agent draft, your team spends more time reading diffs than writing them. GitHub says AI-assisted code review grew more than 90% year over year. Make sure your agent tests its own diff before a human sees it.
Don’t lock into one model. 47 models in use, none above 25%. 61% of Copilot users who pick a model already use more than one. Pick a tool that lets you, or picks for you per task.
Write the agent config file. New repos with one logged 8x the commits of new repos without one. That’s correlation, not proof, but an AGENTS.md or similar file is the cheapest experiment on this list.
Ignore the merge-rate victory lap for hard work. 90% within two weeks says agents handle the work teams give them. It doesn’t say they handle your hardest ticket. Keep measuring reverts on your own repo.
FAQ
What share of GitHub pull requests do AI agents open?
GitHub’s Octoverse 2026 says more than 40% of pull requests on GitHub in September 2026 started with an agent’s first draft, up from 17% in March. That was 110 million agent-started pull requests in September alone, about 6x March’s volume. GitHub calls the number a floor, since not every agent is detected.
Do agent pull requests get merged?
More often than the rest, by GitHub’s measure. In September 2026, 90% of agent-involved pull requests merged within two weeks, against about 86% for pull requests without detected agent involvement. In February, agent PRs were behind, at 83%.
How many pull requests were merged on GitHub in 2026?
1.3 billion in the Octoverse year from October 1, 2025 to September 30, 2026, an average of 3.5 million a day. GitHub’s recalculated figures for the two years before are 538 million (2025) and 402 million (2024).
Which AI model do GitHub Copilot users pick most?
None dominates. In September 2026, no single model carried more than 25% of developer-selected Copilot requests, paid users used 47 distinct models, up from 29 in May, and 61% of users who picked a model themselves used more than one.
Which programming languages see the most AI-written code on GitHub?
Among merged pull requests over the past 12 months, detected AI involvement was highest in TypeScript (44.1%), Python (41.1%) and Rust (40.4%), and lowest in Java (21.2%). GitHub says AI shows up in more than one in five merged pull requests in every language it shows.
Where MegaBrain fits
Full disclosure: we sell an AI coding agent, and these are GitHub’s numbers, not ours. MegaBrain’s terminal coding agent goes from the task to the pull request: it plans the work, writes the code across files, runs your commands and tests, reviews its own diff, then commits and opens the PR. It puts many models behind one plan, with free models on Free, Claude, GPT and Gemini from Pro, your own API keys and local models, and Auto picks one for each task. See the plans. MegaBrain is not affiliated with GitHub or Microsoft.
Sources
All pages opened October 31, 2026.
- GitHub Blog — Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents (October 28, 2026)
- Simon Willison — notes on GitHub Octoverse 2026 from GitHub Universe (October 28, 2026)
MegaBrain · free to start
Try a free AI coding agent in your terminal or your editor.
No card: the free models, your own API keys and local models with no limits from us. 90+ models on one plan from $8 a month.
npm install -g @getmegabrain/cliThen run megabrain in your project and sign in once.
Newsletter
Stay in the loop
Get our latest AI research and guides. No spam, unsubscribe anytime.