40% of GitHub Pull Requests Now Start With an AI Agent: Octoverse 2026
GitHub published its annual Octoverse report on October 28, 2026, at its Universe conference — and one sentence in it describes a genuine changing of the guard in software development: “By September 2026, more than 40% of pull requests on GitHub started with an agent’s first draft, up from 17% in March.” A pull request is a proposed code change that teammates review before merging. So in September, more than four in ten of those proposals began as an AI agent’s draft, not a human’s. Microsoft CEO Satya Nadella shared the figures on stage. Let’s unpack what the numbers actually say — and what they don’t.
What “Agent-Started” Actually Means
Precision matters here, because the headline is easy to misread. GitHub counts an “agent-started pull request” as one whose first draft came from a coding agent that GitHub can detect. That is a narrower claim than “40% of code is written by AI” — it counts pull requests, not lines of code, and it counts the starting draft, not the finished change.
There’s another wrinkle, and GitHub is upfront about it: detected agent use is a floor, not a ceiling. The company’s own glossary calls the agent signal “a floor, since not all agent use may be detected.” The real share of agent-started work is probably higher than 40% — GitHub just can’t see all of it.
The Scale of the Shift
The raw volumes are hard to grasp. Agents started about 110 million pull requests in September 2026 alone — roughly six times March’s volume. The Octoverse year runs October 1, 2025 to September 30, 2026, and across that window GitHub logged 1.3 billion merged pull requests (about 3.5 million a day), 34.3 billion commits, and 17.9 billion pushes. September alone saw about 7.38 billion commits, more than triple the level a year earlier.
Adoption isn’t just growing — it’s compounding. Among monthly active GitHub Copilot users, 55% now use agentic features, up from 35% a year ago. Cloud-based autonomous agents, which run tasks on remote machines instead of the developer’s laptop, grew 230% year over year.
Agents Merge Faster — With an Asterisk
Here’s the finding that will spark the most debate: agent-involved pull requests are getting merged faster than human-only ones. In September, 90% of pull requests with detected agent involvement merged within two weeks, versus about 86% without it. Back in February, the agent figure was 83% — slightly behind. In seven months, agents went from a touch slower to measurably quicker.
Agent-started PRs also tend to be broader: 87% touched multiple files, compared with 79% of the others. That fits the pattern of agents doing sweeping refactors and multi-file implementations that humans then review.
But — and this is the asterisk that matters — GitHub measured merge rates, not quality. The report contains no data on defects, security bugs, or reverts in agent-started PRs. A merge means a change was accepted by a human reviewer. It doesn’t, by itself, prove the change was correct, secure, or free of rework down the line. Faster merging could mean agents write good code. It could also mean reviewers are waving things through. The data can’t tell those stories apart.
The Rest of the Report in Numbers
A few more figures worth knowing. TypeScript held its spot as the most-used language, followed by Python and JavaScript — and among repositories created in 2026 with an observed agent signal, TypeScript was the primary language in 35%, versus 24% without one. Model choice has become a per-step decision: in September, 61% of Copilot users who selected models themselves used more than one model across their sessions, up from 50% four months earlier, and no single model accounted for more than 25% of requests.
What It Means for Developers
Step back and the picture is clear: the bottleneck in software development has moved. Writing the first draft of a change is increasingly the agent’s job; the scarce, valuable work is reviewing — reading agent-produced diffs carefully, catching the subtle bug, asking whether the change should exist at all.
That’s a different skill set from cranking out boilerplate, and teams that don’t adapt their review practices are the ones most likely to learn about agent mistakes from their users instead of their pull requests. The 40% figure is a floor, and it’s still climbing. Plan accordingly.
Outbound links:
Internal links: