A day after GitHub published its Octoverse 2026 report on October 28, one number was everywhere: AI agents started more than 40% of all pull requests on the platform in September, up from 17% in March. That is about 110 million agent-started pull requests in a single month. (Source: LetsDataScience)

It is a striking figure, and it comes from a company that sells those agents. This post lays out the headline numbers, the ones that get less attention, and the reasons to read the 40% with care. If you code, manage developers, or just wonder how much of the software around you is now machine-drafted, this is the clearest public snapshot so far.

The headline numbers

LetsDataScience, summarizing the report, lists year-over-year jumps: pushes rose from 6.7 billion to 17.9 billion, commits from 12.9 billion to 34.3 billion, merged pull requests from 538 million to 1.3 billion, and GitHub Actions jobs from 14.0 billion to 45.4 billion. That works out to about 3.5 million merged pull requests per day and about 1,440 Actions jobs per second. The Stack notes that developers pushed code 17.9 billion times in the past year, more than the previous two years combined, which it attributes mainly to agent use. (Sources: LetsDataScience, The Stack)

One caution on numbers: Simon Willison's post quotes monthly chart figures including 220 million merged pull requests per month, and says merged pull requests tripled in less than a year. That does not line up neatly with 1.3 billion for the year, which is about 108 million per month on average, The likeliest explanation (my inference) is that 220 million is a recent monthly peak well above the 12-month average; Willison's 7.38 billion commits per month similarly exceeds 34.3 billion divided by 12. I use the annual totals here and flag the mismatch rather than resolve it. (Sources: Simon Willison, LetsDataScience)

Agents merge as often as people

According to the report summary, agent-involved pull requests merged within two weeks 83% of the time in February and 90% in September, while pull requests without detected agent involvement held near 86% in both months. About 87% of agent pull requests change multiple files, versus 79% of other pull requests. (Source: LetsDataScience)

The Stack adds that the average new repository created with an agent logged eight times as many commits as the average new repository without one, and LetsDataScience reports that since June about 55% of commits came from brand-new repositories. (Sources: The Stack, LetsDataScience)

Also on LumiChats

AI Tools

Google Let Rival AI Coding Agents Into Android Studio. Here's How It Works

7 min read→

AI Tools

Why AI Agents Get Worse the Longer They Run: 6 Fixes and 2 New Failure Modes

8 min read→

AI Tools

Walmart Will Let AI Agents Shop for You, With Spending Limits You Set

8 min read→

Where AI shows up by language

The same summary reports AI involvement in merged pull requests of 44.1% for TypeScript, 41.1% for Python, 40.4% for Rust and 21.2% for Java. It also reports that 55% of monthly active Copilot users now use agentic features, up from 35% a year earlier, and that the platform has more than 225 million developers, with 38.3 million in the US and 32.0 million in India. (Source: LetsDataScience)

Why to read the 40% carefully

• It is a floor, not a ceiling: GitHub says the figure counts only agents it can detect, so the true share may be higher.
• Merge rate is not review cost: a merged request tells you what was accepted, not how many hours reviewers spent deciding.
• It is self-reported: GitHub sells these agents, and the merge-rate and benchmark claims are its own.
• Review capacity may not have tripled with volume: merged pull requests more than doubled, so more code may be passing through the same human checks.
• A pull request started by an agent can still be heavily edited by a person, so it does not mean a person wrote nothing.

The summary also mentions a product claim from the same event: Copilot Fusion, launched on October 28, routes each step of a task to a model. GitHub says Copilot Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, two points behind Claude Code with Opus 5.5 Max, at about half the cost per task ($5.59 versus $11.42 per Artificial Analysis). GitHub says it funded the evaluation. The 40% time claim appears in only one report. (Source: LetsDataScience)

What it means if you write code

The practical shift is from writing every line to reviewing what an agent drafts. The same summary says 61% of users who chose their own model used more than one in September, and that paid users drew on 47 models with none above 25% of requests. If true, that suggests developers are mixing models instead of loyalty to one. (Source: LetsDataScience)

If you use AI for coding, try the same task on more than one model before you commit. LumiChats offers 40+ premium AI models in one subscription, so you can compare the results yourself.