GitHub Octoverse 2026: Agents Started Over 40% of Pull Requests

Sources checked and dated on publication. See our editorial policy.

On October 28, 2026, GitHub published its 2026 Octoverse report, and one number travelled furthest: by September, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March, or about 110 million in that single month. It is a striking measure of how quickly coding agents moved into everyday work. It is also a vendor-reported, detection-based floor from the company that sells those agents. Here is what the report actually says, what its methodology allows you to conclude, and what it does not.

What the report actually says

Related guide: Flock Safety to Cut ~270 Jobs (18%) as AI Surveillance Backlash Bites

On October 28, 2026, GitHub published its 2026 Octoverse report, subtitled "Commits, pushes, reach new highs in 2026 as developers build with agents." The line that travelled furthest: "By September 2026, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March. That is 110 million in September alone, about 6x March's volume." Developer Tech and Visual Studio Magazine both led with the same figure the following day.

The report is a snapshot of the twelve months ending September 30, 2026, covering public and private repositories unless a metric says otherwise. Alongside the agent share, GitHub reports 17.9 billion pushes, 34.3 billion commits, 1.3 billion merged pull requests (an average the post puts at 3.5 million a day) and roughly 45 billion Actions jobs, with GitHub Actions running 1,440 jobs every second. It says GitHub now serves more than 225 million developers, with the United States at 38.3 million and India at 32.0 million. The Stack summarised the headline as developers pushing more code in 2026 than in the previous two years combined.

These are large numbers, and they are worth separating from the conclusions sometimes drawn from them. The report does not say agents wrote 40% of code, that 40% of developers use agents, or that agent-written changes are better. It says a detectable agent produced the first draft of more than 40% of pull requests opened in one month on one platform.

Why more than 40% is a floor, not a share of code

Related guide: Your Claude API Key Is Now a Primary Target: What Anthropic's September Threat Report Found

GitHub's own glossary does most of the qualifying for us. It defines an agent-started pull request as "a pull request whose first draft came from a coding agent that we can detect," and an agent signal as "an observed sign of agent involvement in telemetry, such as an agent-opened pull request. It is a floor, since not all agent use may be detected." In other words, the 40% is a lower bound on detectable agent involvement, not an upper bound, and not a measure of how many lines of code were machine-written.

That distinction has practical consequences. A pull request that begins with an agent's draft can be heavily rewritten by a person before it merges; the report counts the origin, not the final authorship. The glossary also notes that repositories without an agent configuration file "may still use AI," so the eight-times-as-many-commits comparison the report draws between new repos with and without such a file is a correlation between observable configuration and activity, not proof the file caused the difference. Visual Studio Magazine makes the same point explicitly: "These are GitHub's platform measurements and should not be taken as proof that agent-written changes are more reliable," and calls the eightfold figure "a correlation that does not establish causation."

There is also a self-interest caveat that belongs in the same paragraph. GitHub sells Copilot, the agent tooling these numbers describe. The report is the company's own telemetry, published at its own conference, and its methodology section describes classification by "a fine-tuned multilingual embedding model" for industry data and LLM-based adjustment for geography — estimates, not audited counts. None of that makes the figures wrong; it means they should be read as a vendor's measured view of its own platform.

Merge rates are not a quality score

Related guide: Best VPS for AI agents in 2026: Hetzner vs AWS Lightsail vs EC2

The most misread numbers in the report are about merging. GitHub says agent-started pull requests "more often change multiple files: about 87%, compared with 79% of pull requests without detected agent involvement." It also reports that in February 2026, 83% of agent-involved pull requests merged within two weeks, slightly behind roughly 86% for the comparison group; by September the agent figure had risen to 90% while the comparison group held near 86%. Developer Tech adds the necessary qualifier: "These figures reflect the time taken to merge pull requests, rather than the quality of the resulting code."

A faster merge tells you a change was accepted, not how many reviewer hours it took, whether it was reverted later, or whether it introduced a defect. GitHub's own framing is more careful than the headlines: "Review is where developers shape what comes back," and the report points to growing use of AI-assisted code review and GitHub Advanced Security (both up more than 90% year over year), 5.8 billion push-protection scans for exposed secrets, and more than 30 million merged Dependabot pull requests. It also says more than 264,000 developers had used stacked pull requests, a review workflow that splits related changes into smaller pieces. If volume is rising that fast, the interesting question is whether review capacity rose with it, and the report does not publish a measure of reviewer effort.

Independent commentary has flagged a related consistency issue worth noting rather than resolving. Simon Willison, attending GitHub Universe, posted keynote figures of 220 million merged pull requests per month and 7.38 billion commits per month. Those do not line up neatly with the annual totals of 1.3 billion merged pull requests (about 108 million a month on average) and 34.3 billion commits. The likely explanation is that the keynote numbers are recent monthly peaks against a twelve-month average, but GitHub's post does not publish a reconciliation table, so the honest position is to cite the annual totals and flag the mismatch.

The wider picture: models, review and delegation

Related guide: Nemotron 3.5 Bolt: NVIDIA's 74B and 4GB open models bring agents to local hardware

Beyond the 40% figure, the report describes a workflow shift. Among monthly active GitHub Copilot users, 55% used agentic features, up from 35% a year earlier. The most hands-off mode grew fastest: Copilot cloud agent, where a developer assigns a task and the agent works independently, grew 230% year over year, against 85% for agent chat. The report's own gloss is that developers "still ask agents questions, and increasingly they hand over whole tasks."

Model choice is fragmenting too. In September 2026, 61% of Copilot users who selected a model themselves used more than one model across their sessions, up from 50% four months earlier. Paying users drew on 47 distinct models in September, up from 29 in May, and no single model carried more than 25% of requests. Of every 100 developer-selected requests, GitHub reports 38 went to general-purpose models, 37 to deep reasoning and planning, and 25 to fast and lightweight models. Its methodology section is explicit that these counts "describe observed use, not model quality," and exclude auto model selection, staff usage, background utility models and requests made with a developer's own API key. Separately, 63% of paying Copilot developers used auto model selection at least once in September.

Language data adds a quieter signal. TypeScript, Python and Rust, the fastest-growing communities, have the highest shares of merged pull requests with detected AI involvement, about 40% to 44%; Java, C# and C++ sit lower, between about 21% and 34%. GitHub notes that among repositories created in 2026, TypeScript was the primary language in 35% of those with an observed agent signal versus 24% without one, and offers a cautious interpretation: developers working with agents "gravitate toward languages where agent output is easier to check." The report frames this as a question, whether a fourth influence is emerging, rather than a finding.

What teams should actually measure

Related guide: The UK can test frontier AI models again: what the 17-day freeze and the U-turn actually changed

If you write or review code, the useful output of Octoverse 2026 is a set of questions to ask your own team, not a number to quote. GitHub's data cannot tell you whether agent-drafted changes are cheaper or safer in your repository, because it publishes no measure of review effort, revert rate or defect rate. Those are the metrics that would answer the question: track review time per pull request, the share of agent-started changes reverted within 30 days, and the volume of security findings in agent-authored diffs.

Keep the human decision in the loop. The report closes on exactly this point: "None of that activity changes who decides what the code should do ... someone still has to decide whether it works, whether it belongs, and whether it's good." Practical controls that follow from the report's own findings include keeping secret scanning and push protection on (GitHub reports 5.8 billion scans this year), treating AI-assisted review as an aid rather than an approval, and using smaller, stacked changes when reviewing agent output. None of these require trusting the 40% figure.

And treat the number as what it is: one platform's self-reported, detection-limited measure of where a first draft came from, in a single month, from a company with a product to sell. It is a strong signal that agent-assisted work has moved into the mainstream of software development. It is not evidence that agents write 40% of the world's code, that review has become unnecessary, or that the resulting software is better.

Bottom line: GitHub's Octoverse 2026 reports that agents started more than 40% of September pull requests, up from 17% in March, within a year of record activity across pushes, commits and merges. The figure is a genuine milestone for agentic coding and a genuinely limited one: vendor-reported, detection-based, and silent on code quality. Read it as evidence that review has become the bottleneck worth measuring, not as a verdict on the code.

Prices and benchmark figures are standard published rates as of November 6, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.