GitHub Octoverse: 40 Percent of PRs Start With AI

Posted on by
Contents show
GitHub Octoverse

GitHub Octoverse 2026 records 17.9 billion pushes in twelve months, more than 2024 and 2025 combined, alongside 34.3 billion commits and 1.3 billion merged pull requests. Those numbers describe a platform that doubled in a year, and the report's more interesting finding explains why: by September, more than 40 percent of pull requests began with an agent's first draft, up from 17 percent in March.

That shift is the heart of the GitHub Octoverse data, and it reframes what developer productivity means. Volume growth on a platform can come from more people, and GitHub did add developers, passing 225 million accounts. But the steepest change is in how work starts. A pull request used to begin with a human writing code; it now frequently begins with a human describing intent and an agent producing the first attempt, which turns the developer's job toward specification and review.

The GitHub Octoverse report quantifies that turn in ways that matter for anyone running an engineering organization: agent-started pull requests touch more files, merge faster than they did six months ago, and cluster in specific industries in patterns that reveal where automation has actually landed.

This article covers the GitHub Octoverse numbers that matter most, the eightfold commit gap between repositories with and without agent configuration, what the forty percent figure does and does not mean, how review tooling is absorbing the change, why developers now run multiple models per session, and which industries are adopting fastest.

The GitHub Octoverse numbers, and the June inflection

Four GitHub Octoverse series tell the story, and each broke its own record by a wide margin.

Pushes reached 17.9 billion, up from 6.7 billion the year before and 5.2 billion the year before that. Commits climbed to 34.3 billion, more than three times where they stood two years ago. Merged pull requests more than tripled to 1.3 billion, an average of 3.5 million per day. And GitHub Actions ran 45.4 billion jobs, which works out to 1,440 jobs every second, or 124 million a day, more than three times last year's total.

The GitHub Octoverse growth shape is as informative as its size. Activity climbed steadily from January and then bent sharply upward around June, which the report attributes to a cluster of changes arriving together: better underlying models, better harnesses around them, and agents that became persistent rather than session-bound, capable of tracking progress across days and orchestrating specialized sub-agents. Agents also began doing substantial work in the cloud rather than only on local machines, which removed the constraint of the developer's laptop staying open.

The composition of the GitHub Octoverse surge is worth noting. From June onward, about 55 percent of commits came from brand new repositories, which suggests a wave of new projects rather than only intensified work on existing ones. At the same time, the developers pushing code have been on GitHub for an average of 36 months, and more than 75 percent of those who opened a pull request in August pushed code again in September. New projects and returning developers grew together, which is unusual; platform surges typically come from one or the other.

Public contributions doubled to 2.5 billion, and the report names individual extremes: one developer made more than a million contributions maintaining a machine-readable catalog of public API information using agentic workflows, another made 178,000 maintaining a computer-checked mathematics library. Those figures are outliers, but they indicate what agent-assisted maintenance looks like at the high end.

The GitHub Octoverse eightfold commit gap

The single most quoted GitHub Octoverse statistic is also the most misunderstood, and it deserves careful reading.

Among repositories created this year, the average repository containing an agent configuration file logged eight times as many commits as the average new repository without one. The comparison is between new repositories only, which controls for age, and the metric is commits per repository, which controls for the number of projects.

What the GitHub Octoverse eightfold figure does not control for is selection. Teams that add an agent configuration file on day one are, by definition, teams that intend to work with agents, and they are plausibly also teams with more ambitious scope, more automation comfort, and more willingness to commit frequently. The eightfold gap therefore measures the difference between agent-native projects and everything else, not the causal effect of adding a configuration file.

The GitHub Octoverse causal reading is nonetheless defensible in one direction. A workflow that generates code in small increments, tests automatically, and commits often produces more commits for the same feature than a workflow where a developer writes for two days and commits once. Part of the gap is behavior change rather than output change, and both are real: the repository with an agent configuration is likely doing more work and recording it more granularly.

For engineering leaders, the useful GitHub Octoverse takeaway is not the multiplier but the correlation. Agent configuration files are appearing at the start of projects rather than being retrofitted, which means the next cohort of repositories is being built with delegation as a default assumption rather than an experiment.

What the GitHub Octoverse forty percent actually measures

The agent-started pull request figure is the GitHub Octoverse report's clearest statement of workflow change, and its definition matters more than the number.

By September 2026, more than 40 percent of pull requests on GitHub started with an agent's first draft, up from 17 percent in March. That is 110 million pull requests in September alone, roughly six times March's volume. The measurement detects agent involvement rather than asserting authorship, which means the human role in those pull requests is specification, direction, and review of an agent's initial output.

The characteristics of those pull requests are informative. About 87 percent change multiple files, compared with 79 percent for pull requests without detected agent involvement, which fits the pattern of agents taking on tasks with broader blast radius than a single-file edit. And they merge faster over time: in February, 83 percent of agent-involved pull requests merged within two weeks, slightly behind the 86 percent for the comparison group; by September, agent pull requests were at 90 percent while the comparison group held near 86.

That crossover is the most consequential line in the report. Agent-generated work started slightly less likely to merge quickly and became more likely to merge quickly than human-only work, which suggests that review processes adapted and that the quality of agent output improved. Neither conclusion is certain from one metric, and both are consistent with the data.

The delegation trend shows up in Copilot usage as well. Among monthly active Copilot users, 55 percent now use agentic features, up from 35 percent a year earlier, and the fastest growth is in the most hands-off mode: the cloud agent, where a developer assigns a task and an agent works independently, grew 230 percent year over year against 85 percent for agent chat.

Review is where the job moved, per GitHub Octoverse

If agents write the first draft, review becomes the primary human contribution, and the GitHub Octoverse data agrees, and the GitHub Octoverse data shows tooling reorganizing around that reality.

Stacked pull requests, which split related work into smaller sequential changes that are easier to review, launched in preview on July 30 and were used by more than 264,000 developers by September. That adoption rate for a review-oriented feature is a signal about where pain now lives: reviewing a large agent-generated change is harder than reviewing a small human one, because the reviewer did not build the mental model while writing the code.

Usage of AI-assisted code review and GitHub Advanced Security both grew more than 90 percent year over year, and push protection, which screens for exposed secrets such as API keys and passwords, ran 5.8 billion scans, nearly four times 2025's pace. Secret scanning volume is a proxy for how much code is moving and how much of it is generated quickly enough that credentials slip in.

The GitHub Octoverse review turn has a skills implication that the report does not state directly. A reviewer of agent-generated code needs to evaluate intent against implementation quickly, spot the plausible-but-wrong pattern that models produce, and decide whether to accept, revise, or reject. That is a different skill from writing the code, and organizations that promote their fastest writers into review roles may find the skills do not transfer automatically.

GitHub Octoverse: developers run several models at once

The GitHub Octoverse data on model choice describes a market that has already fragmented, and it undercuts the assumption that one model wins a developer's loyalty.

In September 2026, 61 percent of Copilot users who selected a model themselves used more than one model across their sessions, up from 50 percent four months earlier. Paid users had access to 47 distinct models in September, against 29 in May, and no single model carried more than 25 percent of requests that month.

The report groups models by purpose, deep reasoning and planning, general purpose, and fast and lightweight, and finds requests spreading across all three. That distribution is rational: a developer planning a refactor wants a reasoning model, a developer renaming variables across a repository wants a fast one, and a developer debugging an unfamiliar stack may want whichever model has the best track record on that language.

The practical consequence of the GitHub Octoverse model data is that model-agnostic tooling has become the default expectation rather than a differentiator. Developers are also weighing price against capability, and the report notes that familiarity with a provider influences selection, which is the kind of soft factor that predicts retention better than benchmark scores do.

For teams standardizing their stack, the data argues against single-vendor commitments. A workflow built around one model's quirks becomes expensive to change and fragile when that model is deprecated, and the current market depreciates models on a quarterly rhythm.

The developer map keeps widening

Behind every volume figure is a population, and the GitHub Octoverse geography section describes a community that grew in more places at once.

GitHub now serves more than 225 million developers. The United States remains largest at 38.3 million and India second at 32 million, and each added more than 10 million developers this year. China follows at 12.4 million, Brazil at 10.2 million with 3.28 million added in twelve months, then Japan, the United Kingdom, Indonesia, Germany, Russia, and France.

Last year's report projected India at 57.5 million developers by 2030 and the United States at 54.7 million. One year into that five-year window, every country in the projection has closed at least a fifth of the distance to its own 2030 figure, with the United States at 38.6 percent and India at 28.3 percent. Morocco has closed 54.4 percent, Indonesia 45.4 percent, and Kenya 21.7 percent, which is the slowest in the group.

The regional drivers are specific rather than generic. Brazil's growth tracks a mature fintech sector; Germany's reflects enterprise AI adoption across its industrial base; Indonesia benefits from a booming digital economy; South Korea and Japan from government investment in AI talent and infrastructure. African markets, particularly Kenya, Nigeria, and South Africa, are growing through fintech, agritech, and coding bootcamps.

The forecasting section is unusually honest. GitHub ran more than 40 models forward to 2031, and about half call for developer counts more than tripling, just under half call for doubling to tripling, and only three expect growth under 2x. The company notes that its own models no longer agree on where the curve goes, which is a fair description of a platform that just doubled in a year and has no comparable prior period to extrapolate from.

For anyone planning hiring, tooling budgets, or localization, the takeaway is that the center of gravity is shifting south and east, and the growth is not concentrated in the traditional technology hubs. The next cohort of developers is more likely to be building for a local industry than for a software company, which is consistent with the industry data in the same report.

Which industries are actually adopting

The GitHub Octoverse report classifies public repositories by industry for the first time, and the results complicate the narrative that software companies lead AI adoption.

Measuring the share of repositories that mention AI against the share with agent-opened pull requests produces two different pictures. Administrative and support services, mining, health care, and agriculture all have more than four in ten repositories mentioning AI, while agent-opened pull requests remain rare at under 3 percent. Construction breaks the pattern: about a quarter of its repositories mention AI, but about 9 percent have agent-opened pull requests, the highest share of any sector, with manufacturing next.

The GitHub Octoverse reading is that mentioning AI and delegating work to agents are different adoption stages. Sectors in the first group are experimenting publicly, documenting intent, and building prototypes, while construction and manufacturing have moved to handing tasks to agents, likely because their work involves well-specified, repetitive, verifiable outputs.

Growth against each sector's own 2022 baseline is equally striking. Mining, quarrying, and oil and gas extraction grew roughly 65 times from a small base, administrative services and construction 17 to 18 times, and manufacturing almost tripled. Professional, scientific, and technical services rose from fourth to second place among all industries, behind only information.

Those GitHub Octoverse multipliers describe a long tail waking up. The developers building in mining and agriculture are not software engineers at technology companies; they are domain experts using agents to build tools for problems they understand better than any vendor does, which is precisely the adoption pattern that agentic coding was supposed to enable.

For readers tracking the tooling that makes this workflow possible, the orchestration layer that manages parallel agents is examined in our Meta Conductor analysis, the model economics behind high-volume delegation are unpacked in our GPT-6.1 Terra piece, and the security implications of agent-written code are covered in our GPT-6 Cyber analysis.

What engineering leaders should take from it

The GitHub Octoverse report is a platform vendor describing its own traffic, which makes it useful as a mirror rather than a forecast, and four conclusions hold up under that caution.

First, the bottleneck has moved. When agents produce 40 percent of first drafts, the constraint on a team is no longer typing speed; it is the clarity of task specification and the throughput of review. Teams that invest in issue templates, acceptance criteria, and test coverage get more from the same models, because those artifacts are what an agent needs to produce a usable first attempt.

Second, tooling budgets should shift toward review and verification. Stacked pull requests, AI-assisted review, secret scanning, and dependency automation all grew faster than raw repository creation, which is what you would expect if organizations are learning that generated code is cheap and validated code is not. The teams that skip this investment generate more code and ship the same amount.

Third, model plurality is the safe assumption. With 61 percent of Copilot users running more than one model and no model exceeding a quarter of requests, an architecture that assumes a single provider will need rework. Abstraction at the tooling layer, with models as swappable components, is now a maintenance decision rather than an architectural luxury.

Fourth, the adoption frontier is not where the marketing is. Construction and manufacturing lead on agent-opened pull requests, while sectors that talk about AI most have the fewest delegations. That gap is an opportunity for vendors willing to build for specific industries rather than for the generic developer, and for engineering leaders willing to look outside their own sector for proven patterns.

The honest summary is that GitHub Octoverse documents a platform that absorbed a workflow change faster than it absorbed a population change, and the gap between those two curves is the story. Developers did not triple; commits did. That distinction should shape how organizations hire, train, and measure, because the output per developer is what moved, and the systems around that developer determine how much of the gain is real.

Questions readers are asking

What is the biggest number in GitHub Octoverse 2026?

Merged pull requests more than tripled to 1.3 billion, and commits reached 34.3 billion, but pushes are the headline at 17.9 billion in twelve months, more than 2024 and 2025 combined.

Does the eightfold commit gap mean agents make developers eight times more productive?

Not directly. The comparison is between new repositories with and without an agent configuration file, and it reflects selection as well as behavior: agent-native teams commit more often, in smaller increments, and likely take on larger scope.

What does “40 percent of pull requests start with an agent” mean?

It means the first draft of more than 40 percent of pull requests in September came from an agent, with a human specifying the task and reviewing the result. It measures detected agent involvement, not unattended authorship.

Are agent-written pull requests worse quality?

The merge data suggests they are at least as good by that measure: 90 percent merged within two weeks in September, against about 86 percent for pull requests without detected agent involvement.

Sources
Subscription Form