GitHub Octoverse 2026: 7.4 Billion Commits a Month, 40% of Pull Requests Start With an Agent

GitHub's Octoverse 2026, unveiled by EVP Jay Parikh at the company's Universe conference on October 28, 2026, puts a number on something developers have felt for months: agents are now a major source of commits. In the twelve months to September, GitHub says developers pushed code 17.9 billion times and made 34.3 billion commits. In September alone the platform saw about 7.38 billion commits, and a separate report from The New Stack puts GitHub's infrastructure additions at 4 million CPU cores and 160 petabytes of storage in 2026 so far.

Simon Willison highlighted the keynote charts the same day. This post goes through GitHub's full Octoverse report, separates what the numbers show from what they do not, and covers what changes for teams that review code.

TL;DR: the numbers that matter

Is the agent boom real, or is it noise?

Both. GitHub says activity accelerated after June: from September 2025 to September 2026 commits rose about 6x, Actions jobs about 4.8x, pushes about 4.5x and merged pull requests about 3.9x. About 55% of commits since June came from brand-new repositories, which is a clue: much of the growth is fresh projects, often spun up by or for agents, not just more work in old ones.

GitHub also reports that new repositories with an agent configuration file logged about eight times as many commits as new repositories without one. That is correlation: people who set up agents are probably also the most active builders. Still, it fits the pattern in our guide to loop engineering for coding agents, where a configured agent commits in small steps all day.

A raw commit count also counts everything: abandoned branches, test repos, machine-generated churn. The earlier 2025 report counted 986 million commits for the whole year on a narrower, public basis; the 2026 figures are described as public and private activity, so do not compare them directly with older Octoverse totals.

What the agent pull request numbers say

The most useful findings are about pull requests:

• Over 40% of pull requests in September started with an agent, from 17% in March.
• Agents started about 110 million PRs in September, roughly six times March's volume.
• Agent-started PRs more often touch multiple files, 87% against 79% for others.
• Agent PRs merged within two weeks climbed from 83% in February to 90% in September; non-agent PRs held near 86%.

The merge rate deserves care. A high merge rate can mean agents produce good work, or that reviewers approve quickly, or that agents are assigned easy tasks like dependency bumps. Dependabot alone merged over 30 million pull requests in the year, about double the 2025 pace, and that is automation of a very mechanical kind. None of this measures defects shipped.

For reviewers the practical point is volume. If agent-started PRs are 40% of the queue and span more files, review capacity is the bottleneck. GitHub's own response is visible in the report: AI-assisted code review and GitHub Advanced Security usage each grew over 90% year over year, and stacked pull requests, launched as a preview on July 30, had been used by more than 264,000 developers by September.

Developers are mixing models

One section speaks directly to the model wars. Among Copilot users who selected models, 61% used more than one in September, up from 50% four months earlier. Paying users tried 29 distinct models in May and 47 in September, and no single model exceeded 25% of requests. Of developer-selected requests, 38% were general purpose, 37% deep reasoning and planning and 25% fast and lightweight. And 63% of paying users used auto model selection at least once in September.

That is the backdrop for the product GitHub made generally available the same day: Copilot Fusion, which routes between models for you. The report says Project HydraFusion, the preview we covered as model orchestration over model selection, routed nearly 60,000 sessions before launch. If you want to understand why no model holds more than a quarter of requests, see our look at the tier trade-off between Opus 5.5 and Sonnet 5.5.

Who is building: countries, languages and industries

The geography is shifting. GitHub counts more than 225 million developers, with the US (38.3M), India (32.0M), China (12.4M), Brazil (10.2M) and Japan (7.15M) the largest communities. The US and India each added over 10 million developers this year; Brazil added 3.28 million.

On languages, TypeScript ranks first among contributors, followed by Python and JavaScript; monthly contributors added in the year were TypeScript +3.62M, Python +3.23M and JavaScript +1.82M, and Rust doubled. GitHub reports AI involvement in merged pull requests at 44.1% for TypeScript, 41.1% for Python, 40.4% for Rust and 21.2% for Java, the lowest. That spread suggests agents are more at home in languages with strong types and abundant public examples, a pattern our guide to agent skills touches on when explaining why context matters.

Industry data is stranger. Public repositories grew from about 9 million in 2022 to 25 million in 2026, and "professional, scientific and technical services" rose to second place behind information. Construction has the highest share of agent-opened PRs at about 9%, while in mining, health care, agriculture and administrative services over 40% of repositories mention AI but agent-opened PRs stay under 3%. Interest in AI is not the same as agents doing the work.

A side finding: reverse engineering and clean-room rewrites

One line in the report is easy to miss. Public repositories mentioning interoperability or reverse engineering, including terms like "clean room implementation", "drop-in replacement" and "protocol re-implementation", rose over 300% to more than 130,000. GitHub notes its counting method: it matches those phrases in a repository's name, description, topics or README for repos created each year.

The likely driver is that agents make rewriting an existing tool cheap. A developer can point an agent at a documented protocol and ask for a compatible implementation in an afternoon. That raises licensing and trademark questions that tooling has not caught up with, and it is a trend to watch alongside the headline commit numbers. It also means the next wave of open-source projects may be more copies than inventions, so provenance checks on dependencies matter more than before.

The infrastructure bill

Volume has a cost. The New Stack reports that GitHub says it is handling about 7.4 billion commits per month, up from 2.9 billion in August (a figure from earlier reporting that GitHub's later numbers now supersede), and has added 4 million CPU cores and 160 petabytes of storage in 2026 so far. We could not open that report in full, so those two capacity figures are as relayed by the outlet and by the headline GitHub summary; treat them as reported, not independently confirmed.

The practical impact for users is reliability. When a platform's traffic multiplies several times in a year, outages and rate limits follow, and any CI pipeline that assumes GitHub is always fast inherits that risk. It also ties to the wider story of AI demand reshaping physical supply, as in our coverage of SK Hynix's record quarter and memory shortage.

What this means for teams

1. Budget for review, not just generation. If your agents open more PRs, add reviewers, tests or automated checks first. Merge rate alone is not quality.
2. Cap agent branches. Set limits so abandoned agent branches do not clutter repositories and CI.
3. Watch your Actions spend. GitHub ran 45.4 billion Actions jobs in the year; if agents trigger CI on every small commit, your minutes bill grows with them.
4. Protect secrets. GitHub ran 5.8 billion push-protection scans, nearly 4x the 2025 pace, which tells you how often credentials reach commits. Turn it on, and if your agents hold tokens, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
5. Read the definitions. "Started by an agent" is not "written by an agent". Ask what each statistic counts before you quote it.

What is not in the report

• No measure of defects, security bugs or reverts in agent-started PRs.
• No breakdown of agent share by vendor beyond Copilot usage stats.
• Capacity figures for CPU cores and storage come from press coverage of GitHub's statements, not the report text we could read.

Related reading

• GitHub Copilot Fusion is generally available
• GitHub Copilot HydraFusion: orchestration over selection
• Loop engineering for coding agents
• What are agent skills? A complete guide
• Claude Opus 5.5 vs Sonnet 5.5
• SK Hynix Q3 2026 and the memory shortage
• Official: GitHub Octoverse 2026 and Simon Willison's notes

Figures are accurate as of October 28, 2026 and come from GitHub's Octoverse report and press coverage; they may be revised.