GitHub Octoverse 2026: Over 40% of PRs Start with an AI Agent's First Draft, 4 Takeaways for Beginners

GitHub is a platform where developers store code and work on projects together. Every year it publishes an annual report called Octoverse that sums up what developers did over the past year. The 2026 edition’s title states the theme outright: developers are building with agents.

17.9 Billion Pushes in a Year, More Than the Previous Two Years Combined

First, the scale. Over the past 12 months, developers pushed code to GitHub (a push sends your finished changes up to the platform) 17.9 billion times, more than in 2024 and 2025 combined.

GitHub’s own reading is that activity bent sharply upward around June, just as agents went from “a new toy to try” to “a partner to work alongside.” Over the same period, models got better, the tools around them multiplied, and agents began tracking progress across days and running work in the cloud.

Takeaway 1: Over 40% of Pull Requests Start with an Agent’s First Draft

On GitHub, someone who wants to change a project submits a pull request (PR, a request to merge a change), and a project maintainer reviews it before merging. The first draft used to be written by a person.

The report’s numbers: by September 2026, more than 40% of PRs started with an agent’s first draft, up from 17% in March. This counts only agent involvement that GitHub can detect.

In the report’s words, the first move is now directing an agent, and review is how people shape what comes back.

An everyday analogy: you used to cook the meal yourself, and now you first tell a kitchen assistant what to make, then taste and adjust.

Takeaway 2: New Projects with an Agent Configuration File Log About 8x the Commits

Among projects created this year, those with an agent configuration file logged on average 8 times as many commits as those without one.

A configuration file is a note kept inside the project that tells an AI coding agent how the project works. A commit is a saved snapshot of changes.

This compares the averages of two groups of projects, and the report does not say the configuration file caused the 8x. GitHub’s reading is that people and teams building with agents from day one move at a markedly different pace.

The practical point for beginners: if you want an agent to be useful, start by writing down your project’s rules clearly. To get started with Claude Code, see the Claude Code Beginner Guide.

Takeaway 3: Most People Use More Than One Model

The report also looked at how GitHub Copilot users pick models. In September 2026, among users who chose a model themselves, 61% used more than one model across different tasks.

Their choices spread across three kinds of models: general purpose, deep reasoning and planning, and fast and lightweight. Planning, writing, and checking each place different demands on a model, so developers match the model to the task instead of betting on a single model.

This echoes the idea of several agents dividing up the work. See What Is a Multi-agent Orchestration Framework.

Takeaway 4: People Outside the Software Industry Are Building with Agents Too

This is the first time the report counts public projects by industry. Projects serving mining, construction, health care, and agriculture have each grown several times over since 2022.

The report gives three examples:

A small team at a manufacturer built a way to turn a photo of your tools into a custom-fit, 3D-printable drawer organizer
A university lab in Copenhagen built a pipeline that classifies video of chickens at play
A novelist built a free app that makes a manuscript look like a book as you type

In the report’s words, the people closest to the problem are building the solution themselves, with agents helping along the way.

New Faces on the List Are Mostly AI Coding Tools

The report lists this year’s fastest-growing projects by contributors. Among projects from companies, the list includes Anthropic’s claude-code, OpenAI’s codex, GitHub’s copilot-cli, and Google’s gemini-cli. The report notes that these projects are mostly tools developers work with directly, including AI coding agents, editors, and the programs that orchestrate agents.

What Hasn’t Changed

The report closes with a reminder: however much activity there is, nothing changes who decides what the code should do. Someone still has to come up with the idea, and someone still has to decide whether it works and whether it belongs.

A one-line takeaway for beginners: the agent does the work, and you say clearly what you want and understand what comes back. Before you hand an agent any permissions, read up on its risks in Security Risks of AI Agents.

FAQ

What is Octoverse?

Octoverse is GitHub’s annual developer activity report. This edition covers October 1, 2025 to September 30, 2026, and turns a year of commits, merges, model usage, and more into trends.

Do I need this report if I have never written code?

If you want to know how far AI agents are used in real work, this report is rare large-scale data. It also notes that many people outside software are using agents to build tools for their own fields.

If agents write the first draft, are engineers out of work?

The report concludes the opposite: someone has to come up with the idea, and someone has to judge whether the result works and whether it should be accepted. Both still belong to people. The developer’s role is shifting toward directing agents and reviewing what they hand back.