GitHub Reports AI Agents Now Start More Than 40 Percent of Pull Requests

GitHub used its annual developer conference in San Francisco on October 28 to report that autonomous agents now do a large share of software work on its platform, and it introduced tools meant to coordinate that activity. The company said more than 40 percent of pull requests start with an AI agent rather than a person, up from roughly 17 percent in March. The shift sits at the center of its yearly developer review, Octoverse 2026, which covers activity from October 1, 2025 to September 30, 2026.

The platform reported 34.3 billion commits over that period, compared with 12.9 billion a year earlier, alongside 17.9 billion code pushes and 1.3 billion pull requests. Monthly figures showed about 7.38 billion commits, 220 million merged pull requests and 30.3 million new repositories. Open source and public repositories reached 566 million, more than 30 percent higher than a year earlier. GitHub also said its developer base passed 225 million people, growing at roughly one new developer per second during the year.

Agents take on coordination work

GitHub and Microsoft CoreAI head Jay Parikh told the keynote audience that the software development life cycle is no longer linear and is instead breaking apart, with agents handling much of implementation and verification. He said agent work on the platform rose from 8 billion in 2024 to 13.9 billion in 2025 and reached 45.4 billion this year. Among monthly Copilot users, the share who use agent functions rose from 35 percent a year ago to 55 percent.

To manage that volume, GitHub announced Copilot Projects, which lets an orchestrator coordinate sub-agents across one or several repositories, open pull requests and show results on a single overview screen. Developers can add requirements while agents work, and cloud agents keep running after a laptop closes. Copilot Fusion automatically selects and combines models across single, cascade and critique patterns. Parikh said Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of an Anthropic Claude model, at half the cost per task and 40 percent less time.

The company also showed Copilot Dreaming, which updates instruction files that have gone stale as code bases change, and Copilot Code Review Max for deeper inspection of proposed changes. GitHub said it rebuilt its core Git architecture to withstand heavy writing by agents and expanded infrastructure by 4 million CPU cores, 160 petabytes of high-speed storage and 50 terabits per second of network capacity.

Security pressure builds alongside speed

Faster code generation is raising concerns about review capacity. An analysis of the Octoverse findings by Security Boulevard, a cybersecurity publication and vendor, noted that agent-written code arrives far quicker than people can inspect it, and cited a separate finding that commits assisted by Anthropic's coding tool carried secrets at about twice the baseline rate, 3.2 percent against 1.5 percent across public commits. The same analysis said 64 percent of credentials found exposed were still valid more than four years later. GitHub has stressed that developers set direction, remain accountable for final results and that review demand grows with code volume.

The numbers point to a changing job profile for people in software development. Parikh said developers have moved from typing code by hand, to working with AI assistance, to managing dozens and sometimes hundreds of agents. Microsoft's CoreAI unit and GitHub are positioning the platform around speed, intelligence and quality, aiming to catch problems before code ships while letting teams connect outside tools such as Datadog and Figma through published primitives.

What to watch next

Attention now turns to whether review practices, security tooling and training keep pace with agent output. GitHub said Copilot Projects is available to all users and that Fusion is offered at a 30 percent discount through the end of November. Analysts will be watching whether the share of agent-started pull requests keeps climbing, how organizations staff code review and security checks, and whether older models of developer productivity hold up as agents take on coordination rather than only drafting.