From 17% to 40% in six months
More than 40% of pull requests opened on GitHub in September 2026 began with an AI agent, according to the company's Octoverse 2026 report published on 28 October. That is about 110 million pull requests in a single month, roughly six times the March volume, when agents accounted for 17% of the total.
Developers pushed code 17.9 billion times over the twelve months covered by the report, more than in 2024 and 2025 combined. The measurement window ran from 1 October 2025 to 30 September 2026, and GitHub detected agent activity through platform signals rather than self-reporting by users.
Bigger changes, heavier review
Agent-started pull requests modified multiple files 87% of the time, compared with 79% for those opened by people. That pattern matters for reviewers: a change that touches many files is harder to check than a one-line fix, and review capacity has not grown at the same rate as generation capacity.
GitHub has also reported repeated large-scale outages as processing load climbed. The company expanded its capacity by millions of compute cores during the year, and its engineering posts describe queueing and rate-limit changes pushed out to keep the platform stable during peak hours.
Tools are converging on the same benchmark
The report arrived alongside new product data. GitHub's Copilot Fusion, now generally available, was scored at 60.0 on the Coding Agent Index v2.0 by the independent evaluator Artificial Analysis on 28 October. A comparison configuration of Claude Code with Claude Opus 5.5 Max scored 61.7, a gap of 1.7 points.
Those numbers suggest the leading coding agents are close on benchmark tasks, which shifts competition toward reliability, review quality and cost per merged change. Engineering teams contacted by trade publications say the bottleneck has moved from writing code to verifying it, and several are testing agent-generated pull requests against stricter automated checks before a human reads them.
The next figure worth watching is review hours per merged pull request. If that keeps climbing while output rises, agents are moving work around rather than removing it.