GitHub Octoverse 2026: AI agents now draft over 40% of pull requests

GitHub's Octoverse 2026 report, published October 28 at the company's Universe event, quantifies a change that has been building across development teams: as of September 2026, coding agents produced the first draft of more than 40% of pull requests on the platform. According to a dev.to summary of the findings, the share was 17% in March, and September's volume works out to roughly 110 million agent-started pull requests — about six times what March saw.

What the metric actually measures

The headline number comes with fine print. GitHub classifies a pull request as "agent-started" when its initial draft came from an agent the platform can detect, so the figure captures detected usage and works as a lower bound — the true share is presumably higher. The metric also counts pull requests rather than lines of code, which means it does not support the claim that 40% of all code is now machine-generated.

Agent-assisted work is merging faster

The data also tracks how agent involvement relates to delivery speed. In September, 90% of pull requests with detected agent involvement were merged within two weeks, compared with roughly 86% of those without it. That gap is recent: in February, agent-involved pull requests actually trailed, at 83%. Agent-started changes also tend to be broader in scope, with 87% touching multiple files versus 79% for the rest.

There is an important caveat, though. GitHub measured merges, not correctness. The report contains no figures on bugs, security flaws or reverts, and a merge only establishes that a human chose to accept the change — not that the change was sound.

The platform in numbers

The report's full-year statistics cover October 2025 through September 2026: 1.3 billion merged pull requests, 34.3 billion commits and 17.9 billion pushes, with September alone contributing around 7.38 billion commits. Copilot usage patterns are shifting as well — 55% of its monthly users now engage agentic features, up from 35% a year earlier. TypeScript remains the most-used language on the platform, followed by Python and then JavaScript. Model choice is fragmented: 61% of users who select models use more than one, and no single model holds more than 25% of usage.

Why it matters

The 40% figure marks the point at which agents stopped being autocomplete-style helpers and became first-draft authors on a large slice of real engineering work. Humans remain the gatekeepers — every one of those pull requests was merged by a person — but the framing relayed by dev.to points to where the pressure now lands: the constraint has shifted from producing code to vetting it. If 40% is only the detectable floor and the trend line is still steep, then the scarce resource on engineering teams is no longer code production but the attention and judgment required to evaluate machine-drafted changes. Teams that invest in review capacity, standards and tooling for scrutinising agent output are the ones positioned to benefit; teams that treat merging as a formality inherit exactly the risks — quality, security, maintainability — that this report does not measure.