In September 2026, 90% of pull requests on GitHub with detected agent involvement merged within two weeks, up from 83% in February.

That comes from GitHub's Octoverse 2026 report, published on 28 October and built on GitHub's own telemetry across public and private repositories. Pull requests without detected agent involvement held at about 86% in both months. The same report found that by September more than 40% of pull requests started with an agent's first draft, 110 million of them that month. GitHub frames review as "how developers shape what comes back", and notes that use of AI-assisted code review grew more than 90% year over year.

What the merge-rate figure counts

The figure measures one event: whether a pull request merged within two weeks. It is a platform count, not a survey, so it records what happened rather than what developers say. GitHub says its detection of agent involvement undercounts, since not every agent leaves a signal it can see.

What it doesn't record is how the change got there. A pull request whose agent draft was merged untouched and one that people rewrote line by line over ten days both count as merged. The Octoverse report doesn't say how much of an agent's first draft survived to the merge, or how many commits people added on the way.

What Linux and Microsoft said about AI patches

On 1 November, Linus Torvalds released Linux 7.4-rc1 and called it a big merge window, with almost 20,000 commits against the more common 15,000 or so. He said it was "full of fixes for issues found by AI", and that "more and more the patches have been suggested by AI too, although they often get rewritten". He gave a second reason for the size as well: 7.4 is expected to be the next long-term support release, which tends to draw more changes.

On 11 November, Microsoft's FORGE Lab described what it took to land kernel fixes for bugs its AI scanner found. Two weeks of scanning produced 26,502 raw findings, which agents and experts narrowed to 621 cases with proof-of-concept evidence and 59 upstream reports. Agents helped draft the first fixes, but Microsoft wrote that "human validation, subsystem understanding, and iterating with maintainers during review remained essential". Of the 38 fixes merged so far, half went in on the first submission and the rest after one to four revisions. Reporting the 59 took more than four person-months.

GitHub's own product points the same way. On 5 November it added a count of Copilot-reviewed pull requests that merged with no person requesting changes after Copilot's first review, so that teams can see "whether reviewers still request changes after Copilot reviews". That measures Copilot's reviews rather than agent-written code, but it is GitHub treating a merge on its own as too blunt to say whether people had to step in.

Why a merge doesn't show how a change arrived

A merge rate tells you that agent work is landing, not that agents' drafts were right.

A merge rate tells you that agent work is landing, not that agents' drafts were right. Microsoft's numbers show the gap. Count only merges, and all 38 kernel fixes look the same. Count revisions, and half needed more human work before maintainers would take them. Torvalds's remark describes the same pattern across a much larger body of work, in a project where AI patches arrive by email rather than as GitHub pull requests.

That suggests the rise from 83% to 90% may say as much about review as about agents. More developers are reviewing agent work, with more tools for it, and a reviewer who rewrites a weak draft until it is good enough still produces a merge. The figure can rise because agents got better, or because people got faster at fixing what agents produce, and the count alone can't tell the two apart.

Neither kernel account is a sample of GitHub's 110 million agent-started pull requests, so they show what a merge can contain. They can't say how often a merged agent draft was rewritten.

Counting the human commits behind agent merges

Take twenty agent-started pull requests your team merged last month and count the commits people added after the agent's last one. If most have several, your merge rate is measuring your reviewers' effort as much as your agents' output.

Sources

  1. GitHub, "Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents", 28 October 2026 — https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/
  2. Linus Torvalds, "Linux 7.4-rc1", Linux kernel mailing list, 1 November 2026 — https://lkml.iu.edu/hypermail/linux/kernel/2611.0/00796.html
  3. Microsoft, "Lessons from 59 Linux Kernel Bug Reports", FORGE Lab, 11 November 2026 — https://www.microsoft.com/en-us/security/blog/2026/11/11/lessons-from-59-linux-kernel-bug-reports/
  4. GitHub, "Usage metrics API adds code review details and addressed rate", 5 November 2026 — https://github.blog/changelog/2026-11-05-usage-metrics-api-adds-code-review-details-and-addressed-rate/