🔍 Read the full analysis: AI Agents Now Start 4 In 10 Pull Requests On GitHub. The Bottleneck Has Moved To Review on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
GitHub’s Octoverse report says detectable coding agents started more than 40% of pull requests in September 2026, up from 17% in March. GitHub says 90% of those requests were merged within two weeks, but merging does not establish code quality; review capacity, costs and human responsibility are now central concerns.
Coding agents started more than 40% of GitHub pull requests in September, up from 17% in March, according to the company’s Octoverse report released October 28. The reported total was 110 million agent-started pull requests in one month, a shift that puts more pressure on people and systems responsible for reviewing proposed code before it is accepted.
GitHub defines an agent-started pull request as one whose first draft came from a coding agent the platform can detect. The company says the resulting count is a floor because not all agent use may be detected. The report’s figure includes public and private repositories unless otherwise specified; GitHub did not provide a split between them. A request still counts as agent-started if a person later rewrites part of the draft.
GitHub reported that 90% of detected agent-started pull requests were merged within two weeks in September, compared with 83% in February. The rate for pull requests without a detected agent remained about 86%. Those figures describe whether requests were merged, not whether the changes were correct, safe or free of later rework.
The report also describes a sharp rise in platform activity: GitHub counted 1.3 billion merged pull requests in the 12 months ending in September, compared with 538 million in 2025. It said 55% of Copilot monthly active users used agent features, up from 35% a year earlier. Agent-started requests were more likely to touch multiple files: 87% changed more than one file, against 79% of requests without a detected agent.
Infographic 01 · The agent share
Four in ten pull requests now start with an agent.
GitHub’s own telemetry, March to September 2026. The first draft comes from an agent; review and merge stay with people.
Merged within two weeks
Share of pull requests merged within 14 days
Merged fast is not the same as correct. GitHub publishes no revert or defect rate.
Bigger changes to review
Share of pull requests that change more than one file
More files per change means more for a reviewer to check.
What GitHub counts: an “agent-started pull request” is one “whose first draft came from a coding agent that we can detect.” Undetected agents are missing, so GitHub calls the figure a floor. Metrics cover public and private repositories unless noted.
Source: GitHub Octoverse 2026, 28 Oct 2026 (text, chart captions, glossary and methodology). All figures are GitHub’s own.
Powered by Thorsten Meyer AI
Review Becomes the New Constraint
When agents produce more proposed code, the work does not end at drafting. Each change still needs someone to judge whether it matches the task, fits the surrounding system and should be approved. Review time and reviewer attention can become limiting resources even when code generation is fast, particularly when requests span several files.
GitHub’s product announcements reflect that shift. The company introduced Copilot code review Max, which uses several AI reviewers from different model providers and combines their feedback; it also said Copilot approvals can count toward required approvals. Copilot Fusion routes different steps to selected models. These features may help teams handle review volume, but they do not establish that generated code is reliable or that AI review can replace human judgment.
Review has a direct budget dimension as well. GitHub estimates an average Copilot review costs $0.25 for Lite, $1.50 for Balanced and $10 for Max, plus GitHub Actions minutes. Organizations will need to weigh those costs against reviewer time and track outcomes such as rework, reversions and incidents—not just the number of pull requests created or merged.
Products Worth Considering
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How GitHub Counts Agent Work
The figures come from GitHub’s 2026 Octoverse report, its annual account of activity on the platform. This edition covers October 1, 2025, through September 30, 2026. GitHub says it recalculated figures for 2024 and 2025 using updated definitions and time windows, so older Octoverse totals are not directly comparable with the revised series.
The report’s terminology varies: it uses “agent-started,” “agent-generated,” “agent-authored” and “agent-opened” in different places. Its glossary definition is narrower and more precise: the first draft must come from an agent GitHub can detect. That definition matters when interpreting the headline share; it is not a count of every line of code written by AI, nor a measure of how much humans subsequently changed.
GitHub also reported that activity rose sharply around June, when it said agents were moving from single-session tools to persistent cloud-based tools that can work across days. Use of Copilot’s cloud agent grew 230% year over year, according to the report. Those are company-reported platform metrics, not an independent assessment of the quality or productivity impact of agent-generated code.
“A pull request whose first draft came from a coding agent that we can detect.”
— GitHub, in its Octoverse report
What Merge Rates Cannot Show
GitHub’s figures do not show whether merged agent-started changes were correct, how often they needed substantial human rewriting, or whether they later caused bugs or security issues. The reported two-week merge rate is a workflow measure, not a quality assessment. The source material provides no independent audit of the agent detection method or code outcomes.
The platform also did not publish an open-source versus private-repository breakdown for the agent share. Because GitHub says the count includes only agents it can detect, the exact proportion of pull requests that began with any coding agent remains unknown. And while the report gives price estimates for Copilot reviews, it does not establish how those costs compare with human review time across different teams.
Teams Must Measure Review Outcomes
GitHub’s newly announced review and routing tools will give teams options for handling agent-generated proposals, but the report does not specify when every feature will be available to all customers. The practical next step for organizations is to set rules for when an AI review may count toward required approvals and which changes still require direct human sign-off.
As agent use expands, teams will need to monitor review workload, rework, reversions and production incidents alongside request and merge counts. Developers will still need to check diffs critically and give agents clear tasks. GitHub’s report establishes the scale of the change in its own detected activity; whether that activity produces lasting productivity gains without added quality or maintenance costs remains to be measured.
Key Questions
What does GitHub mean by an agent-started pull request?
GitHub counts a request when its first draft came from a coding agent the platform can detect. A person may edit the draft afterward and it still counts. GitHub says the measure is a floor because some agent use may go undetected.
Does the 90% merge rate mean agent code was correct?
No. The figure means 90% of detected agent-started requests were merged within two weeks in September, according to GitHub. It does not measure correctness, later defects, security or rework.
How much does an AI code review cost?
GitHub estimates an average Copilot review at $0.25 for Lite, $1.50 for Balanced and $10 for Max, with GitHub Actions minutes charged separately.
Who is responsible for code written by an agent?
GitHub says developers remain accountable for code that ships. The source material also says its cloud agent cannot approve or merge its own pull requests, and the person who requested a change cannot approve it either.
The report’s methodology says metrics include public and private repositories unless otherwise noted, and the agent-share figure carries no exception. But GitHub counts only agents it can detect and did not publish a public-versus-private breakdown, so the number does not capture all agent use with certainty.
Source: ThorstenMeyerAI.com
