- GitHub says agents wrote the first draft of more than 40% of pull requests on its platform in September, up from 17% in March.
- The report shows output rising fast but says little about who checks agent-written changes, so review capacity is the open question.
- Ask your team how many merged changes start with an agent, who reviews them, and whether secret scanning is on everywhere.
A pull request is a proposal to change a company's code. For years, one person wrote the change and a colleague checked it. GitHub's Octoverse 2026 report points to a shift in the first step. Software now drafts a large share of changes. The skill that matters more is the second step: judging what comes back.
What GitHub reported
GitHub says an agent wrote the first draft of more than 40% of the pull requests on its platform in September. In March the figure was 17%. GitHub counts 110 million such requests for September, roughly six times the March volume.
Overall volume is up sharply as well. GitHub reports 17.9 billion code pushes in the last 12 months, which exceeds 2024 and 2025 added together. Over a two-year span, merged pull requests rose to 1.3 billion, more than three times the earlier level. That works out to about 3.5 million merges a day.
GitHub also says 55% of its monthly active Copilot users now try agent-style features, against 35% a year earlier. The fastest growth, 230% year over year, is in Copilot cloud agent. There a developer hands over a whole task and leaves the agent to finish it.
How the shift works
GitHub's telemetry shows activity jumping around June. The company links this to a change in the agents themselves. Earlier tools worked within one session. The newer ones persist, follow progress over several days and can coordinate other specialised agents. They also began running work in the cloud, not only on a developer's own machine.
The developer's job shifts with this. Instead of typing the change, they describe the goal and direct the agents that carry it out.
GitHub also compared repositories created this year. Those with a configuration file for an AI coding agent averaged eight times the commits of new ones without such a file. A file like that tells the agent how to work in that project.
This compares two groups of repositories. It does not show that the file caused the extra activity.
Review is where the work moved
GitHub says agent-started pull requests more often touch several files: about 87%, against 79% for those without detected agent involvement. By September, 90% of agent pull requests merged within two weeks. The comparison group stayed near 86%.
Quick merging means a reviewer accepted the change. It does not show the change was right. The report does not measure defects. The lesson here is that a team's output now depends on how well it can check work, not how fast it can write.
GitHub points to tools built for checking. Stacked pull requests break a large change into a series of small ones that a reviewer reads one after another. That preview began on July 30, and GitHub says over 264,000 developers had tried it by September.
GitHub also says use of AI-assisted code review, and of GitHub Advanced Security, each grew by more than 90% from a year earlier. It gives no starting figures for those rises. Push protection, which checks for leaked secrets such as passwords and API keys, ran 5.8 billion scans this year, close to four times the 2025 pace.
So automated review and scanning are growing fast. GitHub does not break that growth down by who or what wrote the change. It does not say how much of an agent's work these tools check. It does not say whether a person makes the final call. That human role is the open question, and each organisation has to answer it for itself.
Beyond software teams
For the first time, GitHub sorted public repositories by industry. In administrative and support services, mining, health care and agriculture, over 40% of repositories mention AI. Yet GitHub says agents opened pull requests against fewer than 3% of them. Construction has the highest share of any sector, about 9%.
GitHub describes people who know their fields building tools with coding agents. That is a real gain. It also invites an argument of our own. If more code is written by people outside software teams, those teams may have fewer trained reviewers. GitHub does not report on that, and its own data shows agent-opened pull requests are still rare in these sectors.
What the numbers cannot tell you
The figures come from GitHub, which sells the tools it measures. It calls its agent signal "a floor", because not all agent use is detected. A language model sorted random samples of repositories into industries. GitHub also says its own forecasting models no longer agree on where growth goes next.
Questions for your team
Ask what share of your merged changes started with an agent. Ask who reviews them, and how long that takes compared with a year ago. Ask whether anyone tracks defects and security findings by where the change came from. Ask whether secret scanning is switched on for every repository, since more changes means more chances to leak a key. Finally, ask whether review capacity has grown as fast as output.
Writing code got cheaper. Trusting it did not.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: GitHub.
