GitHub’s Octoverse 2026 report, published on 28 October 2026, puts a clear number on a change developers have felt all year. By September 2026, more than 40% of pull requests on GitHub began with an agent’s first draft, up from 17% in March. That meant about 110 million agent-started pull requests in September alone, roughly six times the March volume. The report covers 1 October 2025 to 30 September 2026 and spans public and private repositories.
The numbers behind the shift
The scale of activity behind that headline is striking. Developers pushed code 17.9 billion times during the year, more than in 2024 and 2025 combined, while commits reached 34.3 billion. Merged pull requests more than tripled to 1.3 billion, or about 3.5 million a day, and GitHub Actions ran 45.4 billion jobs, roughly 1,440 every second.
GitHub now serves more than 225 million developers, led by the United States with 38.3 million and India with 32 million. Among repositories created this year, the average repo with an agent configuration file logged eight times as many commits as the average new repo without one. That is an association, not proof that agents alone caused the difference, but it shows how closely agent setup and intense activity now travel together.
Review is the new bottleneck
When agents write the first draft, human review becomes the work that shapes the result. In September, 90% of pull requests with detected agent involvement merged within two weeks, compared with about 86% without detected involvement; in February, the agent figure was 83%. Agent-involved pull requests also more often touched several files, about 87% versus 79%, so reviewers are increasingly asked to judge broader changes rather than small, isolated edits.
A merge shows that a change was accepted, not that it was correct, so reviewers still need to read diffs, check tests, and question assumptions. The same discipline matters when building agentic AI systems, where a clear task and a careful review loop matter as much as the first draft. It is also worth remembering that GitHub can only count the agents it can detect, so its figures are a floor and the real share may be higher.
Model choice is now a per-task decision
Developers are not settling on one model. In September, 61% of GitHub Copilot users who selected a model themselves used more than one model across their sessions, up from 50% four months earlier, and no single model carried more than 25% of requests. That spread suggests teams are matching models to tasks instead of treating one model as a default for everything.
Among monthly active Copilot users, 55% now use agentic features, up from 35% a year earlier. The practical skill is shifting from memorising one tool to writing a clear task, choosing the right model for it, and verifying the output. To see how these agents are set up and used, read GitHub’s Copilot agents documentation.
What this means for your career
Start with habits that make agent work easier to trust. Learn to write precise task descriptions, keep agent work in small, reviewable chunks, and require tests and CI checks before merging. Keep agent configuration files in the repo so behaviour is repeatable, and build the review habit: read every diff as if a junior colleague wrote it, even when the code looks polished at first glance.
Those habits show up in real-world AI success stories, where teams combine ambitious automation with human oversight, and they are essential for anyone shipping production AI systems. Developers who can direct agents, spot weak code, and own the final result will be in demand; developers who paste output unchecked will not.
The caveats worth keeping
These figures describe GitHub’s own platform, not the whole software industry, and they measure observed usage, not code quality. Treat the headline as a signal about how work starts, not a verdict on how good agent-written code is. The sensible response is to adopt agents deliberately, measure your own merge, revert and bug rates, and keep humans in charge of review.
The first draft is increasingly written by an agent; the judgment, the tests and the responsibility still belong to the developer. That is the skill to invest in now.
