GitHub says AI agents now start more than 40% of pull requests
GitHub's 2026 Octoverse counts 1.3 billion merged pull requests in a year, more than double 2025, and says agents now open a large share of the work.
GitHub's annual Octoverse report says developers merged 1.3 billion pull requests in the 12 months to September 2026, about 3.5 million a day and more than double the year before. The bigger change is who starts the work. GitHub says more than 40% of pull requests in September were opened by an AI agent, up from 17% in March. The report is GitHub's own count of activity on its platform, not a measure of all software written everywhere.
- 1.3bnmerged pull requests in the 12 months to September 2026
- 17% → 40%+agent-started share of pull requests, March to September
- 32.0mdevelopers in India on GitHub, second to the US
The numbers GitHub reported
GitHub counted 17.9 billion pushes and 34.3 billion commits over the year. Merged pull requests rose from 402 million in 2024 to 538 million in 2025 and then to 1.3 billion in 2026. Public contributions reached 2.5 billion, up from 1.12 billion. GitHub Actions, its automation service, ran about 1,440 jobs a second, 124 million a day.
The platform says it now serves more than 225 million developers. The United States has the most at 38.3 million, followed by India at 32.0 million; GitHub says each added more than 10 million this year. Brazil added 3.28 million, the third-largest increase.
Automation shows up even without large models. Dependabot, which opens pull requests to update vulnerable or outdated dependencies, had more than 30 million of its pull requests merged, about double the pace of 2025. A great deal of maintenance work was already machine-started before the current wave of coding agents.
What an agent-started pull request is
A pull request is a proposed change to a codebase, offered for review before it is merged. In the classic workflow a developer writes code on a branch, opens a pull request, and colleagues read it, run the tests and ask for changes. The request is the unit where quality control happens.
An agent-started pull request is one where an AI system wrote the change and opened the request, usually because a person assigned it a task: fix this bug, add this test, update this library. A person is still expected to review and merge it. GitHub's figure counts who opened the request, not whether the change was good, accepted unchanged, or later reverted.
That distinction matters for reading the 40% figure. It tells us agents are now a normal part of how work enters the review queue. It does not tell us that 40% of software is written without people, or that the agents' changes are of the same quality as a human's. The report does not claim either.
Why the volume more than doubled
When starting a change becomes cheap, people start more changes. Agents can be pointed at small tasks a human would leave for later, such as renaming, adding a missing test or fixing a lint warning. Each becomes its own pull request. The count rises faster than the amount of new functionality.
Volume moves the bottleneck to review. A team can now produce more proposed changes than it can carefully read. That puts weight on automated tests, continuous integration and clear ownership, because a reviewer who cannot trust the test suite will either slow down or approve changes they have not fully understood.
The Actions figures point the same way. Every pull request typically triggers builds and tests. Running 124 million jobs a day is the infrastructure cost of a workflow in which code is proposed constantly and checked by machines before a person looks at it.
What this means for the job
The skills that rise in value are the ones around the code. Writing a precise task description is now a technical skill, because an agent does what it is asked. Reading a diff quickly and spotting what it breaks is more valuable than typing the same diff. Designing tests that catch a wrong change is the cheapest protection a team has.
Junior roles change shape rather than vanish. Much entry-level work used to be exactly the small tasks agents now start. A new graduate who can review, test and explain a change will find more use than one who can only produce it. Understanding the system, its data and its failure modes becomes the entry ticket.
Open-source contribution remains one of the best ways to show those skills, because every review comment is public. A student whose pull requests show careful testing and clear reasoning is demonstrating the work employers increasingly need.
Why it matters to engineering students in Nepal
Nepal's software industry is mostly export services, building and maintaining code for clients abroad. Those clients use the same tools. If their own teams start 40% of changes with agents, they will expect outsourced teams to work the same way and to be paid for review, testing and judgment rather than hours of typing.
GitHub's growth figures for India show how quickly the region's developer base is expanding. The report mentions a Nepali developer's flood-alert project as an example but gives no figures for Nepal. A Nepali engineer competes in that same pool, and the advantage will go to those who can direct and check automated work well.
There is also a cost note. Agent tools and cloud automation are usually billed in dollars. A Nepali student or small studio pays for them through cards with annual dollar limits, a constraint covered in this news section's story on Google Play merchant accounts.
What to study if this interests you
Software engineering courses are the direct fit: version control, code review, testing strategies and continuous integration. Practise on a real repository. Set up a test suite and a CI pipeline for a project, then ask an AI assistant to make a change and review it as strictly as you would a colleague's.
Distributed and cloud computing explains the infrastructure under the numbers, from build runners to artefact storage. Project management courses help with the other half of the shift: deciding what work to assign, how to split it, and how to know when it is done.
Keep the measurement habit. GitHub's report counts activity on GitHub. For your own projects, measure what matters: how many changes broke something, how long review took, and how many bugs reached users. Those are the numbers that tell you whether automation helped.
Words in this story
- Pull request
- A proposed change to a code repository, reviewed and tested before it is merged into the main branch.
- Coding agent
- An AI system that can take a task, edit code across files, run tools and open a pull request for human review.
- Continuous integration
- Automatically building and testing every proposed change so that problems are caught before merging.
Where this comes from
- GitHub Blog, 28 Oct 2026
Written in our own words; no sentence is copied from these reports. Researched with AI assistance and added 17 October 2026; no member of faculty has reviewed this story yet. If you spot a mistake, call 01-5091616 and we will correct it and say so.