GitHub’s annual Octoverse report for 2026 describes a turning point in how software is made, with review, not writing, becoming the central human task: by September, more than 40% of pull requests on the platform started with an agent’s first draft, up from 17% in March. The report, published on October 28, covers the year to September 30 and records record volumes across pushes, commits and merged changes.
A year of record volume
Developers pushed code to GitHub 17.9 billion times in the past twelve months, more than in 2024 and 2025 combined, according to the report. Commits across public and private repositories reached 34.3 billion, more than three times the level of two years earlier. Merged pull requests more than tripled to 1.3 billion, an average of about 3.5 million every day.
GitHub Actions, the platform’s automation system, ran 45.4 billion jobs in the year, or about 124 million a day. That pace, roughly 1,440 jobs every second, shows how much routine work now runs without a developer watching it. Dependabot, which opens pull requests to update dependencies, merged more than 30 million changes, about double its pace in 2025.
Activity did not rise evenly through the year. GitHub says its telemetry shows steady growth from January, then a sharp surge around June, when better models, stronger tooling and more capable agents arrived together. Agents moved from single-session helpers to persistent tools that could track progress across days and orchestrate other specialised agents.
From June onward, about 55% of all commits came from brand-new repositories, as new projects started with agents. Yet established developers were not standing still: those who pushed code in the past year had been on GitHub for an average of 36 months. The average developer also worked across 10% more repositories than a year earlier.
Open-source contributions reached 2.5 billion in 2026, more than double the 1.12 billion recorded last year. GitHub points to individual maintainers carrying extraordinary loads with agentic workflows, including one contributor with more than one million contributions across public repositories. Scale, in this report, is a story about people as much as machines.
The first draft now belongs to agents
The headline change is who writes the first draft. A pull request is a proposed change to a codebase, submitted for review before it is merged. By September, more than 40% of pull requests started with an agent’s first draft, up from 17% in March. That is about 110 million in September, roughly six times March’s volume.
Agent-started pull requests also tend to be broader. About 87% of them changed multiple files, compared with about 79% of pull requests without detected agent involvement. That breadth matters for reviewers: a change touching many files can be harder to check than a small, focused fix, even when each individual change looks reasonable on its own.
Among Copilot’s monthly active users, 55% now use agentic features, up from 35% a year ago. The fastest-growing mode is the cloud agent, where a developer assigns a task for independent work, up 230% year over year against 85% for agent chat. GitHub’s Copilot now runs projects across repositories, the same move from answering questions to carrying out whole tasks.
GitHub is careful about what its agent figures can and cannot show. Its glossary defines an agent signal as an observed sign of agent involvement, and warns that it is a floor, since not all agent use may be detected. The true share of agent-involved work could therefore be higher, while the comparison group may itself contain undetected agent use.
That caveat cuts both ways. If detection misses some agent work, the 40% figure understates the change. But it also means the comparison between agent and non-agent pull requests is not perfectly clean. Readers should treat the gap between the two groups as indicative, not as a precise measurement of an agent’s effect on any single change.
Review becomes the real work
As agents take over first drafts, review becomes the place where developers shape what comes back. In February, 83% of agent-involved pull requests merged within two weeks, slightly behind the roughly 86% for pull requests without detected agent involvement. By September, agent pull requests had reached 90%, while the comparison group held at about 86%.
A merge is not proof of quality. It shows that a change was accepted, not that it was correct, secure or designed. GitHub’s framing points the same way: for a growing share of work, the first move is now directing an agent, and review is how developers shape what comes back. The bottleneck moves from writing code to judging it.
Tooling for that judgement is growing fast. Stacked pull requests, which split related work into smaller changes reviewed in sequence, launched in preview on July 30 and had been used by more than 264,000 developers by September. Use of AI-assisted code review and GitHub Advanced Security each grew by more than 90% year over year.
Security automation is scaling with the volume. Push protection, which screens for exposed secrets such as API keys and passwords, ran 5.8 billion scans in the year, nearly four times its 2025 pace. When agents can open changes at machine speed, the checks that catch secrets and vulnerabilities must run at machine speed too.
Model choice moves at developer speed
Developers are not settling on a single model. In September 2026, 61% of GitHub Copilot users who selected a model themselves used more than one model across their sessions, up from 50% four months earlier. Paid Copilot users drew on 29 distinct models in May and 47 in September, and no single model carried more than 25% of requests.
Out of every 100 requests, GitHub says 38 went to general-purpose models, 37 to deep reasoning and planning models, and 25 to fast and lightweight models. Those figures cover only developer-chosen requests, excluding automatic selection, staff use and background models. GitHub’s Fusion picks the right model per task, routing each step to the best-suited model.
New models are used before most developers have tried them, GitHub says, with early adopters driving a larger share of requests than their share of users. When a model’s request share falls, the share of developers still using it falls more slowly, suggesting models stay in rotation after their peak. Developers blend kinds of models rather than committing to one.
Automatic selection is common too: in September, 63% of paying Copilot developers used auto model selection at least once, letting Copilot pick the model for the task. The picture is therefore mixed, with deliberate choice and delegated choice side by side. A developer may hand-pick a model for a hard problem and let the system choose for routine work.
Language choice shows a similar pattern. TypeScript remained the most used language by contributors, ahead of Python and JavaScript, and Rust grew fastest by percentage among the top languages, doubling in a year. GitHub raises a new question: when a developer does not specify a language, the language an agent picks may become a fourth influence on what gets built.
What this changes, and what it does not
The most important change in Octoverse 2026 is not a benchmark score. It is a change in the default starting point of software work. Writing the first draft, long a developer’s core task, is increasingly directed rather than typed. That changes what skill looks like: specifying a task well, and reviewing the result critically, matter more than before.
That shift also changes where risk concentrates. A single developer can now propose far more changes, across more files and repositories, than before. If review does not keep pace, errors, flaws and design debt can enter codebases faster. The report’s review and security figures are GitHub’s answer to that worry, not proof it is solved.
The volume figures need the same care. Pushes, commits and merged pull requests are activity measures, not quality measures. A repository with an agent configuration file logging eight times as many commits as one without it may be moving faster, or generating more churn. GitHub presents the association without claiming agents alone caused it, and readers should hold that line.
There is a competitive angle the report only hints at. When model choice spreads across dozens of models and no single model carries more than a quarter of requests, no provider can take developer loyalty for granted. Copilot’s value shifts from the model underneath it to the workflow around it: routing, review tooling, security checks and the agent harness itself.
For rival coding tools, the message is uncomfortable. GitHub’s scale gives it a view of developer behaviour that smaller tools cannot match, and its report turns that view into a marketing asset. Yet the same data shows developers spreading their model use widely, which suggests the underlying models remain interchangeable in many tasks, even if the workflow platform is not.
For teams hiring and training developers, the report points to a practical conclusion. Junior developers will need to learn review, testing and specification earlier, because those are the skills an agent-heavy workflow rewards. Senior developers, meanwhile, become the bottleneck: their judgement is what stands between a flood of plausible drafts and a codebase worth keeping.
Why It Matters
Octoverse matters because it measures a change easy to exaggerate and hard to quantify. Claims about agents replacing developers have outrun evidence for years. Here, the world’s largest code host puts a number on one concrete step: more than four in ten pull requests now begin with an agent’s draft, up from fewer than two in ten in March.
It matters for software quality too. If review is now the main human contribution, then review tooling, test coverage and security scanning become the critical infrastructure of software development. Organisations that invest only in generating more code, without investing in checking it, are likely to feel the cost in bugs and incidents first.
It matters for the model market as well. A world in which developers routinely use several models, and let the system route between them, rewards models that are good at specific steps rather than models that claim to do everything. That fragmentation may suit GitHub, which sits above the models, more than any single model provider.
Finally, it matters because the numbers come with stated limits. GitHub says its agent signal is a floor, its industry figures are estimates from sampled repositories, and its forecasts disagree widely about the future. A report that shows its working, even imperfectly, is more useful than a cleaner story that hides its uncertainty.
Background
Octoverse is GitHub’s annual census of activity on its platform, covering the twelve months to September 30. The 2026 edition counts public and private repositories unless noted, recalculates prior years on the same definitions, and draws on telemetry, developer surveys of usage patterns and sampled public repositories for its industry breakdown.
Source: GitHub Octoverse report
0 Comments