开发活动创新高,超 40% 的 PR 由智能体起草
Key Highlights
GitHub released its 2026 Octoverse report, covering October 2025 through September 2026. The takeaway is clear: developer activity and agent penetration both hit new highs at the same time, and the two trends are now entangled.
What Happened / The Numbers
Over the 12-month window, developers pushed code to GitHub 17.9 billion times — more than the combined totals of 2024 (5.2 billion) and 2025 (6.7 billion). Commits jumped to 34.3 billion, more than triple two years prior. Merged pull requests reached 1.3 billion, averaging 3.5 million per day. GitHub Actions ran 45.4 billion jobs, about 1,440 every second. The scale alone shows the platform is not just growing; the shape of work on it is changing.
Technical Details
Agents are now deeply embedded in the workflow. As of September 2026, over 40% of pull requests are first drafted by an agent, and a large share of commits carry agent-generated code. That is the headline shift: humans still review and merge, but the first draft — the part that used to consume the most junior engineering time — is increasingly machine-written. The report frames this as a new baseline for how software gets made, not a future scenario.
vs Competitors
GitLab and Bitbucket tell similar stories from narrower bases, but GitHub's scale gives its numbers outsized signal value. The agent-drafted PR metric matters because it is a concrete, measurable proxy for AI's entry into the core loop of software development, beyond chat assistants and doc generation. When 4 in 10 PRs start with an agent, the question flips from "will AI write code?" to "how do we review AI-written code well?"
Industry Impact / Use Cases
For engineering leaders, Octoverse 2026 is a planning document: budget for agent tooling, invest in review discipline, and treat agent-written first drafts as the norm. For developers, the junior-grunt work shrinks while the review-and-architecture bar rises. The open risk is quality and provenance — more code, drafted faster, demands better testing and clearer attribution so a team still knows what it shipped and why.
Practical Notes
For engineering leaders, Octoverse 2026 is less a celebration than a planning document. The 40%-of-PRs-agent-drafted number should trigger a budget line for agent tooling and, just as importantly, for review discipline. If machines draft more than humans, the bottleneck moves to review quality, test coverage, and provenance. Teams that treat agent output as automatically safe will accumulate debt; teams that build review gates will scale cleanly.
The 17.9 billion pushes also hint at a measurement problem: raw activity is up, but "more code" is not "better software." Leaders should track defect rates and rework alongside volume, or the metric becomes a vanity signal. For developers, the junior-grunt work shrinks while the architecture and review bar rises — a healthy shift if you invest in the latter. The bottom line: the report confirms AI is now in the core loop of software development. The open question is whether review and testing practices evolve as fast as generation does, because that gap is where the next wave of incidents will come from.
What to Watch
Watch the agent-drafted-PR percentage next year; if it crosses 50%, the review layer becomes the most valuable part of the developer stack, and tooling for agent-code review will be the next startup wave. Also watch whether defect rates rise with volume — that is the tell for whether teams are reviewing or rubber-stamping. The bottom line for 2027: generation is solved; verification is the frontier, and Octoverse will keep measuring the gap.
Takeaway
Octoverse 2026 confirms AI is inside the core loop of software development, not hovering around it. The work left for humans is review, architecture, and verification — exactly the skills teams should be investing in. Generation is solved; trust is the frontier, and the data will keep measuring the gap.
Context
The report's timing matters: it lands as every major platform races to embed agents in the developer workflow, and the 40% figure gives teams a concrete benchmark to measure themselves against. If your org is below that, you are behind the curve; if above, the risk is review quality, not volume. Either way, the number is now a shared reference point for the industry.