What changed
On October 28, 2026, GitHub published Octoverse 2026 covering October 1, 2025 to September 30, 2026. It reports that more than 40% of September pull requests started with a detectable agent first draft, compared with 17% in March, representing about 110 million agent-started PRs in September. The report also counts 17.9 billion pushes, 34.3 billion commits, 1.3 billion merged PRs and 2.5 billion public/open-source contributions across its measurement year. Agent-involved PRs were merged within two weeks 90% of the time by September versus about 86% without detected agent involvement. These are GitHub's own platform measurements; the report does not establish that agents caused the overall activity increase or improved code correctness.
Why it matters
The practical shift for software teams is that code review and verification may become a more important constraint than producing a first draft. Agent-started changes are often multi-file and arrive quickly; teams need tested review boundaries, CI gates, clear ownership and checks for insecure or unmaintainable changes. The report also shows developers selecting multiple models and increasing demand for agent-compatible tools, but GitHub's telemetry detects only observable agent signals and does not measure off-platform work. This is evidence of changed workflow on GitHub, not an independent causal study of developer productivity.
Agent first drafts now account for a large measured share of PRs
GitHub says detectable agent-started PRs rose from 17% in March to more than 40% in September 2026, about 110 million in September. The definition is a PR whose first draft came from an agent GitHub can detect. It is not a census of all AI-assisted work: undetectable agents and AI used only inside human-authored changes are not necessarily counted.
Merge speed rose, but it is not a quality benchmark
By September, GitHub reports 90% of agent-involved PRs merged within two weeks, versus about 86% for the comparison group; in February, agent PRs were at 83%. Agent-involved PRs more often changed multiple files (87% versus 79%). The observational figures are not controlled for task difficulty, team practice or code correctness. They cannot establish that agents make changes safer or better.
Activity counts are huge and definitions changed
The report lists 17.9 billion pushes, 34.3 billion commits, 45.4 billion Actions jobs and 1.3 billion merged PRs. GitHub recalculated 2024/2025 comparisons using the new October–September window and public/private scope, unlike some earlier Octoverse editions. A push can carry multiple commits; Actions jobs are not minutes. Rising raw counts can reflect automated activity and do not map directly to independent developer output.
Developers use multiple models rather than a single default
Among GitHub Copilot users who explicitly selected models, 61% used more than one in September, up from 50% four months earlier. Among qualifying developer-selected requests, 38% used general-purpose, 37% deep reasoning/planning and 25% fast/lightweight models. These counts exclude auto selection and bring-your-own-key requests. Separately, 63% of paying Copilot developers used auto model selection at least once in September. Model selection patterns are not quality rankings.
TypeScript and open-source infrastructure are prominent
TypeScript remained GitHub's most-used language by monthly participating accounts, followed by Python and JavaScript. In new 2026 repositories with detected agent involvement, 35% were primarily TypeScript versus 24% of repositories without that signal. GitHub's community ranking includes small or specialist projects such as cc-switch, Omarchy, llama.cpp, SGLang and vLLM, while product-led rankings feature coding agents and editors. These rankings describe contributor activity and engagement, not a controlled comparison of project quality.
What engineering teams should change
Treat an agent's pull request as an untrusted first draft that needs the same tests, ownership and review standards as human changes. Prefer smaller reviewable diffs, reproducible builds, regression checks and permission boundaries for autonomous agents. Monitor the cost and volume of CI runs and review queues, not only commits or PRs created. The report supports these as prudent operational implications, not proof of a universal productivity gain.