Octoverse 2026: agent drafts open 40% of PRs

GitHub published its annual Octoverse 2026 report on October 28, 2026, and the headline is a structural shift, not a growth anecdote: by September, more than 40% of pull requests on the platform started as an agent’s first draft — up from 17% in March, with roughly 110 million agent-drafted PRs in September alone, about six times the March volume. The report covers every push, commit and merge on GitHub between October 1, 2025 and September 30, 2026, private repositories included.

Key points

Developers pushed code to GitHub 17.9 billion times over the past 12 months — more than 2024 and 2025 combined. The year also saw 34.3 billion commits and 1.3 billion merged pull requests (3.5 million a day), each more than triple its level of two years ago.
New repositories created with an agent configuration file logged eight times as many commits as the average new repo without one; 55% of Copilot monthly active users now use agentic features, versus 35% a year ago.
Agent-heavy repos skew toward the most verifiable language: among repos created in 2026, 35% of those with an observed agent signal have TypeScript as their primary language, against 24% without a signal.

Background

Octoverse is GitHub’s yearly census of activity on the platform, and past editions were cited mostly for one thing: the language rankings. The 2025 edition made headlines when TypeScript overtook Python and JavaScript for the first time. The 2026 edition changes the subject — the official blog post is titled around commits and pushes reaching new highs “as developers build with agents,” and agent participation is promoted from a side note to a main indicator.

Two methodology notes matter before reading any number. First, GitHub recalculated the 2024 and 2025 historical figures under this year’s definitions — a window that includes private repositories — so they are not comparable with the numbers quoted from older reports. Second, GitHub calls the agent signal “a floor, since not all agent use may be detected.” But it publishes no detection error rate or estimate of missed activity, so readers cannot infer the true share or quantify the gap between actual use and the reported figure.

The numbers in the report

The headline metrics, all from GitHub’s official blog post, covering October 1, 2025 through September 30, 2026:

Metric	Value	Comparison
Code pushes	17.9B	More than 2024 and 2025 combined
Commits	34.3B	More than 3x two years ago
Merged PRs	1.3B (3.5M a day)	More than 3x two years ago
Agent-drafted PR share	40%+ (September)	17% in March; ~110M in the month
New repos with agent config	8x more commits	Vs average new repo without one
Copilot MAU on agentic features	55%	35% a year ago
TypeScript as top language, new repos with agent signal	35%	24% without a signal
Total developers	225M+	India 32M second; US 38.3M first

Two cuts of the same data deserve a closer look. By language: TypeScript added 3.62 million monthly contributors year over year, Python 3.23 million, and Rust doubled in the past year — the fastest percentage growth among the top languages. The report finds roughly 40–44% of merged PRs in TypeScript, Python and Rust show AI involvement, a concentration in the languages that are easiest for a machine to verify. By industry, the contrast is sharper: construction has the highest share of repos with agent-opened PRs (about 9%), while health care and mining — sectors that talk about AI constantly — see four or more repos in ten mentioning AI, but agent PRs are rare, under 3%.

GitHub also showed its own uncertainty. It ran 43 time-series and regression models to project activity out to 2031: about half project “extreme” growth (3x or more), just under half “substantial” (2x to under 3x), and only three models land under 2x. The report’s own words: “our own forecasting models no longer agree on where it will go next.”

What others say

Digit, the Indian technology outlet, led its coverage with the country data: India now hosts 32 million GitHub developers, second globally (China is third with 12.4 million, Brazil fourth with 10.2 million), added 10.08 million developers in a year, and grew 46% against the US’s 37%. GitHub’s own projection has India reaching 57.5 million developers by 2030, past the projected 54.7 million for the US.

The same piece delivers the most systematic skepticism published on the report so far, on four grounds: GitHub is grading its own platform; country attribution rests on self-reported developer locations; the recalculated 2024–2025 baselines reset every year-over-year comparison; and the India-overtakes-US story leans on forecasts where half the models say “extreme growth” while disagreeing with each other. Digit’s advice to readers: treat it as “a signal rather than a census.” Visual Studio Magazine adds a separate caveat: these are GitHub’s platform measurements, not evidence that agent-written changes are more reliable. It also notes that repositories with agent configuration files averaged eight times as many commits as those without them, a correlation that does not establish causation.

Not every independent number survives scrutiny, though. Digit describes the March-to-September rise as “twice higher than in March” — by GitHub’s own numbers, 17% to 40%+ is more than a 2.3x move. Independent coverage garbles figures too; where the two disagree, this entry follows the primary source.

Agent PRs merge faster

One under-reported detail says more about day-to-day reality than any headline total: agent-involved PRs now merge slightly faster than human-only ones. In September, 90% of agent-involved PRs merged within two weeks, against roughly 86% for non-agent PRs. In February the same comparison ran the other way — 83% versus 86%, with agent PRs slower.

The report offers no causal explanation for the six-month flip, and none should be invented on its behalf: tooling matured, submitters learned which tasks to delegate, or review workflows adapted — any mix of these could be true. What is verifiable is the trend itself. A year ago, “AI-generated code is hard to merge” was the standard complaint; on GitHub’s own platform, that gap has since reversed. The friction has moved elsewhere — into review and verification, exactly where Anthropic’s published Claude Code review pipeline shows a mature team spending its effort.

Our take

For developers and engineering managers, these figures are clues about platform adoption, not direct proof of team productivity: GitHub reports that more than four in ten PRs start from an agent draft, while new repos with an agent configuration file averaged eight times as many commits as new repos without one. A configuration file shows that a repository contains agent configuration; it does not prove how routinely a team uses it or that the configuration caused more commits. Visual Studio Magazine likewise cautions that GitHub’s data do not show agent-written changes are more reliable. For the receiving end of that shift, our earlier coverage of Indeed’s developer-hiring data offers a point of comparison.

The industry spread is the second mirror. Construction repos lead in agent-opened PRs while health care and mining mostly talk about AI without running it in their codebases — an AI strategy in a slide deck is a different thing from one in a repository. For teams building coding agents, that gap is the market: by GitHub’s own count, under 60% of PRs still start without an agent draft.

Three limits should be stated plainly. The window is one platform, with private-repository counting defined by GitHub alone and not independently auditable; GitHub calls its agent signal a detection floor but publishes no error estimate, so it does not let readers calculate the true share; and GitHub’s own forecast models disagree with each other, so any “by 2031 it will be X” quote deserves to be treated cautiously. If you want a number you can actually check, look at your own repository: what share of last month’s PRs started as an agent draft?