Agents push GitHub's numbers up — and expose the limits of measurement
On 28 October 2026, alongside its GitHub Universe conference, GitHub published its annual Octoverse report covering the period from 1 October 2025 to 30 September 2026. The company's self-reported data show activity roughly tripling: pushes reached 17.9 billion (more than 2024 and 2025 combined), commits totalled 34.3 billion, and merged pull requests 1.3 billion, around 3.5 million a day. The most disruptive element concerns coding agents: in September 2026 more than 40% of pull requests started from an initial draft generated by an agent (up from 17% in March), for a total of roughly 110 million agent-initiated pull requests in September alone, about six times the March volume. In the report, the company states that this percentage “is a floor, since not all agent use may be detected”. Alongside that statement, though, it is worth noting that the note does not describe the tracking method and counts only the initial draft, without clarifying how much of the final code was written by AI or how many of these requests were actually merged into repositories. According to The Next Web, the acceleration in agent activity was concentrated from June onwards, when the tools became able to run over several days and in the cloud.
Geographically, the platform counts more than 225 million developers in total, led by the United States (38.3 million) and India (32.0 million), each adding more than 10 million new users over the year; Italy does not appear in the report's top-ten country ranking. Among languages, TypeScript ranks first, followed by Python and JavaScript. Rust stands out with the highest percentage growth among the major languages: the report's text describes it as having “doubled”, although the charts in the same document show a 162% year-on-year increase. The way people interact with language models is changing too: in September, 61% of Copilot users who manually pick a model used more than one (compared with 50% four months earlier), with 47 distinct models in use among paying users and no single system above 25% of requests.
Alongside the report, the company announced the availability of Copilot Fusion and a credit-based payment model for Enterprise customers. According to GitHub, on the Artificial Analysis Coding Agent Index the Fusion Max variant lands within two points of Opus 5.5 Max, at half the cost and taking 40% less time per task. These are the company's own figures and have not been independently verified. In the same post, Jay Parikh, EVP of CoreAI at Microsoft, wrote that the software written so far will be “less than 1% of the software we'll write in the next five years”.
Measuring the transformation of software by counting a dizzying rise in automated drafts risks making us mistake volume of output for sound system architecture. A million extra generated lines do not automatically add up to better engineering. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official Octoverse 2026 page on github.blog (dates, period, pushes, commits, pull requests, the 40% and 17% shares, 110 million, the ‘floor’ caveat, developers by country, languages, multi-model use) and Jay Parikh's GitHub Universe 2026 post (the quote, the Fusion Max claims, credit-based pricing). The figures match those reported by The Next Web on 28/10/2026. I set aside topics already covered, unverified allegations and personal legal matters.
- Incertezze
- The data come from GitHub (Microsoft), which sells the very agent tools being measured. The 40% counts only detectable agents and is presented as a minimum, but the detection method is not explained. An ‘agent-initiated’ PR only means the first draft: it says nothing about how much final code the AI wrote or whether the PR was merged, and the report does not say how many of the 1.3 billion merged PRs came from agents. On Rust the text says ‘doubled’ while the chart shows +162% year on year. Copilot Fusion benchmarks are company claims, not independently verified. No data on Italy.
- Perché pubblicarla
- It is the first large-scale, official figure on how far AI agents have already entered real-world software development: from 17% to over 40% of pull requests in six months, on a platform of 225 million developers. Useful for anyone working in software or following AI's impact on work, as long as the methodological limits are clearly explained.