What GitHub means by ‘agent-started’

GitHub defines an agent-started pull request as a pull request whose first draft came from a coding agent that it can detect. That glossary line is the entire definition, and the phrase carrying the most weight is “that we can detect.”

GitHub describes the underlying evidence as an “agent signal”: “an observed sign of agent involvement in telemetry, such as an agent-opened pull request.” Its glossary adds that the signal “is a floor, since not all agent use may be detected.” The floor wording matters. GitHub is telling readers that its measured share is a lower bound rather than a count of every agent-assisted pull request on the platform. Its October 2026 release reports that more than 40% of pull requests in September started with an agent’s first draft, up from 17% in March.

The post also uses four labels that are not interchangeable. Agent-started means the first draft came from a detectable agent. Agent-involved describes a comparison group in a merge-speed analysis. Agent-opened appears as a signal in the industry section. “Merged pull requests with detected AI involvement” labels a language chart. Each measures something different, and a reader who swaps them will misstate what moved.

One question stays open. GitHub never says whether a session a person starts and an agent drafts counts toward the metric, or whether the agent has to act on its own. Its own sentence makes the developer the actor who starts more than 40% of pull requests “with an agent,” which is consistent with a human trigger, and the detection method is not published at all. The careful reading is narrow: the metric turns on where the first draft came from, and GitHub has not said more than that.

‘Agent-started’ is not ‘agent-written’

The 40% answers a question about process. In September 2026, how many pull requests began with a draft an agent produced? It does not answer who wrote the code that shipped. A pull request can open with an agent’s draft and then be rewritten, reviewed, or replaced by the person responsible for it. The metric records the first version.

GitHub’s closing section makes the same distinction: “None of that activity changes who decides what the code should do.” Someone still has to come up with the idea, and someone still has to decide whether the result works and belongs. Reading 40% as the share of code written by AI, or as a productivity or quality result, goes past what the telemetry supports.

Pushes, commits, contributions and merged pull requests are different units

GitHub’s glossary draws lines between five things the Octoverse release leans on at once. A push sends saved changes to GitHub, and one push can carry many commits. A contribution counts each push once, so commit totals can exceed contribution totals. An opened pull request and a merged pull request are also separate counts.

The numbers illustrate the gap. The year to 30 September 2026 saw 17.9 billion pushes and 34.3 billion commits, alongside 1.3 billion merged pull requests and 2.5 billion public contributions. Pushes and commits are not two views of one quantity. A push carries one or more commits, and a contribution registers the push once no matter how many commits it carried. Grouping them as though they describe the same activity inflates the impression of a single trend.

GitHub’s Universe post pushes back on reading volume as output: “Completed work, not more code, is the output that matters.”

What changed in the 2026 counting

The 2026 edition moved two things at once. The Octoverse year now runs 1 October to 30 September; earlier reports covered 1 September to 31 August. The base widened from public activity to public and private repositories, while industry, open-source contribution and open-source project figures stay public only.

GitHub recalculated 2024 and 2025 on the new definitions and window, so the year-over-year statements inside the 2026 post are internally consistent. They are not comparable with earlier published Octoverse figures. The discontinuity is large: the 2025 report gave 518.7 million merged pull requests and about 986 million commits, and the 2026 report gives 1.3 billion merged pull requests and 34.3 billion commits.

The agent share inherits the same caution. GitHub does not flag the more-than-40% figure as public-only, and its rule is that metrics include public and private repositories “unless noted,” so the figure plausibly rests on a larger and less externally checkable base than older public Octoverse numbers. That reading follows from GitHub’s stated rule; GitHub has not confirmed it. For scale, the 2025 report noted that 81.5% of contributions happened in private repositories.

GitHub publishes no detection method for the agent signal, no false-negative or false-positive rate, and no coverage estimate beyond the glossary’s “floor.”

The 8x figure has no published base sizes

Among repositories created this year, GitHub reports, the average repo with an agent configuration file logged eight times as many commits as the average new repo without one. An agent configuration file is a repository file that sets up or directs a coding agent; GitHub notes that repositories without one may still use AI.

The 8x is a mean commits per repository, limited to repositories created inside the Octoverse year. It is not a total, not a per-developer figure, and not per unit of time. GitHub publishes no count of repositories in either group and no absolute average, so the ratio cannot be converted into a share of all commits, and the size of the comparison groups is unknown. The gap is also confounded: adding an agent configuration file is a self-selected marker of teams already committed to agent workflows, and the comparison group includes repositories that use AI without the file. Crypto Briefing put the standard caveat on it: “an eightfold gap shows correlation, not proof that the file itself drives activity.”

What the telemetry cannot show

The numbers come from GitHub’s own platform data. They are interested-party telemetry, self-reported, and this review found no external audit of them. That does not make them wrong, but it does mean the claims rest on GitHub’s account of its own product.

GitHub did not repeat in its 2026 post the caveat its 2025 report carried, that such signals are observational rather than causal and “more work is needed to understand the full impact AI is having.” The 2026 post calls the 8x figure a “sign.” Nothing in the release establishes that agent use caused the rise in started pull requests, or that it changed code quality.

Coverage published on 28 and 29 October treated the numbers as more settled than the definitions are. Developer Tech News rendered the metric as “an AI agent-generated draft” and noted that the figures “reflect the time taken to merge pull requests, rather than the quality of the resulting code.” Crypto Briefing attributed the figure to Satya Nadella’s Universe talk, in its report of the talk rather than a verified transcript, and wrote that “the bots didn’t just show up. They took over a meaningful chunk of the workflow.” The Stack described the year as “driven primarily by agent use.” None of that coverage examined what the metric counts. GitHub’s glossary supplies the only published definition, and it does not describe how the signal is detected.