I've been analyzing a corpus of 558 repos that already had an AGENTS.md file, but that's a biased way to measure adoption. To find out if these files are actually common, I sampled GitHub using two different methods. The results are wild: 6.2% of active repos (those pushed in the last 90 days, not forks, and not archived) have one, but across all public repos, that number drops to just 1.0%.
That's a 6x difference depending on whether you look at the total "stock" of repos or just the ones people are actually working on.
A few things caught me by surprise during this process:
- CLAUDE.md is just as common. It shows up in 5.4% of active repos, which is statistically the same as
AGENTS.md. Neither format has "won" yet. - GitHub is mostly a graveyard. 93% of public repos haven't seen a push in 90 days, 29% are forks, and 8.3% are totally empty.
- Existing files are often abandoned. Out of the 9
AGENTS.mdfiles found in the general population sample, 7 belonged to repos that haven't been touched in three months.
Why the specific sample size matters #
I realized I was falling into a sampling trap. When I said "85.7% of files prohibit things," the denominator was only the files that already existed. That doesn't answer the real question: should you actually bother writing one?
To get an honest answer, you need a sample drawn independently of whether the file exists. I broke this down into three frames to avoid "confident nonsense":
- Stock rate: Answers "is this mainstream?" by looking at every public repo.
- Active rate: Answers "is this what working projects do?" by looking at repos pushed in the last 90 days.
- Trend: Answers "is it spreading?" by looking at rates based on the repo's creation year.
How I pulled the data #
I used a recursive tree call for every repo to ensure I didn't miss files tucked away in folders like docs/AGENTS.md.
GET /repos/{owner}/{repo}/git/trees/HEAD?recursive=1
I processed roughly 1,900 repos using seed 20260918. For the active rate (Frame A), I sampled random slices of creation time from 2010–2026 and queried created: pushed:>2026-06-20 fork:false archived:false. I pulled every result from those slices and then randomly picked 50 repos per cohort to avoid the bias of GitHub's "relevance" ranking.
For the stock rate (Frame B), since GitHub doesn't offer a random sort, I binary-searched the maximum repo ID (1,375,203,308) and sampled IDs uniformly. Interestingly, only 35% of those IDs actually pointed to an existing public repo—the rest were deleted or private. Getting 1,000 usable repos required 2,920 probes.
Next Claude Code Projects and Jev are shifting how we think about agent orchestration →