How Many Repos Use AGENTS.md? 6,006 of 22,854 on October 16, 2026

On October 16, 2026, 6,006 of 22,854 popular, active GitHub repositories had a root AGENTS.md, a root CLAUDE.md, or both. That is 26.3%.

Pawel Zarzycki’s census is arXiv:2610.31715, version 2, revised October 26, 2026. The pages that rank for the filename repeat a round number from the format’s homepage. The homepage, fetched November 9, 2026, still says the format is used by over 60,000 open-source projects. The census says how many repositories cleared a star floor, a recency floor, and a language list, and then had one of two root files. Both sentences can be true. They are not the same count.

What is AGENTS.md?

A plain Markdown file that coding agents are expected to read before they change a repository. The homepage calls it a README for agents.

The page fetched November 9 opens with “A simple, open format for guiding coding agents, used by over 60k open-source projects.” It says a README is for humans and that AGENTS.md holds build steps, tests, and conventions that would clutter that README. The sample on the page has setup commands and a code-style section. The compatible-agent list on that page includes Codex, Cursor, Copilot, Gemini CLI, Aider, Devin, Zed, and others. This note did not test whether each named tool loads the file.

The spec repository agentsmd/agents.md returned 25,408 stars, 1,949 forks, the SPDX id MIT, and pushed_at 2026-09-10T18:46:16Z. The HTTP Date on that response was 13:39:04Z on November 9, 2026. Stars on the spec repository are not the adoption count. The paper cites the homepage as accessed on the census day, October 16.

A harness is the software around the model: the loop, the tools, the permissions, the stop rules. An instruction file is one input to that loop. What is an AI agent harness? describes a short AGENTS.md used as a map. Thoughtworks’ write-up of the same pattern is harness engineering at scale.

How many popular repositories had a root file on October 16?

6,006 of 22,854. Adoption is higher in the repositories with the most stars.

The search that built the frame asked for public repositories with at least 1,000 stars, a push in the previous 90 days, that were not forks and not archived, in thirteen languages. GitHub returned 22,854. A repository counts if the default branch has a root AGENTS.md or a root CLAUDE.md in that exact case.

89 + 3,059 + 3,087 + 16,619 = 22,854. The bottom tier is 23.2%. The top tier is 59.6%, and it contains 89 repositories. The paper’s Figure 1A is this table.

Of the 6,006 adopters, 2,566 have only AGENTS.md, 1,052 have only CLAUDE.md, and 2,388 have both. 2,566 + 1,052 + 2,388 = 6,006. Among the 2,388, CLAUDE.md defers to AGENTS.md through a link, an import, or a pointer in 1,540 repositories. The reverse pointer appears in 182. The paper says 598 keep two files that each have content of their own. 1,540 + 182 + 598 = 2,320, which is 68 short of 2,388. The paper’s sentence does not account for those 68 in the three buckets it names. This note leaves the gap as the paper left it.

New adopters moved from starting with CLAUDE.md to starting with AGENTS.md. The paper says 73.7% of new adopters from January through August 2025 started with CLAUDE.md, and that the CLAUDE.md-first share since May 2026 is 17.1%. It dates Claude Code’s support for AGENTS.md to September 18, 2026, from changelog version 2.1.277, and draws that date as a dashed line on the monthly chart. This note did not re-open the changelog. The census day is four weeks after the date the citation uses.

What else did the census count, besides the two root names?

6,652 repositories, 29.1%, had at least one of the instruction paths the paper listed. 646 had one of the other paths and no root AGENTS.md or CLAUDE.md.

4,954 equals 2,566 + 2,388, the only-AGENTS.md repositories plus the both-file repositories. 3,440 equals 1,052 + 2,388. The root-file census and this wider list meet. 6,652 minus 6,006 is 646, the paper’s count of repositories that use another format and no root file. Copilot’s instruction path is the largest of those other formats, at 1,024 repositories, 4.5%. Cursor’s .cursor/rules/ is 347. A case-variant filename such as agents.md at the root is 47 repositories, and only 8.5% of those also have the exact-case AGENTS.md. The exact-case rule in the main count is doing work.

What do the files contain?

Rules, in almost every instruction set. The median length is 818 words. Most directives do not say why.

Sonnet’s codes, reweighted to the population, put at least one rule in 98.2% of instruction sets (95% CI 97.3–98.9). A repository’s statements average 52.6% rules, 29.8% descriptions, and 17.6% references. The rules cover code and its artefacts in 88.7% of repositories, workflow in 85.4%, the agent’s own behaviour in 79.3%, and the instruction files themselves in 52.7%.

The subject shares the paper prints for rules: code conventions and quality 74.4%, build, environment, and delivery 72.6%, agent workflow and conduct 67.3%. Project knowledge appears in 86.2% of instruction sets, and 54.3% contain a rule on that area. The most common specific rule subjects are verification before declaring work done (49.7%), test writing (48.2%), and the instruction files themselves (44.3%). Each of those three is under half of repositories.

The 10th and 90th percentiles of length are 163 and 3,010 words. A file can be a short command list or a long manual. The median, 818, is the figure to remember.

Do the rules give a reason?

Most instruction sets contain a reason somewhere. Most directives do not carry one. 74.0% of directives have none.

82.9% of instruction sets give a reason at least once. 10.8% cite an incident or evidence. 80.9% mark at least one rule as mandatory, and 77.8% prohibit something, yet prohibitions average 15.8% of a repository’s directives. Unmarked imperatives are the most common form. Rules are tied to a moment in the workflow in 54.7% of instruction sets, and 53.7% state an exception. 8.6% of repositories name a specific agent somewhere. 8.3% address human contributors. Vague directives appear in 40.5% of repositories and are 3.5% of directives.

These shares are Sonnet’s codes on the reweighted sample. The human review, below, is why this note keeps the label on them.

On collaboration files, Haiku’s uncorrected codes name concrete requirements. Of files with a rule on pull-request descriptions, 27.2% require a test plan, 24.2% require the project’s template, and 23.3% require linked issues. Of files with a commit-message rule, 36.7% require Conventional Commits. Test-writing rules prohibit trivial tests in 8.1% of those files and prohibit weakening tests so they pass in 4.9%.

Do projects agree about AI attribution?

Commit-message rules prohibit it more often than they require it. Pull-request rules more often require disclosure. The paper prints both a file table and a repository sentence, and the percentages are not identical.

The 7.5% and the 7.4% are close and they are not the same statistic. One is repositories, with an interval. One is files, uncorrected. The “59 require it” is a count. The sentence does not print how many repositories had a commit-message rule, so this note does not turn 59 into a percentage. The pull-request prose says 5.8% require disclosure and 1.9% prohibit it, mostly in different repositories. The file table says 5.7% and 1.9%. Trailer-based measures of AI use inherit the disagreement: a project that forbids the trailer and a project that demands it will not leave the same trace for the same amount of agent work.

Claude Code’s “Don’t ask again” files are a different instruction surface, permission prompts rather than a root markdown file. The census does not count them.

Where does the text come from, and do projects edit it?

15.8% of instruction sets match an /init signature. 81.8% were edited after they were created.

948 of 5,998 instruction sets match a signature the paper associates with an /init command. The opening line of claude /init is in 692, which is 11.5%. Codex’s heading set appears in 146. The share falls with popularity, from 17.1% in the lowest star tier to 9.4% in the highest. The paper says the detector mainly measures Claude Code, because other tools leave no verbatim line.

93 instruction sets, 1.6%, contain a block a tool writes into the file. The Next.js development server accounts for 44 of those. 3.4% share a paragraph verbatim with the files of at least two other owners. 17.7% were created in an AI-marked commit, and 41.7% have at least one AI-marked commit. Among the 5,942 instruction sets with a file history, 33.1% carry a text trace or were created in an AI-marked commit. The paper’s “a third” is that 33.1%.

81.8% were edited after creation. Sets with a generator signature were edited about as often as sets without one, 82.4% against 81.7%, and the paper does not adjust those two shares for age. By file, 80.0% of AGENTS.md files and 67.7% of CLAUDE.md files have more than one commit. 24.5% of the 106,279 commits that touched agent files are AI-marked. The marker is a trailer. It does not identify which lines a model wrote.

Which languages have the files?

TypeScript is the highest share, 40.8% of 4,048 repositories. C is the lowest, 13.0% of 1,272. Python, the largest group, is 24.3% of 5,015.

The thirteen repository counts sum to 22,854. JavaScript’s either-share, 17.5%, is well below TypeScript’s 40.8% even though the names are often discussed together. The percentages are the paper’s Table 4. They are rounded, so a reader who multiplies them back by the repository count will not always land on an integer.

Which of the paper’s hypotheses held?

Three of six prespecified hypotheses held. More stars go with more rules about pull requests or AI contribution policy. Compiled languages go with more build-and-run statements. In a repository that has both files, CLAUDE.md is more descriptive than AGENTS.md.

The tests use 5,587 repositories with at least five coded statements. A hypothesis counts as supported only if every one of its tests points the predicted way and the adjusted p-value is under 0.05. H1 and H4 are significant in the wrong direction, which is why they are not supported. The paper says H3, H6, and the pull-request half of H2 still hold on the Sonnet sample and when either of the two largest languages is left out.

The exploratory models, 99 tests on 5,586 instruction sets, associate commits in the past year with more of the content outcomes than the other predictors: 11 of 13 tests on that predictor stay significant after the false-discovery correction. LLM and agent projects are 27.9% of adopters and 13.3% of matched controls, an odds ratio of 2.76 within star tier and language. A post hoc check that holds age and activity fixed, and drops agent-tooling projects, shrinks that ratio to 1.54 (95% CI 1.28–1.84). Libraries are under-represented, odds ratio 0.58. The paper classifies project type with Haiku on repository metadata. On 29 repositories the reviewer agreed on type at kappa 0.43 and on AI focus at kappa 0.80. The type contrasts are the paper’s “indicative” results.

How much did the human review agree?

The reviewer accepted the statement boundaries in 55 of 60 cases, 92%. Agreement on the subject was 43% to 50%.

The 60 statements were coded before the reviewer saw the model labels. Restatements and extracted elements were accepted in 56 of 60, 93%. The free-text subject fit in 57 of 60, 95%. Against the models, description versus rule agreed in 82–83% of statements once references were excluded. The three-way split into description, rule, and reference agreed in 62–67%, kappa 0.33–0.40. Subject area agreed in 43–50%. The fine subject agreed in 32–37%.

The paper’s own reading of that gap: the models agree with each other more closely than with this reviewer, and subject categories stay model-defined. One reviewer cannot produce a human-human kappa. The adoption counts in the first table do not depend on these codes. The 74.0%, the subject shares, and the attribution stances do.

What would a fresh count have to change?

A count that wants to update the 6,006 has to keep the frame or publish a new one. “Over 60,000” does not update it.

The homepage number, read November 9, has no date, no star floor, and no language list on the sentence itself. The census has all three, and a day. Adding nested AGENTS.md files, dropping the star floor, or including archived repositories would move the numerator and the denominator together. The paper’s data-availability section says the dataset is on Hugging Face and does not print a URL. This note did not locate the dataset from that sentence, and it did not re-run the search.

Which libraries people actually install around these agents is a different November 9 count: Which AI agent framework do people install?. Coding-agent prices and the dates already on vendor pages are the 30-day coding-agents brief. The sourced product table is Best AI coding agents. How the loop uses a file like this one is How do AI agents work?.