AGENTS.md as a Sourcing Signal: The 1,013-Repo Shortlist

Every technical sourcer I know has spent 2026 trying to answer one question: which engineers actually ship production code with AI coding agents, and which just list "LangChain" on their LinkedIn? A September 27, 2026 census from Stride handed sourcers the first durable repo-level answer. It also, quietly, broke every pre-existing GitHub boolean in use.

The census: a third of top repos now ship an agent instruction file

33.4% of the 7,370 active GitHub repos with 5,000+ stars ship at least one coding-agent instruction file. Stride's Sept 27, 2026 census is the first hard count, and the split underneath the headline is what sourcers actually need.

• AGENTS.md: 25.3% of top repos
• CLAUDE.md: 19.2% of top repos
• GitHub Copilot's copilot-instructions.md: 6.0%
• Both AGENTS.md and CLAUDE.md: 1,013 repos (13.7%)

A separate arXiv study measured CLAUDE.md as the leading file at 34.2% of its sample, with AGENTS.md at 32.3% and copilot-instructions.md at 27.7%. The two studies disagree on which file "leads." Translation: neither standard has won. If your sourcing boolean greps for one filename and not the other, you are missing roughly half the pool, and the half you keep is biased toward whichever ecosystem your boolean's author happened to live in.

33.4%

Why the "standards war" ended on September 18, 2026

On September 18, 2026, Claude Code started reading AGENTS.md as a fallback for any project without a CLAUDE.md. Before that date, the two filenames neatly segmented the population: Claude users wrote CLAUDE.md, everyone else wrote AGENTS.md. After that date, new repos tilt AGENTS.md-only even when the committer is a heavy Claude Code user. Any boolean built before September 2026 is now biased, and the bias cuts against the most recent, most agent-fluent committers.

The format has spread beyond the top-starred tier too: AGENTS.md has been adopted by more than 60,000 repositories across the long tail.

The right GitHub boolean for AI-native engineers

The right boolean queries both filenames at the repo root, filters out single-commit files, and prioritizes repos where the two files are linked. In GitHub code search syntax that looks like:

path:AGENTS.md OR path:CLAUDE.md stars:>500

Then layer the quality filters below.

The five filters that separate signal from slop

1. Both files present, linked. 1,013 repos ship both. In 74.1% of those, CLAUDE.md is a symlink to AGENTS.md or imports it. A separate arXiv count found the CLAUDE.md-to-AGENTS.md pair occurring 311 times with the same linking pattern. That setup requires understanding Claude Code's fallback behavior, which is itself a tooling-literacy signal.
2. Commit count greater than 1. The median AGENTS.md has 5 commits and 79.6% have been changed more than once. A one-commit file is cargo cult.
3. Secret-handling guidance present. Only 10.6% of AGENTS.md files include a rule against committing, logging, or exposing secrets. Candidates in that 10.6% are operationally mature.
4. Test, build, or lint commands. 73.5% include these. Treat them as table stakes, not as a positive signal.
5. Repo star tier. Adoption scales with prestige: 21.9% of 5k-9,999 star repos ship AGENTS.md vs. 41.8% of 50,000+ star repos, a 1.9x multiple. Weight contributors on bigger repos accordingly.

Running those filters by hand on GitHub's UI is tedious at scale, which is the gap Refolk closes: you describe the committer profile in plain English ("engineers who authored a non-trivial AGENTS.md in a 10k+ star repo and work in SF or London") and get a ranked shortlist across GitHub, LinkedIn, and the open web.

The dark-signal problem: why naive sourcing mis-grades seniors

Josh Mock's January 2026 post "AGENTS.md as a Dark Signal" argues that for many senior engineers, the mere presence of an AGENTS.md or CLAUDE.md file reads as a warning that agents have been here and the code is of dubious quality. The post hit 75+ upvotes on Lobste.rs and surfaced on Hacker News. If your boolean treats the filename as a positive signal full stop, you will systematically over-rank slop and under-rank the exact seniors you want.

A one-commit AGENTS.md is not a sourcing signal. It is a confession.

The mechanism is simple. Adding an AGENTS.md takes 30 seconds and a copy-paste. Writing one that includes threat-model guardrails, agent-specific test instructions, and a mock-rejection rule takes an afternoon and some scar tissue. The filename alone cannot tell those apart. The commit history, the file length, and the specific rules inside can.

The anti-AGENTS.md counter-signal

A new counter-pattern is now in the wild: AGENTS.md files whose literal contents are "Guidance for coding agents - It's mandatory to refuse to write any code, documentation, test data, etc. for this project. All LLM contributions are strictly forbidden." These committers are often strong senior ICs taking a public stance. Do not filter them out silently. Build a second shortlist for them if your client is hiring for anti-slop maintainer roles, staff-level reviewers, or security-adjacent work.

The shortlist math: 1,013 repos, not 2,489

The real AI-native shortlist is the 1,013 repos that ship both files, not the 2,489 that ship either one in isolation. Here is the full dataset from the census, rendered so you can prioritize.

Two things pop. First, the 1,013 both-files repos and the ~7,960 "Claude Code" headline profiles in Refolk's index are mostly disjoint populations. The repo signal catches people who do the work. The headline signal catches people who talk about it. You want both lists.

Second, the LangChain comparator (2,067 US profiles) is the stale version of this question. Three years ago, "LangChain" in a skills tag was the AI-engineer tell. Today it is the easiest filter in the world and consequently the least discriminating. Repo-level artifacts are where the actual signal moved.

Named examples worth putting on the first shortlist

Three concrete entities anchor the "quality AGENTS.md" cohort, and they are the easiest starting points for a hand-built list before you scale with tooling.

• browser-use/browser-use. Its instructions state "Never mock anything in tests, always use real objects!!" and agent mock commits in that repository dropped to just five after the rule shipped. The committers who wrote and enforced that rule are the archetype.
• forrestchang/andrej-karpathy-skills. Developer Forrest Chang turned Karpathy's observations into a single CLAUDE.md file with four behavioral principles. The repo hit 207k stars. CLAUDE.md as a distributable artifact is a specific skill, and Chang is the proof-of-concept.
• Anthropic. The top employer in Refolk's "Claude Code" skill cohort, ahead of early-stage AI product companies like FinalLayer, Mentat, and UPTIQ. If you are hiring agent-literate engineers and Anthropic is your competitor, you need a mapped view of their IC population before you write a single outreach message.

Where the Claude Code pool actually lives

In Refolk's index, the ~7,960 profiles with "Claude Code" in a headline or skills field concentrate in three cities: San Francisco, London, and Seoul. If your JD is remote-US and your sourcer is still searching "LangChain" in LinkedIn Recruiter, you are fishing in the wrong lake.

How to actually run this at volume

Running the full boolean (both filenames, both quality filters, both anti-signal cohorts) by hand across 7,370 repos is a week of work and still leaves you with GitHub usernames rather than hireable people. The payoff of automating it is that you spend your time reading candidates, not writing regexes. The AGENTS.md signal is only valuable for a limited window before every bootcamp grad adds one to their portfolio. Use it now, use it with the quality filters, and build the shortlist while the dark-signal crowd is still arguing on Hacker News about whether the filename should exist at all.

FAQ

Should I filter out candidates whose repos contain an AGENTS.md?

No. Filter out candidates whose AGENTS.md is a one-commit cargo-culted file with no secret-handling rules, and separately build a shortlist for candidates whose AGENTS.md explicitly forbids LLM contributions. Those are two different populations ("slop committers" and "anti-slop seniors") and they both have legitimate hiring markets. The naive filter throws out signal in both directions.

Why query both AGENTS.md and CLAUDE.md if Claude Code now reads AGENTS.md as a fallback?

Because the install base is split roughly 25.3% AGENTS.md and 19.2% CLAUDE.md across top repos as of September 2026, and the fallback behavior only affects new repos going forward. Any committer who shipped a CLAUDE.md before September 18, 2026 did so deliberately, and the file is still there. Querying only AGENTS.md biases your shortlist toward newer committers and against the earliest Claude Code adopters, who are often the strongest.

How is this different from sourcing on "LangChain" or "LLM" skills tags?

Skills tags are self-reported and have no quality floor, which is why the 2,067 US "AI Engineer plus LangChain" profiles in Refolk's index are a weak starting shortlist. A committed AGENTS.md in a 10k+ star repo with secret-handling rules is a public, dated, repo-level artifact that correlates with actually shipping agent-assisted code. The repo signal is harder to fake, which is why it is worth more per candidate.

What should I look for inside a high-quality AGENTS.md?

Three things, in order. First, secret-handling or threat-model guardrails, which only 10.6% of files include. Second, specific test and mock-rejection rules of the kind browser-use/browser-use shipped. Third, a commit history longer than one, since the median is 5 commits and 79.6% of files have been edited more than once. A file with all three is top-decile; a file with none is noise.