GitHub's June 17 Agent Finder Just Made AGENTS.md a Higher Signal Than LinkedIn

If you have been writing Boolean strings with "LLM" and "agentic" and "RAG" for the last year, June 17, 2026 quietly broke them. GitHub shipped Agent Finder, the first product implementation of the open Agentic Resource Discovery (ARD) spec co-published with Google, Microsoft, Hugging Face, GoDaddy, and seven other companies. Combined with the AGENTS.md file convention that landed at Universe 2025, every repo that opts in now broadcasts, in machine-readable form, who is actually shipping agentic workflows in production.

That is a sourcing signal that did not exist 30 days ago. And it lands at a moment when resume databases are not just stale, they are empty. Querying a major professional-profile index for U.S. engineers with "MCP server" or "AGENTS.md" in the headline returns essentially zero matches. The skill is too new for LinkedIn's taxonomy. It is not too new for GitHub.

What actually shipped on June 17

Agent Finder is a Copilot feature that lets agents discover other agents, tools, and skills across the open web. Under the hood it implements ARD, a spec for how AI agents find and verify capabilities. The mechanics matter for sourcing:

• An organization publishes an ai-catalog.json file at a well-known path on its own domain. The file lists tools, MCP servers, agents, and APIs the org actually ships.
• Registries crawl those catalogs. Domain ownership verifies the publisher.
• Cisco tied its AGNTCY Agent Directory (Linux Foundation) to the same spec, broadening the public contributor graph.

The combined effect: a company that has shipped real agentic infrastructure now leaves a public, verifiable fingerprint at a predictable URL. Not a press release. Not a fundraising deck. A file.

For sourcers, that means hiring-target identification becomes a crawl. You no longer have to guess which Series B is actually building agents versus which one is just saying so on the careers page. The ai-catalog.json either exists or it does not.

The AGENTS.md file you should already be reading

AGENTS.md is the other half of the picture. GitHub shipped it at Universe 2025 as a source-controlled file that tells coding agents how to behave in a repo: "prefer this logger," "use table-driven tests for all handlers," "do not touch the migrations directory without approval." It is the Dockerfile of the agent era.

Per The Prompt Shelf, AGENTS.md is now in 60,000+ repos as a multi-tool standard, read by Copilot, Cursor, Claude Code, and the coding agents from Anthropic, OpenAI, Google, Cognition, and xAI that now run inside GitHub directly. Matt Nigh's GitHub Blog analysis of 2,500+ public AGENTS.md files (November 2025) gave us the first real rubric for what a good one looks like: explicit commands, boundary rules, persona frontmatter, test conventions.

That rubric is the new sourcing rubric. An engineer who authored a non-trivial AGENTS.md in a repo with real traffic has demonstrated three things at once: they understand their own codebase well enough to constrain an agent, they have shipped agentic workflows past the demo stage, and they have opinions about how LLMs should behave in production. You will not get that from a resume.

Why Copilot usage is now worthless as a signal

GitHub reports 180 million developers on the platform and that 80% of new developers use Copilot in their first week. "Uses Copilot" is the new "knows Git." It tells you nothing.

80%

The differentiator has moved one layer up the stack. Authoring AGENTS.md. Maintaining an MCP server. Committing to the agentfinder organization. Publishing to an ai-catalog.json. These are the artifacts that separate the engineer who pastes prompts into a chat window from the engineer who is rewiring their team's development loop around agents.

The finite, nameable pool

This is the part founders should print and tape to the wall. The expert pool here is small enough to enumerate.

As of May 24, 2026, GitHub's Search API returned 15,926 repositories tagged with mcp-server. The reference modelcontextprotocol/servers repo alone has 86,148 stars and 10,799 forks. Anthropic's December 2025 ecosystem update cited more than 10,000 active public MCP servers, each with a maintainer graph. The Python and TypeScript SDKs see roughly 97 million monthly downloads combined.

Sounds huge. Now look at who is actually steering the standard:

• 9 core lead maintainers of MCP under the Agentic AI Foundation (Linux Foundation, donated by Anthropic in December 2025).
• 58 total active maintainers across the MCP working groups.
• 2,900+ contributors in the official Discord, with public attribution.

That is the entire steering pool. Add the committers to github.com/agentfinder (Azure skill definitions for airunway-aks-setup, appinsights-instrumentation, azure-kubernetes and the rest), the contributors to Hugging Face's Discover Tool reference ARD implementation, and the public maintainers of Cisco's AGNTCY directory. You can build a named list in an afternoon.

The expert pool fits in a spreadsheet. The arbitrage window does not fit in a quarter.

This is the "first commit to Kubernetes" moment. The engineers who landed PRs in early k8s became the most aggressively recruited infra ICs of the late 2010s, and their comp benchmarks took three years to adjust. The MCP and Agent Finder maintainer pool is the same shape, at the same stage.

What to actually search for

Stop typing ("LLM" OR "agentic" OR "RAG") AND ("production" OR "shipped") into LinkedIn. It returns 400,000 people, all of whom updated their headline last Tuesday.

Try these instead, on GitHub:

1. Contributors to modelcontextprotocol/servers with more than two merged PRs in the last 180 days.
2. Authors of AGENTS.md files in repos with 200+ stars (you want production weight, not a weekend toy).
3. Maintainers of any MCP server published by a vendor whose product your candidate would actually integrate: Stripe, Figma, Sentry (GitHub called these out at Universe 2025), HashiCorp Vault, Auth0.
4. Anyone whose org publishes an ai-catalog.json at a verifiable domain path. That is your shortlist of companies, not just people.
5. Committers to the github.com/agentfinder organization repos.

The friction is that these are five different queries, across three platforms (GitHub code search, GitHub API, the open web for ai-catalog.json discovery), and the people you find still need to be matched back to a contactable identity. Which is exactly the friction Refolk was built to remove: you describe the engineer in plain English ("maintainers of MCP servers at vendor companies with shipping production deployments in the last 90 days") and get a ranked shortlist with the public profiles already joined.

The AGENTS.md quality filter

Not every AGENTS.md is a hiring signal. The Matt Nigh study made the gradient explicit. Skip files that are 20 lines of "be polite, write good code." Look for:

• Explicit command lists (pnpm test, make migrate, etc.) with expected outputs.
• Boundary rules (which directories an agent may not modify, which secrets it may not read).
• Persona frontmatter or role definitions.
• Test conventions, particularly table-driven or property-based test requirements.
• Logger and observability conventions (if they tell the agent which logger to prefer, they have been burned before).

An engineer who wrote that file has run agents in anger against their own codebase and lost. Then they wrote the rules so they would not lose again. That is the person you want.

ARD turns company sourcing into enumeration

The hiring-target half of this story matters as much as the IC half. ARD's ai-catalog.json convention means you can now build, in code, a list of every company that has published verified agentic infrastructure. Not "has a Show HN about agents." Has a catalog file at a verified domain that a third party would actually crawl.

For a founder hiring their first three agent engineers, this is the cleanest market map you will get. For a recruiting agency doing a retained search, it is a defensible sourcing universe. For a corp-dev team, it is an acquisition pipeline.

This is the second place Refolk earns its keep on this topic: you can ask for "engineers at companies with a published ai-catalog.json that lists at least one MCP server they authored," and the answer is a real list, not a saved search you have to babysit.

The 30-day half-life

Be honest about the window. ATS vendors will catch up. Greenhouse will add an "MCP" skill tag. LinkedIn will autopopulate "Agent Engineer" titles. The AGENTS.md authorship signal will be parseable by every sourcing tool in the category within two quarters, and the comp benchmarks for the named MCP maintainer pool will adjust shortly after.

60,000+

That growth curve is the arbitrage clock. The signal is highest right now because (a) the convention is adopted enough to be meaningful, (b) resume databases have not caught up, and (c) the spec is fresh enough that the maintainer pool is still finite and named. All three conditions degrade with time.

If you are hiring agent engineers in Q3 2026, treat AGENTS.md authorship and MCP contribution as primary screens, not nice-to-haves. By Q1 2027 they will be table stakes that everyone screens for, badly. Refolk's bet is that the plain-English query layer ("show me the 58 active MCP maintainers, minus anyone currently at a FAANG, minus anyone who has changed jobs in the last 12 months") is the durable interface even after the underlying signals commoditize. Get the names now anyway.

What changes Monday morning

Three concrete changes to your workflow this week:

1. Replace one "LLM engineer" Boolean string with a GitHub query for AGENTS.md authors in repos with 200+ stars. Compare reply rates over two weeks.
2. Pull the public contributor list for modelcontextprotocol/servers and the github.com/agentfinder org. De-dupe against your CRM. The delta is your immediate outreach list.
3. Build a small crawler (or ask Refolk to) for ai-catalog.json files across your target hiring-company universe. The companies that have published one are your hot market. The ones that have not yet are your sleeper pipeline.

The 30-day-old spec is doing your top-of-funnel work for you. Run the play before everyone else does.

FAQ

How is AGENTS.md different from a normal README for sourcing purposes?

A README describes what a project does for humans. An AGENTS.md describes the rules an autonomous coding agent must follow inside the repo. Authoring one means the engineer has actually run agents against their own codebase in production, hit failure modes, and codified the guardrails. That is a much harder credential to fake than "interested in AI" on a profile, and the Matt Nigh 2,500-repo study from November 2025 gives you a public rubric for separating serious specs from cargo-culted ones.

Is "MCP server contributor" really a finite enough pool to be useful?

Yes, if you focus on the steering layer. There are 9 core lead maintainers of MCP under the Agentic AI Foundation, 58 total active maintainers across working groups, and 2,900+ named Discord contributors. The broader 15,926 repositories with the mcp-server topic and 10,000+ active public servers give you a wider IC pool, but the maintainer graph is small enough to enumerate by name in an afternoon and is the highest-signal slice for senior or staff roles.

How long will this sourcing arbitrage last?

Probably one to two quarters. ATS vendors and LinkedIn will eventually parse AGENTS.md commits and add an MCP skill tag, and the named maintainer pool will get expensive once comp benchmarks adjust. The growth from roughly 2,500 AGENTS.md repos in November 2025 to 60,000+ by Q2 2026 (a 24x jump) tells you both how fast the signal is becoming meaningful and how fast it will commoditize. Run the play now.

What if my target company has not adopted ARD or AGENTS.md yet?

That is also a signal, just an inverse one. If a company claims to be "AI-native" in their fundraising or careers page but has no ai-catalog.json and no AGENTS.md in their public repos, you have a real prior on how mature their agentic work actually is. For candidates, it can mean greenfield opportunity. For acquirers, it can mean the team has not shipped past the demo. Either way, the absence of the file is information you did not have 30 days ago.