GitHub's June 17 Agent Finder Just Named 15,926 AI-Native Engineers
GitHub's Agent Finder and the ARD spec turned "AI-native engineer" into a verifiable public artifact. Here's how to source it before LinkedIn indexes it.
On June 17, 2026, GitHub shipped Agent Finder, a plain-language search over an index of MCP servers, skills, agents, and tools built on the new Agentic Resource Discovery (ARD) spec co-developed with Google, GoDaddy, Hugging Face, and Microsoft. For the first time, "AI-native engineer" is not a résumé claim. It is a signed, domain-verified public artifact. If you are hiring agentic engineers in Q3 2026, this is a sourcing surface that did not exist three weeks ago, and it is going to get scraped, enriched, and picked over by every sourcing tool in the market within 90 days.
The play right now is simple: source directly from ARD publishers, MCP registry contributors, and the ards-project GitHub org before LinkedIn keyword search catches up. This post is a working guide.
What Agent Finder actually is, in sourcing terms
ARD is a v0.9 draft spec. The mechanics matter for recruiters because they define the signal.
A publisher hosts an ai-catalog.json at a well-known path on their own domain, listing the tools, MCP servers, agents, or APIs they expose. Registries (GitHub's Agent Finder, Hugging Face's Discover Tool, Google's forthcoming Agent Registry in Gemini Enterprise) crawl those catalogs and answer discovery requests. Publisher identity is verified through domain ownership, and production publishers can attach trust metadata so an agent can confirm the publisher's cryptographic identity before connecting.
Translation: every valid ARD catalog is cryptographically tied to a real domain owned by a real human or team. You cannot fake it without controlling DNS. Compare that to "AI Engineer" in a LinkedIn headline, a phrase that has been diluted since 2023.
Eleven companies published ARD at launch: Google, Microsoft, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, NVIDIA, Salesforce, ServiceNow, and Snowflake. Apache 2.0 licensed. Data model maintained by a Linux Foundation working group.
Why the signal is unusually clean right now
Two weeks after launch, an independent field test of all 11 launch partners found that none of them were serving a catalog on their main domain yet. The one exception was Hugging Face, running a live registry on a subdomain. Everyone else was still in slide decks.
That means the earliest ARD publishers in the wild are not enterprise laggards. They are indie developers and small teams who read the spec on June 17, understood it, and shipped by June 19. That is the population you want to be sourcing.
Suganthan Mohanadasan is the archetype. He shipped ARD on his own site within two days of the spec drop, built a public Agentic Resource Discovery Checker and a federated ARD registry, and filed a bug against Hugging Face's own deployment. That is a 72-hour ship cycle on a v0.9 spec with cryptographic trust metadata. If your JD says "self-directed, ships fast, comfortable with ambiguity," you wrote it about Suganthan.
The protocol-fluency filter
To publish a valid ARD catalog, a developer has to understand MCP, A2A (agent-to-agent), Skills, and trust manifests. They have to correctly implement Ed25519 signing and a revocation URL if they want production trust metadata. This is not a checkbox exercise. Ramanathan Guha, Microsoft's ARD lead, put the underlying problem this way on the Command Line blog:
AI can only use what it's been explicitly wired to use. Everything else may as well not even exist.
Publishers who wire it correctly are, by construction, the engineers you want building your agent stack. They have shown they can read a draft spec, implement it against a moving target, and reason about the trust boundary between an agent and an external tool. That is a much sharper filter than "starred langchain."
The five sourcing surfaces to work this quarter
1. The ards-project GitHub org
Contributors to ard-spec, connectors, and docs are a public, high-signal candidate list. The connectors repo alone ships client-side integrations for Claude, ChatGPT, Copilot, and Gemini. Anyone with merged PRs there is doing production-grade cross-vendor agent work in the open.
Pull the contributor graphs, dedupe against employer domains, and filter out the obvious Google, Microsoft, and Hugging Face committers if you cannot poach from those companies. What is left is your shortlist.
2. .well-known/ai-catalog.json crawls
Because ARD uses a well-known path, you can crawl for it. Point a crawler at the top 100k developer-adjacent domains and log every 200 response. Every hit is a self-declared AI-native engineer or team whose domain you now have. This is the surface that will disappear first, because it is exactly what Google's Agent Registry and Hugging Face's Discover Tool are built to index.
For most recruiting teams, standing up a crawler and joining it to employer, seniority, and location data is not a weekend project. This is where Refolk fits: you describe the person in plain English ("engineers who published an ARD catalog on their own domain since June 17 and have MCP server repos on GitHub") and get a ranked shortlist across GitHub, LinkedIn, and the open web. The crawl, the join, and the enrichment happen underneath.
3. The MCP registry pool
The official MCP Registry API returned 9,652 latest server records and 28,959 server/version records as of May 24, 2026. Anthropic cites more than 10,000 active public MCP servers. One third-party registry, Glama, indexed close to 20,000 servers by 2026. MCP SDKs pulled on the order of 97 million downloads a month.
number: 97M
label: Monthly MCP SDK downloads in 2026
note: The adjacent talent pool. Not every downloader is a builder, but every builder is a downloader.