GitHub Shipped .agent.md on May 13. LinkedIn Still Returns 14 People.

If you have a req open for an engineer who can actually build with MCP servers and orchestrate agents, you already know the LinkedIn search is broken. A US-wide query for "MCP server" in headlines returns roughly 14 people. The population shipping production MCP code is orders of magnitude larger. As of May 13, 2026, GitHub made that hidden population searchable, and most recruiters haven't noticed yet.

What actually shipped on May 13 and 14

On May 13, GitHub rolled global .agent.md support into Copilot for JetBrains and shipped the Copilot CLI agent with a unified sessions view. Global custom agents now live as .agent.md files under ~/.copilot/agents, available across every workspace. That standardizes the filename and the path developers will commit to their repos.

Same day, GitHub launched the Agent Tasks REST API. Copilot Business and Enterprise users can now kick off cloud agent tasks programmatically. The cloud agent works in the background in its own environment, makes and validates code changes, and opens a PR. On May 14, the Copilot App hit technical preview, a desktop client that orchestrates parallel agents, each in its own git worktree and branch, all working the same repo without knowing the others exist.

Then on May 18, the audit API exposed something even more important for sourcing: a repository's Copilot cloud agent setup, including its MCP server configuration, enabled tools, GitHub Actions workflow policy, and firewall configuration. MCP server config is now a first-class repo artifact. It is also, by accident, the single highest-signal hiring filter for agent-fluent engineers that exists right now.

Why this is a sourcing signal, not a Copilot trivia question

A committed .agent.md, AGENTS.md, or .github/agents/*.md file is behavioral proof. It is not "Copilot user" in a bio. It is a developer who sat down, wrote out which MCP servers they trust, declared their model preferences (Claude, Codex, GPT-5.3), set their test framework and package manager defaults, and configured auto-approve rules with a real security posture. You are reading their working style in their own words.

A résumé claims "experience with AI agents." An AGENTS.md says: "Use the Postgres MCP server for read queries against the staging branch, never prod. Prefer Claude for refactors over 500 lines. Run pnpm test:integration before any PR. Do not auto-approve writes to /infra." That is the most useful pre-qualification artifact a recruiter has ever had access to, and it sits in public GitHub.

14

60,000+

The 60k projects number matters because the file format is already mainstream among the people writing agent code. The 14 number matters because none of them have updated their LinkedIn headline to reflect it. The arbitrage is the gap.

The exact GitHub queries to run today

GitHub code search indexes filenames and contents. Start with the file itself.

File-name queries

• path:**/.agent.md returns every committed global custom agent definition. As of May 13, this filename did not have a stable convention in public repos. Anyone with one now is, by definition, early.
• path:AGENTS.md returns the broader open format. Filter by stars, recency, and contents.
• path:.github/agents/*.md returns enterprise-style custom agent definitions stored in .github-private style layouts. These are usually committed by senior platform engineers.
• path:AGENTS.override.md OR path:.agents.md OR path:TEAM_GUIDE.md catches Codex's discovery order variants. A single contributor's GitHub footprint can contain multiple variants of the same fingerprint.

Content queries that actually filter for MCP fluency

• path:AGENTS.md "mcp" narrows to AGENTS files that mention MCP at all.
• path:AGENTS.md "mcpServers" finds files that declare server config inline.
• path:AGENTS.md "@modelcontextprotocol" finds devs naming specific MCP packages.
• path:AGENTS.md "auto-approve" finds devs thinking about security posture, not just plumbing.

The May 18 audit API confirms what the queries imply: MCP server config is being treated as security-relevant infrastructure. That is the strongest possible signal that operating MCP servers is becoming a distinct engineering specialty, not a Copilot power-user gimmick.

The repos to crawl first

Three orgs concentrate the talent.

github/awesome-copilot is the canonical reference. The repo defines the agents/ directory convention explicitly: an agents/ directory of Custom GitHub Copilot agent definitions stored as .agent.md files. PR authors and contributors here are a sourcing goldmine, and the file structure tells you exactly what to grep.

microsoft/skills ships 128 skills with 1,158 test scenarios. Contributors demonstrate MCP plus skills fluency at Microsoft scale. Most are FTE, but the external PR authors are the ones to message.

agentsmd/agents.md is the open-format repo itself. PR authors are a curated list of the people thinking about agent-format standardization, which is roughly the top decile of the population.

Then there are the small AI-tooling startups: Mastra, Factory, Builder.io, Letta. Their engineers regularly author public AGENTS.md examples and MCP server code. Their GitHub orgs are small and concentrated. You can read every recent PR by hand in an afternoon.

A résumé claims experience with AI agents. An AGENTS.md tells you which MCP servers they trust, which model they reach for, and what they refuse to auto-approve.

One more name worth crawling: intellectronica/ruler, referenced as the tool that distributes MCP configurations across agent files. Its contributor graph identifies people building MCP tooling, not just consuming it. That is a meaningfully different population from "wrote an AGENTS.md once."

The part where Boolean stops working

The reason this is hard is not the queries. It is the join. You can pull a list of 4,000 GitHub usernames who committed an AGENTS.md referencing an MCP server in the last 90 days. Now you need: current employer, seniority, location, work authorization, whether they have shipped a real MCP server (not just consumed one), whether their stack matches yours, and a contact channel that is not a LinkedIn InMail they will ignore.

That is the friction we built Refolk for. You describe the person you want in plain English ("staff engineer in the US who has committed an AGENTS.md file mentioning MCP in the last 60 days and currently works at a sub-200-person AI tooling company") and get a ranked shortlist with GitHub, LinkedIn, and open-web profile data joined. The .agent.md sourcing pattern is exactly the kind of signal that lives outside LinkedIn's schema, which is the entire point.

What good agentic developer recruiting looks like now

The interview question writes itself. Send the candidate their own committed AGENTS.md and ask them to walk you through why they configured it that way. Why that MCP server and not the other one. Why they auto-approve reads but not writes. Why Claude for refactors and Codex for boilerplate. You learn more in 20 minutes than two technical screens used to surface.

For find AI agent engineers GitHub workflows, this also fixes the "ghost candidate" problem. The best agentic engineers are often senior ICs at small AI-tooling startups who haven't touched LinkedIn in two years. They ship publicly visible AGENTS.md commits weekly. GitHub code search sees them. Hiring Assistant does not.

Context: why the searchable population just multiplied

Until March 2026, agent mode was VS Code only. That excluded a huge chunk of Java, Kotlin, and Python developers who live in JetBrains. The March GA on both editors, followed by the May 13 global .agent.md rollout in Copilot for JetBrains, means the searchable population just multiplied. Engineers who never touched VS Code are now committing the same fingerprint file.

Adjacent tools matter too. Anthropic's redesigned Claude Code desktop launched in April 2026. Cursor 3 arrived weeks earlier with an Agents Window that runs many agents in parallel. OpenAI shipped native Codex for Windows in March. Engineers fluent across these tools are exactly who hiring managers want. They typically commit AGENTS.md or CLAUDE.md. Sometimes both, in the same repo. That dual commit is itself a filter: cross-tool fluency.

One archetype worth naming. The DevOps Journal profiled a developer claiming 368 PRs merged in the last 30 days using the Copilot App. That is the agent-orchestrating engineer archetype. Three parallel agents per worktree, the human in the steering role. If you are hiring for a platform team that needs to 10x throughput on a small headcount, this is the GitHub Copilot CLI hiring signal you want to filter on, and committed .agent.md files are how you find them at scale.

The window is short

.agent.md is two weeks old as a public convention. Right now, anyone with a hand-written one referencing MCP servers is by definition early to the pattern. Once the Copilot App hits GA (likely within 6 to 12 months), the file will become as common as .eslintrc, and the signal-to-noise will collapse. The arbitrage exists for one hiring cycle, maybe two.

If you are running sourcing MCP engineers as a workflow this quarter, the order of operations is:

1. Run the four path: queries above weekly.
2. Diff the new committer list against last week's. Net-new committers are your hottest leads.
3. For each, read the actual file. Pre-qualify on stack fit before the first message.
4. Open with the file. "I saw your AGENTS.md in your-repo. The way you handle auto-approve on the Postgres MCP server is exactly what we need for X." Response rates on opens like that beat generic InMails by a wide margin.

This is also where Refolk earns its keep: the diff-and-rank step is tedious by hand and trivial to ask for in plain English. You describe the cohort, you get the shortlist, you keep the time for the conversation.

FAQ

Is .agent.md the same as AGENTS.md?

They are part of the same family but not identical. AGENTS.md is the open format used by 60,000+ projects, stewarded by the Agentic AI Foundation. .agent.md is the specific filename GitHub standardized on May 13, 2026 for global custom Copilot agents, stored under ~/.copilot/agents. Codex also looks for AGENTS.override.md, TEAM_GUIDE.md, and .agents.md. For sourcing, search all of them. They surface overlapping but non-identical populations.

Will LinkedIn add a field for this?

Eventually, but not soon enough to matter. LinkedIn's skills taxonomy lags behind tooling reality by 12 to 24 months. There is currently no "MCP servers" skill, no "agent fluent" headline category, and no way to query "has committed AGENTS.md to a repo with greater than 50 stars in the last 90 days." GitHub code search can answer that today. By the time LinkedIn catches up, everyone will have one and the signal will be gone.

What about developers whose .agent.md files are in private repos?

You will miss them, which is fine. The point of this signal is not coverage, it is precision. The developers committing .agent.md to public repos are over-indexed on open-source contribution, public reputation, and willingness to be contacted. That is exactly the population recruiters convert on. For private-repo engineers, you fall back to enterprise audit-API data (if you are an internal recruiter at their company) or to adjacent public signals like conference talks and blog posts.

How do I avoid sounding creepy when I reach out about a file they committed?

Reference the file by name and ask a real engineering question about it. "I noticed in your AGENTS.md you route writes through a separate MCP server. Was that a perf decision or a blast-radius decision?" That is a peer question, not a recruiter ping. The engineers writing these files are flattered when someone has actually read them. They are insulted when you paste "Hi, I see you have GitHub experience" into their inbox.