GitHub's June 9 .agent.md Files Are a Sourcing List for AI-Native Engineers

On June 9, 2026, GitHub shipped custom agents for Copilot CLI, defined by Markdown files committed directly to repos. That sounds like a developer-experience update. For anyone hiring engineers who actually ship with agents, it is a sourcing list with a 6 to 9 month head start on LinkedIn.

The artifact is small and grep-able: a file named something.agent.md, living at .github/agents/ in a public repo. The author is a name. The commit history is a resume. And because of a billing change that landed eight days earlier, the people iterating on these files are not just experimenting. They are optimizing for cost.

What actually shipped on June 9

GitHub's announcement (bylined by Jacklyn Lee) introduced custom agents for Copilot CLI as Markdown files that specify how an agent should operate, which tools it can use, what standards it should follow, and what outputs it should produce. Each file uses the .agent.md extension. Name an agent "Security expert" and you get security-expert.agent.md.

The file format itself is the interesting part for sourcers. It is a Markdown document with YAML frontmatter that declares the agent's name, description, available tools, and MCP server configurations. Authors who write tight frontmatter and scoped tool lists are demonstrating MCP literacy, prompt-engineering taste, and an understanding of subagent context windows. That is a much narrower cohort than "uses Copilot."

The prompt body is capped at 30,000 characters. Engineers who stay well under that with crisp, tool-scoped instructions are showing taste, not just adoption.

Where the files live

This is the part that matters for sourcing AI engineers on GitHub. Agent files are placed in .github/agents/ and version controlled. Push to the repo and every team member gets them automatically. That convention is the search surface:

• path:.github/agents extension:agent.md returns every public author of a Copilot CLI custom agent.
• filename:AGENTS.md returns the broader, cross-vendor agentic cohort (more on that in a second).
• Filtering for commits in the last 30 days isolates the post-June 9 adoption wave.

You do not need a private dataset to run these queries. You need to know they exist.

The June 1 billing change is what turns presence into quality

Eight days before custom agents shipped, GitHub moved every Copilot plan to usage-based billing. The announcement came from Mario Rodriguez, GitHub's Chief Product Officer. The mechanics: each plan includes a monthly allotment of GitHub AI Credits, where 1 AI credit equals $0.01 USD, and usage is calculated on token consumption (input, output, and cached) at listed API rates per model. Copilot Business is $19/user/month including $19 in credits. Copilot Enterprise is $39/user/month including $39 in credits.

The reaction was immediate. One developer on the $39 Copilot Pro+ plan reported burning roughly 8% of their monthly AI credit allotment in two hours. Another reported spending more than $6 on a single change request and called consumption "impossible to predict."

8%

Custom agents directly attack this problem. GitHub's own docs explain why: work performed by a custom agent runs in a subagent with its own context window, so parts of a task can be offloaded without cluttering the main agent's context. The main agent then handles higher-level planning. In plain English: tight .agent.md files reduce token spend.

That is the filter you want. Anyone can commit an .agent.md. Engineers who iterate on those files (look at commit history, not just file existence) are optimizing for cost-efficiency under metered billing. That is the rare trait.

Why LinkedIn will not show you these people for months

Search "AI Engineer" on LinkedIn today and you get an ocean of ML and data folks who self-titled in 2023 and 2024. The shipping cohort, the people writing agent profiles and wiring MCP servers into their CI, mostly have titles like "Senior Software Engineer," "Platform Engineer," or "Staff DevEx." They have not retitled yet. They will, in 6 to 9 months. Until then, the GitHub artifact is the only public signal that catches them.

This is the structural reason sourcing tools built on LinkedIn keyword search miss the cohort entirely. The label has not caught up to the work. We built Refolk for exactly this gap: you describe the engineer you actually want (someone shipping .agent.md files in the last 60 days, prior MCP server work, lives in Berlin) and get a ranked shortlist across GitHub, LinkedIn, and the open web. The artifact does the labeling that the headline has not done yet.

The label has not caught up to the work. The artifact is doing the labeling for you.