Sourcing "Phase 3" Agent-First Devs From Outside GitHub

On August 7, 2026, GitHub shipped a Copilot impact dashboard update that finally names the developer every eng leader wants to poach: "agent-first," split into Phase 2 and Phase 3 cohorts, with cost-per-dev and PRs-per-dev attached. The problem is that the label lives inside the target company's dashboard. If you are sourcing from outside, you need external proxies that map to the same person.

This post reverse-engineers those proxies. Every one of them is a GitHub filename filter, an MCP registry lookup, or a repo-structure pattern you can run today.

What GitHub actually shipped on August 7

GitHub added a "Potential return on investment" section to the Copilot impact dashboard that formally splits developers into two tiers. Passive users and Phase 1 sit on chat plus completions. Phase 2 and Phase 3 are the "agent-first" cohort, and each card exposes cost per dev per month from AI credit consumption, that cost as a share of developer payroll, and average pull requests per developer per month.

The consequence: "agent-first" is now a compensable category, not a persona label. If a hiring manager benchmarks a candidate against their own Phase 3 cohort, they are comparing PR throughput and credit spend against a named internal tier.

One day earlier, on August 6, GitHub also shipped enterprise MCP allowlists. New allowedMcpServers and deniedMcpServers keys in enterprise managed settings let owners name which Model Context Protocol servers Copilot clients may run. Model Context Protocol (MCP) is the open standard for connecting AI assistants to tools and data. The people who get servers onto internal allowlists are the same people who publish public MCP servers, and that authorship is externally visible.

4.7M

Why LinkedIn headlines miss the entire cohort

The headline-search era is over for this role. In Refolk's index, 5,252 U.S. professionals list LangGraph as a skill, but the count with "MCP," "Claude Code," or "AGENTS.md" in their headline rounds to zero. Sourcing AI-native engineers on headline keywords will miss almost the entire pool.

Two mechanisms explain this. First, the artifact layer moves faster than the self-description layer. Devs ship a .github/agents/ directory months before they update a headline. Second, the top of the LangGraph-skilled U.S. pool skews to CTO, Co-Founder & CTO, and Founding Agentic Platform Architect titles. Founders do not rewrite their headline every time they adopt a new tool.

Which means the poach-able ICs are underindexed on LinkedIn and overindexed on commits. That is the exact gap Refolk closes: you describe the person in plain English ("US engineers who publish MCP servers and maintain AGENTS.md at monorepo scale"), and get a ranked shortlist pulled from GitHub, LinkedIn, and the open web rather than from headline text alone.

The external proxy stack, in order of signal strength

Here are the public artifacts that map to GitHub's internal Phase 2/3 label, ranked from weakest to strongest.

1. A single root AGENTS.md file. Table stakes. The convention was formalized in August 2025 through collaboration between OpenAI, Google, Cursor, Factory, and Sourcegraph, and is now adopted by over 20,000 repositories. Treat filename:AGENTS.md like filename:Dockerfile circa 2015.
2. .cursor/rules, CLAUDE.md, .github/copilot-instructions.md. Multi-vendor evidence. Experienced developers use an average of 2.3 AI coding tools, so filter with OR, not AND.
3. .github/agents/*.md. Direct Phase 3 evidence. This is where Copilot custom agents like @docs-agent, @test-agent, and @security-agent are defined.
4. 5+ AGENTS.md files in one repo. Monorepo-scale maintenance. OpenAI's own openai/codex repo ships 88 AGENTS.md files across its directory tree. File-count beats file-presence.
5. hooks/ directory or hooks.json alongside AGENTS.md. The Phase 2 vs Phase 3 dividing line. Prose guidance alone gets 25 to 40% agent compliance; enforced runtime hooks hit ~95%. A candidate shipping hooks is meaningfully more advanced than one shipping prose.
6. Authored server in the MCP Registry under io.github.username/*. The strongest signal. Reverse-DNS names are tied to verified GitHub accounts, so publication is legally-attributable identity, not a self-claim.

The awesome-copilot taxonomy as a search cheat sheet

The github/awesome-copilot repo lays out the canonical directory taxonomy: agents/, instructions/, skills/ (with SKILL.md), hooks/, workflows/, plugins/, and extensions/. Each of those directory names is a distinct GitHub filename filter that maps to a different level of agent fluency. Sourcing AI coding agent builders becomes a matter of composing these filters, not scrolling LinkedIn.

How thin the addressable pool actually is

The public-artifact cohort is roughly 1 in 235 of the paid Copilot base. AGENTS.md repos (20,000) divided by paid Copilot seats (4.7M) equals ~0.4%. Even assuming a typical Phase 2/3 dev maintains three AGENTS.md-carrying repos, the unique-author universe is likely low six figures globally. That is small enough that cold outreach math works and large enough that you cannot brute-force it by hand.

Here is the table-ready view from Refolk's index and public sources:

1,384

Where these people already work

The employer distribution in Refolk's index for this cohort clusters in two groups. First, big-tech AI orgs: Meta, AWS. Second, agent-native scaleups and platforms: Voya AI, Turing, inVia Robotics, Notabene, plus stealth-AI startups concentrated in the SF Bay Area, Palo Alto, and NYC (the top LangGraph-skill regions).

Two implications for outbound. The stealth cluster means job-title filters will underperform: someone building an MCP server at a stealth SF startup may show as "Founding Engineer" with no agent-adjacent keyword. And the big-tech contingent is more likely to be a Phase 3 dev inside an existing Copilot deployment, which means the compensable-cohort argument in an outreach message lands harder there than at a scaleup that has not yet adopted the dashboard.

A workflow you can run this week

1. Pick the strongest proxy your ATS can hold: MCP Registry authorship, then hooks/ next to AGENTS.md, then .github/agents/*.md.
2. Pull authors and contributors from modelcontextprotocol/registry, modelcontextprotocol/servers, github/awesome-copilot, and agentsmd/agents.md. These four repos are the primary graph.
3. Cross-reference against openai/codex-style monorepos in your target verticals to find devs who maintain agent workflows at scale rather than shipping one root file.
4. Do not filter on LinkedIn headline. Filter on artifact, then enrich with title and location.
5. For the ranked shortlist step, describe the person to Refolk in plain English instead of building a boolean string that will miss the artifact-layer identity.

The reason this order matters: each step upstream removes false positives that would otherwise burn recruiter hours on Phase 1 devs cosplaying as Phase 3. MCP registry authorship alone eliminates roughly everyone who wrote a demo AGENTS.md and never shipped a tool their agent could call.

FAQ

What is the difference between GitHub Copilot Phase 2 and Phase 3?

Both sit inside the "agent-first" tier of the August 7, 2026 Copilot impact dashboard, but GitHub's public changelog does not publish a per-phase capability definition. Externally, the cleanest proxy for the Phase 2 to Phase 3 step is enforcement: a candidate whose repos contain a hooks/ directory or hooks.json alongside AGENTS.md is enforcing agent behavior at runtime (~95% compliance) rather than describing it in prose (25 to 40%). Treat hooks as the dividing line until GitHub publishes its own.

Is AGENTS.md a Copilot-only signal?

No. AGENTS.md was formalized in August 2025 as a vendor-neutral spec through OpenAI, Google, Cursor, Factory, and Sourcegraph, and it is stewarded by the Linux Foundation's Agentic AI Foundation via the agentsmd/agents.md repo. Copilot custom agents in .github/agents/*.md are Copilot-specific, but a root AGENTS.md file is cross-vendor evidence. Given developers average 2.3 AI coding tools, expect the same person to also ship CLAUDE.md or .cursor/rules.

Why is MCP Registry authorship a stronger signal than GitHub stars or npm downloads?

Because the registry uses reverse-DNS namespaces (io.github.username/*) tied to verified GitHub accounts or domains. Only the legitimate owner of that account can publish under that namespace. Unlike npm or PyPI, where handles can be squatted or ambiguous, a name in the MCP Registry is a first-party identity primitive. You are sourcing a person, not disambiguating a handle.

How do I estimate my realistic addressable pool?

Start from the 4.7M paid Copilot base as of January 2026. Assume 10% agent-first as an upper bound, which yields ~470K globally. Filter to your country and target titles. For a U.S.-only search over engineers with LangGraph as a listed skill, Refolk's index shows 5,252 people; adding an MCP skill and an agent-engineer title narrows the global pool to 1,384. The public-artifact cohort (AGENTS.md authors) is roughly 0.4% of paid Copilot seats, so plan outreach volume accordingly.