Only 93 People Call Themselves "AI Agent Engineer." Awesome-Copilot Has 346.

Every engineering org you talk to is hiring an "AI Agent Engineer" right now. Almost none of them can find one. The reason is mechanical: the title barely exists yet, and the people who can actually build agents don't call themselves that on LinkedIn. They call themselves Staff Engineer, or DevRel, or nothing in particular. But they have a git history that says otherwise, and most of it is sitting in one repo.

The 93-person mirage

Run a clean LinkedIn title search in the United States for "AI Agent Engineer," "Agent Engineer," and "Agentic AI Engineer." You get roughly 93 profiles. Sierra holds 13 of them. The rest scatter across Profound, Liberate, Planet DDS, Accenture, Day AI, and a long tail, with NYC and SF taking most of the geography.

Every Series B founder, every FAANG platform team, every consultancy with a generative AI practice is hunting that same list. The contact rate is brutal, the comp expectations are silly, and the conversion is worse than usual because Sierra and a handful of others are paying above market to keep them.

93

This is the part most sourcing teams misread. The shortage isn't of agent-building engineers. The shortage is of agent-building engineers who happen to have updated their LinkedIn headline. The actual population that has shipped working agent code is at least three to four times larger. They're just findable through a different graph.

What changed in the last 13 months

On October 28, 2025, GitHub shipped custom agents for Copilot. The mechanic is simple: drop a configuration file into .github/agents in any repo, and you've defined an agent persona with its own prompts, tool selections, and MCP server wiring. On November 18, 2025, the same feature went to JetBrains, Eclipse, and Xcode. At GitHub Universe 2025, the company announced AgentHQ, a deployment surface for agents that run inside the GitHub environment itself.

Then in May 2026, GitHub expanded agent skills and rolled out partner agents in public preview. The named partners are a tour of the modern infrastructure stack:

• Observability: Dynatrace Expert, Elasticsearch agent
• Security: JFrog Security Agent, StackHawk Security Onboarding
• Databases: MongoDB Performance Advisor, Neon Migration Specialist, Neon Performance Analyzer, Neo4j Docker Client Generator
• DevOps and IaC: Terraform Agent, Arm Migration Agent, Octopus Release Notes Agent, DiffBlue Java Unit Test Custom Agent
• Incidents and PM: PagerDuty Incident Responder, Monday Bug Context Fixer
• Feature flags: LaunchDarkly Flag Cleanup, Amplitude Experiment Implementation
• Automation and APIs: Apify Integration Expert, Factory.ai Code Spec Custom Agent, Lingo.dev Internationalization

Each of those agents was shipped by named engineers at the partner company. Their commits are public. That alone is a sourcing list you can't buy.

Why awesome-copilot is the cleanest sourcing pool of 2026

The github/awesome-copilot repo currently sits at 33.8K stars with 346 contributors. Microsoft's own developer blog describes its current contents as 175+ agents, 208+ skills, 176+ instructions, 48+ plugins, 7 agentic workflows, and 3 hooks, all contributed by the community.

346

A contribution to this repo is not a star. It's not a fork. It's a Markdown file with YAML frontmatter, an .agent.md extension, MCP tool wiring, and a git blame trail. The repo's own contributor-report.mjs script under eng/ already classifies contributors by file path, meaning the maintainers themselves have decided that "shipped an agent" is a different category from "shipped a skill" or "fixed a typo."

For sourcing purposes, that's a gift. You don't have to read every PR. You can rank candidates by the type of artifact they merged.

What counts as a real signal

Look at the PagerDuty Incident Responder entry. The description reads: "Responds to PagerDuty incidents by analyzing incident context, identifying recent code changes, and suggesting fixes via GitHub PRs." That's not a toy prompt. It's a workflow spec with tool calls, error handling, and a defined output contract. The engineer who wrote it has demonstrably built and debugged an agent that touches a production incident pipeline.

The MongoDB Performance Advisor and JFrog Security Agent are the same shape. So is Neon's migration specialist. These are people whose first interview should be a conversation about tradeoffs, not a screen on whether they understand what an LLM is.

A merged .agent.md file with MCP wiring is a code artifact that ran in production. A langchain star is not.

The four-graph workflow

If you only mine awesome-copilot, you'll get the agent designers. If you only mine MCP repos, you'll get the protocol engineers. The dual-skilled engineers, the ones founders actually want, sit in the intersection. Here are the graphs to cross-reference:

1. github/awesome-copilot for agent, skill, plugin, and workflow contributors.
2. modelcontextprotocol/servers for the reference MCP servers maintained by the MCP steering group.
3. microsoft/mcp for Microsoft's official catalog (Azure MCP, Foundry, AKS, M365 Agents Toolkit).
4. modelcontextprotocol/registry plus the curated wong2/awesome-mcp-servers list.

Cross-reference any two of those and you've narrowed to engineers who understand both the agent layer and the tool protocol underneath it. Cross-reference three and you're talking to people who could plausibly run an agent platform team.

The MCP protocol itself was, per GitHub's own write-up, one of the fastest-trending open-source projects on the platform in its first week, with contributors building custom servers, tool registries, schema validators, and language-specific clients. Credit the creators David Soria Parra and Justin Spahr-Summers, plus registry maintainers Tadas Antanavicius from PulseMCP, Alex Hancock from Block, Toby Padilla (Head of MCP at GitHub), and Adam Jones from Anthropic. Their follower graphs alone are worth an afternoon.

The boolean nobody runs

Here's the search every recruiter is running today:

("AI Agent Engineer" OR "Agentic AI Engineer" OR "Agent Engineer") AND ("LangChain" OR "LangGraph" OR "MCP")

It returns the same 93 people, plus or minus noise. Now consider the search nobody runs:

contributor:github/awesome-copilot AND file:*.agent.md
AND (contributor:modelcontextprotocol/servers OR contributor:microsoft/mcp)

GitHub doesn't expose that as a single query. You'd have to script it. This is exactly the friction we built Refolk to remove: you describe the engineer in plain English ("people who merged an .agent.md file in awesome-copilot and also have commits in any MCP server repo, based in NYC or remote US"), and you get a ranked shortlist with the LinkedIn, GitHub, and open-web identities already stitched together.

The partner-agent shortcut

If you want to skip the protocol cross-reference and just go after engineers who have shipped commercial agent code, the partner list does most of the work for you. Every named partner agent in the May 2026 release was authored by employees of the partner company. That means:

• Want to hire from Dynatrace's agent team? Pull the contributors on the Dynatrace Expert agent.
• Building an incident-response product? The PagerDuty Incident Responder authors are your bullseye.
• Need someone who understands feature-flag tooling and agent orchestration? LaunchDarkly Flag Cleanup contributors.
• Database-adjacent infra hire? MongoDB Performance Advisor, Neon Migration Specialist, Neo4j Docker Client Generator.
• DevOps with real IaC chops? Terraform Agent, Arm Migration Agent, Octopus Release Notes Agent.

These aren't guesses. The contributors shipped customer-facing agent code at those companies. Their PRs are public. Their commit emails often resolve to corporate domains. Their LinkedIn titles, predictably, say "Senior Software Engineer" or "Principal Engineer," which is exactly why title-search recruiters are missing them.

A useful Refolk query here is something like "engineers at Dynatrace, JFrog, MongoDB, Neon, LaunchDarkly, or PagerDuty who have public commits to github/awesome-copilot in the last 12 months." That returns the partner-agent author list with current employer and seniority, ready for outreach.

How to actually read a contributor

Not every merged PR is equal. Use this triage when you open a candidate's profile:

Tier 1: shipped a domain agent

A merged .agent.md file in a specific domain (security, observability, databases, incidents) with MCP server wiring and a non-trivial tool list. This is the strongest signal in the repo. The engineer designed an agent persona, picked tools, handled the integration with an external system, and got it through review.

Tier 2: shipped a plugin or workflow

The repo's 48+ plugins act as a default marketplace for both Copilot CLI and VS Code. A plugin contribution is closer to "shipped a product" than "wrote a prompt," because it has install paths, error states, and a user surface.

Tier 3: shipped a skill or instruction

The 208+ skills and 176+ instructions are smaller artifacts. Useful as a co-signal, but not enough on their own. Pair with MCP repo activity before you spend an outreach slot.

Tier 4: typo fixes, docs

Real contributors, but not the signal you came for. Skip unless they show up in Tier 1 or 2 elsewhere.

What this changes for your pipeline

If you're hiring for an agent team in Q1 2026, the math is simple. The literal-title pool is roughly 93 in the US and getting picked over weekly. The awesome-copilot contributor pool is 346 globally, with adjacent MCP repos multiplying it several times over, and almost no one is sourcing from it because the workflow requires cross-referencing graphs that LinkedIn doesn't expose.

You don't need a new ATS or a new comp band. You need a different first query. Mine the contributor graph for the artifact, not the title. The engineers who actually shipped agents in 2025 and 2026 will be in that graph whether or not they ever updated their headline.

FAQ

How do I know an awesome-copilot contributor is real and not a drive-by PR?

Check the file they touched. A merged .agent.md with MCP tool wiring, a plugin in the marketplace, or an agentic workflow is real shipping work. The repo's own contributor-report.mjs already classifies contribution type by file path, so you can filter to the tiers that matter. Drive-by typo fixes are obvious in the diff.

Won't every recruiter start doing this once the article is out?

Some will try. Most won't, because the workflow requires cross-referencing four GitHub graphs, dedoping against current employer, and stitching identities to LinkedIn for outreach. That's a multi-day project by hand. Tools like Refolk collapse it into a single plain-English query, but the underlying insight, that agent-builders don't self-title as agent-builders, will hold for at least another hiring cycle.

What about contributors outside the US?

The 346 figure is global, and a meaningful share of contributors are in Europe, India, and Latin America. If you're hiring remote, that's a feature. If you're hiring onsite NYC or SF, filter by location after the graph query, not before, because the strongest contributors often don't list a location on GitHub.

Is starring langchain-ai/langgraph or similar repos a useful signal?

Weakly. A star is a bookmark. A merged PR with code that integrates an MCP server and defines an agent persona is a production artifact. Use stars as a tiebreaker between two otherwise-equal candidates, never as a primary filter.