what I learned mapping 941 AI agent skills

I run an AI agent system called NORD. It orchestrates specialists across departments, enforces standards, and handles the structural overhead I don't want to do by hand. I wrote about how I built it in an earlier post.

Once the system was running, I had a practical question: what does the landscape around it actually look like? What agent skills exist, who's building for which industries, and where is the open ground? I needed a real answer, not a surface read from scrolling GitHub on a Tuesday.

I ran a research project to find out.

the project

Two research waves. Thirty-three lanes. Nineteen functional lanes covering cross-industry capabilities like software development, security, finance, marketing, legal, and HR. Fourteen vertical lanes scanning underserved industries for white space. Every entry required a verifiable link before it made the database.

The sources were public: MCP registries including awesome-mcp-servers, Glama, and the official Anthropic registry. Claude Skills directories from ComposioHQ, VoltAgent, and Anthropic. Automation libraries from n8n and Zapier. GitHub search. The OpenAI GPT store. Both waves completed on the same day.

The final deduplicated count was 941 unique skills. The breakdown by type: 549 MCP servers, 204 Claude Skills, 129 automation recipes, 22 agent tools, 10 GPT/Actions, and a handful of specialized variants.

I built the anti-fabrication discipline into the methodology on purpose. Every entry needed a real link. Anything unconfirmable got tagged for re-verification. By the end of the project, zero entries were outstanding.

what the map looked like

The cross-cutting finding across all 33 lanes was pretty clear: the data and integration layer exists. There are public APIs and MCP servers for reading data from the dominant platforms in most industries. Single-step automation recipes are widely available.

What is almost entirely missing is the workflow-orchestration layer. The multi-step agent that ties data retrieval, reasoning, decision routing, document generation, and system write-back into one coherent operational flow. That layer is open in vertical after vertical.

The fourteen vertical scans ranked the industries from most to least underserved. The barely-served end of the list included wellness and mindfulness, construction and trades, agriculture, energy and sustainability, and healthcare niches. In those spaces, the public agent skill ecosystem had almost nothing. The incumbents for those industries, companies like Procore in construction, are large established platforms that own the customer relationship and the system of record. Their open MCP footprint was thin or nonexistent.

The first-order reading of this map was straightforward: lots of open verticals, lots of open workflows, lots of room to build.

I didn't trust that reading.

the red-team

Before drawing any conclusions from the map, I wrote a deliberate critique of it. Not a soft one.

The foundational problem was this: the entire research effort measured supply, not demand. A low count of agent skills in a vertical can mean an untapped market. It can also mean nobody shops in that aisle. Absence of tooling is at least as likely to be a verdict as an invitation. The research told me where supply was thin. It said nothing about whether demand existed, who held the budget, or whether anyone was actively trying to solve the problem.

The method had optimism bias built in. Each of the fourteen vertical scans was set up to surface opportunity, so each produced one. When you ask fourteen researchers where the opening is, you get fourteen openings, whether or not they're real. That's an artifact of the prompt design, not a finding about the market.

The moat argument was the other weak point. If orchestration means composing commodity MCP servers, anyone can compose them: the client, an open-source contributor, or the incumbent that already owns the data and the customer relationship. Orchestrating commodity parts is configuration. It isn't defensible. A new entrant here gets squeezed from above as incumbents add their own agent layers and from below as open source makes the building blocks cheaper.

The incumbents named in the vertical scans were already moving. Procore, Toast, Salesforce Agentforce, Guidewire, Blackbaud. They own the system of record and the channel. Any independent orchestration layer is competing on someone else's home field, and the window before those gaps close is twelve to eighteen months at the outside.

The capacity math was also a problem worth naming. We are a small team already committed to active products and client delivery. The map presented fourteen verticals as if they were a menu. They are not actionable in parallel. A single vertical done properly would consume the team.

The last critique was the uncomfortable one. The research project started from some skepticism about recycled AI skill packs being sold as market insights. What the map pointed toward was, in effect, becoming a more credible version of the same thing. A single MCP wrapper or skill file is cheap to build and cheap to copy. The value in that game is distribution and domain credibility. We don't have either in the industries that showed the most open space.

what actually holds up

Three things survived the critique intact.

The most defensible reading of the orchestration insight is not a product line. It's a positioning move. Offering an agent layer over the tools a client already runs, as part of an existing engagement, costs no speculative R&D. It tests real appetite with buyers I already reach. That's a much lower-risk way to find out whether the thesis has legs.

The database itself is genuinely useful in the role it was originally built for: internal intelligence. A reference of what's available so I can read the competitive landscape, build the right NORD skills, and avoid re-inventing things that already exist. Where the research overreaches is as a market-entry roadmap. That's not what it is.

The discipline that comes out of the red-team is demand validation. The map identifies where supply is thin. The missing half is finding out whether five real buyers in a vertical would pay, who holds the budget, and how long the sales cycle is. Until those questions have evidence-based answers, a vertical is a watchlist item, not a build.

why this is worth writing about

I'm not sharing this because the research produced some groundbreaking finding about AI agents. I'm sharing it because the process of doing honest research and then actively trying to talk yourself out of the conclusions is a useful one, and I don't see it discussed much.

Most AI market analysis I read leads with the opportunity and never gets to the critique. The supply/demand gap, the optimism bias in the method, the capacity math problem. Those are the honest observations, and they're the ones that actually change what I do.

What I am doing because of this research: using the database as ongoing intelligence, positioning the orchestration layer as an upsell within existing work, and treating vertical-specific builds as watchlist items until demand validation passes.

What I am not doing: chasing fourteen verticals while two products are mid-build and a revenue target is on the books.

That's the point. The real discipline is knowing what to set down.