You'd think AI would be great at running a D&D game. The whole game's structured hallucination, and the rules are written down. The first week was magic. Then it got boring, repetitive, and kind of bad at remembering anything that happened more than an hour ago.

Fixing that's what got me into the actual guts of agentic AI. And after a while it stopped being a D&D problem and started being the same problem people at work keep asking me about: where does this capability go? Is it a prompt I paste in? A skill the model loads when it feels like it? An MCP server I stand up and let any AI client call? Who owns it when it breaks?

This talk's about this adventure and its twists and turns: Agent Skills vs. MCP servers, with D&D as the running example because the failure modes are funny instead of career-ending. I'll walk through what I built and learned along the way.

There's also the boring half, which is the half that matters for the real world: governance. Who owns the capability. How it gets versioned and shipped. Where the secrets live. Who's allowed to call it. How you know what it's doing. What it costs. How you roll back when somebody pushes a change that makes every NPC speak in haiku.

This'll give you a working mental model for how to architect and manage your corporate AI projects. And give you an excuse to play D&D on company time.