Capability, Context, and Skills with Sebastien Levert

The word “capability” is everywhere this year — and most of the time the listener is left to figure out what anyone actually means by it. For episode 25 of the Naftiko Capabilities Podcast we sat down with Sebastien Levert, a product manager on Microsoft 365 Copilot extensibility, and asked him for the working definition that survives contact with real developers building real things.
Sebastien comes to capability from inside one of the largest agent platforms in production. Before Copilot extensibility he worked on the Microsoft Graph — one of the biggest enterprise APIs out there — and his answer here is grounded in that lineage. A capability inside M365 Copilot is a native, scoped agent that lives inside the experience the user is already in, with access to the full connected data estate underneath it. You don’t jump to another product to get the job done — you stay in the experience and the agent brings the work to you.
The conversation moves from there into Work IQ as a context-engineering primitive. Work IQ is the “pre-AI” endpoint that gives an agent the connected data graph — people, documents, meetings, conversations, tasks — without making each developer rebuild it. Two ways to contextualize Work IQ surface naturally: prompt it well in a single shot, or scope a swarm of specialized agents at it and let each one give you back a narrower slice. Sebastien draws the parallel to the fleet-of-coding-agents pattern people are running in their terminals today, applied to the productivity side of work instead of the code side.
The episode lands on where Agent Skills fit relative to MCP, and Sebastien’s framing is the cleanest I have heard yet: MCP is the tool set on the workbench, Skills are the guide that tells you which tools to pick up for which job. Skills are the orchestration layer between raw tools and actual workflows — the agent takes on the hat of a scenario, knows which sequence of MCP calls to make, and gets it done. Capability, context, and skills are showing up in every enterprise conversation right now, and this episode is a short, dense take on why those three words point at the same architecture from three different angles.
This is part of an ongoing conversation with Sebastien — we’ll have more from him in upcoming episodes about how the M365 Copilot ecosystem is reaching developers and where he sees agent extensibility going next.