mcp
Protocol, Capability, Loop
Three Anthropic courses, taken mostly to check my own habits against someone else's definitions. Read side by side they describe how a model reaches the world, what it can do, and who actually runs the loop.
MCP, Agent Skills, Claude Code ... three courses that turn out to describe how a model reaches the world, what it can do, and who actually runs the loop.
I took all three, mostly out of a suspicion that I had been building these things by feel for a while and ought to check my work against someone else's definitions. The certificates are on my profile now, which is a slightly odd sensation after four decades of writing software, but the papers are not really the point.
The point is what the three of them look like when you put them side by side. MCP is the protocol layer: how a model gets access to tools and data at all, in a way that is not a pile of bespoke glue. Agent Skills are the capability layer: instructions and resources packaged so an agent can pick them up when the situation calls for it, rather than being told everything in advance. Claude Code is where both meet the actual work, which is to say the loop, the context, the moment something goes wrong and you have to decide whether the model was confused or you were.
Read in that order it becomes rather obvious that the model itself is the smallest part of the arrangement. Everything that decides whether the thing is useful sits around it ... the protocol, the packaging, the harness, the person who still has to know what "good" looks like.
The courses did not tell me that. They just made it hard to keep pretending otherwise.