so I played with MCP for a while last night and I think MCP is great as a layer to pull custom tools into the existing claude/desktop/chat experience. But at the end of the day its just basic agentic loop over tool calls.

If you want to tell a model to send a message to slack, sure, give it a slack tool and let it go wild. do you see a way how MCP applies for outer-loop or "headless" agents in a way that's any different from another tool-calling agent like langchain or crewai? IT seems like just another protocol for tool calling over the stdio wire (WHICH, TO BE CLEAR, I FIND SUPER DOPE)

I think the fact that MCP is two ways, and tool call is asynchronous (since it's over network), means that it should be possible to code an approval flow with a prompt and a slack approval tool?

Yeah it still takes a bit of work, but it's more extendable (can swap slack to email) and more versatile (human can message agent to interrupt or ask for clarification).

I haven't used langchain (beyond RAG) or crewai, not sure what was capable before.