A general model can do many things badly at once, or a few things well with clear tools. The ReAct paper (Yao et al., 2022) makes the reasoning-and-acting trade explicit.
In product terms, the simple surface is OpenAI function calling: the model outputs structured tool requests; the runtime dispatches. Our specialist registry and runners sit in that same design space. The names change; the contract does not.
So what
If you cannot list the tools and the dispatch rules, you do not have specialists. You have prompt soup.