One can easily imagine a generalist LLM outputting an action token which represents prompting the specialized LLM, which then gets routed to the specialized LLM, then the response is formatted and put into context by the generalist.