> * A thin glue layer orchestrating all of the above: injecting the list of available tools into the LLM context, parsing LLM output to detect tool calls and invoke them with appropriate args, and inject results back into LLM context

Yeah llm rules. You think there must be something more to it. There's not.