It should be quite easy - first fast model evaluates complexity of the task and suitable model to perform it, then it executes on best suited model and returns answer to the user. Obviously multi-step models are not great for simple questions.