Picks the cheapest model that can handle each task.
A router agent that classifies each request's difficulty and dispatches it to the lowest-cost capable model — Mini for simple Q&A, Pro for complex reasoning. Cuts token costs 40–70% on mixed workloads.