Controlling Performance and Budget of a Centralized Multi-agent LLM System with Reinforcement Learning