Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning