Toward Evolutionary Intelligence: LLM-based Agentic Systems with Multi-Agent Reinforcement Learning