Multi-Agent Reinforcement Learning Needs More Than Better Rewards

Multi-agent RL does not mainly have a modeling problem. It has a systems-design problem. Real-world coordination needs explicit task stages, selective communication, local adaptation, and hard safety layers. Until those become standard design assumptions, multi-agent intelligence will keep looking more ready than it really is.