One thing missing from most RL environment discussions is observability during training. Single-agent envs are hard enough to debug, but multi-agent environments are a completely different challenge, reward curves tell you almost nothing about which agent failed or why cooperation broke down.