The bottleneck isn't managing the 'army', it's the interaction latency of a single agent. An AI 'employee' that makes a customer wait 800ms for a voice response feels broken. Getting that full STT->LLM->TTS->render loop under 250ms is the real barrier. This is the problem we're focused on with Nyx: giving the agent a real-time face and voice. An agent that can't hold a conversation isn't much of an employee. Any feedback appreciated: NyxClaw.ai