DeepSeek's Super-agent essay says human attention, not model capability, now limits multi-agent work. Is it an obituary for the attention economy?