Anthropic says fold most subagents into one agent loop. We agree for tasks. For always-on AI employees, memory changes the answer. Here is why.