The architecture mistake that makes multi-agent systems fail
If you're building with multiple AI agents, here's something most people get wrong: They think an agent and a sub-agent are interchangeable — just different names for the same thing. They're not. An Agent is a persistent brain. It has a workspace, memory files, identity. It stays alive and works on evolving problems. A Sub-Agent is a temporary worker. You spawn it for a specific task, it completes the task, and it disappears. No memory. No persistence. Gone. The critical difference: If you try to set up a sub-agent like a persistent agent (with workspace files, memory infrastructure, ongoing state), you're adding complexity to something designed to be simple. If you use agents as simple disposable workers, you're wasting their potential. Match the tool to the job: • Agent → ongoing, evolving work • Sub-Agent → isolated, one-off tasks This shapes how your whole system works.