Until an LLM or some new form of AI can manage an entire architecture itself, I say we need a middle ground. Something will always go wrong, and understanding of the system is necessary to troubleshoot.