Agent Vs Chain

When is using an agent architecture an engineering mistake, even when LLMs are involved?

What signal tells you to use a fixed LCEL chain instead of an agent loop?

When should you use a fixed DAG pipeline instead of a dynamic agent loop?