Cost Compounding

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

How do you define a cost per task budget for an agent and enforce it at runtime?

Why does token cost grow super-linearly with agent step count, and what architectural decisions reduce it?

Why does token cost in a naive agent loop grow quadratically with step count?

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