How do you cap runaway costs in agentic loops without breaking the agent?

Our agent can call tools and loop through reasoning steps. When it gets confused it sometimes enters a spiral — 20+ steps, $4 in inference, no useful output. We need hard cost controls but naive step-count limits break valid multi-step tasks that legitimately need depth. How are teams distinguishing "agent is working hard" from "agent is stuck in a loop" and intervening appropriately?