How to stop AI agents from looping when a tool call fails

My customer support agent spent twenty minutes last Tuesday calling the same non‑existent database endpoint over and over. The agent burned through half my API budget before I noticed the terminal screen spinning uncontrollably.

The issue happened because the language model treated a timeout error as a transient glitch rather than a permanent dead end. Without a failure limit in the system prompt the orchestrator kept trying to fix the orchestrator’s own request using slightly altered query parameters.

You can fix this by setting a maximum retry count of two inside your workflow configuration. If the external tool fails twice force the agent to drop the subtask and elevate the task to a human reviewer instead of letting the agent guess again.

That simple constraint cut my token costs by eighty percent overnight and stopped the silent background crashes entirely. Reliable automation requires boundaries just as much as smart reasoning to stay productive.

What threshold do you use to stop agents from burning resources, on broken tasks and how do you catch these silent failures before they impact production?