Ever had an AI agent burn through your API budget in a silent loop?

Had a pretty painful moment last week when one of our support agents got stuck calling a dead endpoint over and over. It just kept spinning and burning tokens in the background until I finally noticed the terminal.

We fixed it simply by capping the retry limit to two in our workflow config and routing it to a human if it fails twice. Cut our runaway token costs right away.

How do you usually handle safety limits or failure thresholds for your agents so they don't get stuck on broken tasks?