How do you stop an AI agent from running in a loop on simple tasks?

Month one of our customer support agents got stuck trying to refund a canceled order three hundred times. The reason was a status code error from the API. The agent kept retrying every two seconds. It did this repeatedly until our server usage spiked and our cloud bill jumped significantly for that day.

Most builders set a token limit. They assume the model will handle errors gracefully. In practice large language models often misinterpret validation failures as glitches. They then attempt the same broken action over and over again.

To fix this we introduced a strict failure counter directly inside the tool-call execution wrapper. If any specific tool returns the error message three times in a row the agent is forced to pause execution. It then routes the ticket to a queue.

We also added a stop condition, on state changes. This ensures that financial actions can never execute twice within the session. Since making that change we have had zero loops. Our average response accuracy has improved.

What guardrails do you use to keep your agents from looping on bad data?