What limits do you put on your AI agents before shipping?

Many agent demos work well. Then in production, one confused loop calls the same tool 40 times and the bill jumps overnight.

Before I let an agent go live, I add three limits:

1. Max steps: stop after about 10 loops and hand over to a human.

2. A timeout on every tool call, so one slow API cannot hang the whole run.

3. A cost cap per run, checked after every model call.

These are boring, but they turn a scary failure into a clean "I could not finish this" message.

Curious what others do. Which limits or guardrails do you add to your agents? And has a missing one ever surprised you?