I've been building agents for a few months and the failure that scared me wasn't a crash, it was cost. An agent loops, every individual call looks fine, and you find out from the bill. So I built a proxy that sits in front of the agent and stops it. Budget cap that cuts mid-stream, but also: it notices when the agent is sending the identical prompt over and over and breaks the loop before the call goes out, and it drops tool calls you've blocked (like delete_*) before they reach your code. ...