An LLM Has No Idea It's Killing Your Infrastructure. That's Your Problem to Solve.

I was mid-way through an architecture review for one of our internal Slack agents when something stopped me. We were sketching out the agentic loop (model receives a message, decides on an action, calls a tool, repeat) and I asked a question nobody had a quick answer for: What happens if a user asks it to do something that never terminates? Not a crash. Not a thrown exception. Just a loop that keeps going, one perfectly rational step at a time, until the cost notificati