What's the first thing that breaks when an AI agent reaches production?

Everyone talks about building AI agents. Far fewer talk about what happens after they're deployed.

In your experience, what's usually the first thing to go wrong?

• Infinite loops?

• Tool misuse?

• Hallucinated actions?

• Rising API costs?

• Poor observability?

• Something else entirely?

We're building Clyro around runtime governance, but I'm curious what problems other teams are actually seeing in production. I'd love to hear real stories and lessons learned.