What breaks when your agent runs unsupervised?

I'm looking into the real problems teams run into when they have AI agents in production. I mean agents that use tools, APIs, MCP servers, or databases. I want to hear what has actually gone wrong, not just what could go wrong. If your response is a good fit, I'll read it myself and send you a short note with what I think may be causing it and how I would start approaching a fix.