Are AI agents ready for real production workloads?

A lot of the agent demos are impressive, but I’m struggling with reliability once I move past the tutorial stage. Autonomous loops burn tokens, hallucinate tool calls, and end up in useless retries for simple tasks. I’ve had some luck by breaking workflows into smaller, more deterministic steps and adding a human checkpoint, but that defeats the purpose. For those running agents in production: what are the failure modes you actually hit? Do you rely on strict schema validation, timeouts, or something else? Curious if anyone has found a framework or orchestration pattern that genuinely keeps agent costs and error rates under control while handling messy real-world inputs.