Building an agent that works in a controlled demo is one thing; keeping it reliable in production is another. For those who have actually deployed AI agents, what has caused the most problems? Tool/API reliability? Context management? Memory? Authentication and permissions? Evaluation? Hallucinations? Cost and latency? Observability? Human-in-the-loop workflows? Integration with legacy systems? I'm especially interested in what changed between the prototype and production. What problem did...