Agent systems that perform well in controlled environments can behave very differently once they meet real users, changing data and unpredictable workflows. Tool calls fail, context degrades, edge cases multiply and seemingly minor errors compound across multi-step tasks. This panel will examine the gap between designing an agent and operating one reliably in production. We'll unpack where agent systems most commonly break, why failures are difficult to reproduce and how engineering teams can build for recovery, observability and control from the outset. Key takeaways: → Identify the failure modes that emerge only under production conditions. → Understand how context, tools and multi-step workflows create compounding errors. → Design agents that can recover safely when actions fail or outputs become unreliable. → Build observability and evaluation into the system before scaling deployment.