Agent systems rarely break in the way the architecture diagram suggested they would. On paper, the workflow is clean: context comes in, the agent reasons, tools are called, actions happen, and the system improves. In production, inputs are messier, tool calls fail, context drifts, users behave unpredictably, and small errors compound across multi-step workflows. This closing panel looks at what actually breaks once agent systems move into real environments. From evaluation gaps and observability blind spots to orchestration failures, latency, cost, and recovery, the discussion will unpack why production exposes problems that design rarely anticipates and what teams are learning about building agent systems that can hold up under real usage. Key takeaways: → Where agent systems most often break once they move beyond controlled testing → Why evaluation, observability, and debugging become harder in multi-step workflows → What teams are doing to build more reliable, recoverable, production-ready agent systems