Agent systems can fail in ways that are difficult to see and even harder to reproduce. A wrong answer may originate several steps earlier from lost context and flawed reasoning to an unsuccessful tool call or an unexpected state change. This panel will examine how engineering teams can make autonomous agents easier to understand and debug in production. Speakers will explore what teams need to trace, how to diagnose failures across multi-step workflows and how observability can turn unpredictable agent behaviour into actionable engineering insight. Key takeaways: → Identify what needs to be captured across agent reasoning, state and tool use. → Trace failures back through complex, multi-step workflows. → Distinguish between model, orchestration, data and integration issues. → Use production traces to improve agent reliability and performance.