An agent can return a clean answer and still have taken a path nobody would sign off on. Finding out why is the hard part. Once agents plan multi-step work, call tools and adjust mid-task, standard logs and traces capture the output but miss the decisions behind it. So how do you see inside an agent's reasoning when something goes wrong? This session brings together engineers running agents in production to unpack how they instrument agent runs, trace decisions across tool calls and pinpoint where a run first went off course. Through real failures they've had to debug, they'll share the tracing, replay and tooling approaches that actually work, and the gaps that still don't have good answers. Key takeaways: → How to instrument agents to capture decisions, not just outputs → Replay and tracing techniques for isolating where a run went wrong → The core components of an observability stack for production agents