Building an agent that demos well is easy; building one that holds up in production is a different sport. Drawing on lessons from shipping the Scifinder agent, this talk covers the gotchas that show up across any agent application: context rot and message trimming, when prompt caching helps versus hurts, designing observability from day one, and detecting and handling the hallucinations. Whatever type of agent your building, you'll leave with concrete things to instrument and traps to avoid.
Video Stream
