Debugging and evaluating LLM agents in production

A technical conversation about running AI applications for customers. How do you debug an LLM application or agent that misbehaves in production? Talk through what you trace and log, how you reproduce a bad run, how evaluations catch regressions, and how you would answer a customer who wants to move their tracing and evaluation to a different tool. Expect to be asked about a hard production issue you handled yourself.