What does a production-grade observability stack for AI agents look like? What metrics, logs, and traces are essential? How would you debug a scenario where an agent produces correct outputs 95% of the time but fails unpredictably?

When would you use a "Manager Agent" to delegate tasks to "Worker Agents" rather than having one single agent do everything?

Agents generate a lot of "internal thought" and tool logs. How do you keep the context window from filling up with irrelevant logs during a long multi-step task?

A user asks "What were the sales in 2023?" How do you prompt the LLM to separate the semantic search ("sales") from the metadata filter ("year == 2023")?