[Observability] for Long-Running AI Agents: A Field Guide to What to Log

The first time we ran an AI agent for 30 days straight, we logged nothing useful. The agent worked. Then the agent stopped working. We had no idea why. That's the typical observability story for AI ag