The Observability Problem: How Do You Debug an Agent?#

Your agent stops responding. Or worse — it keeps responding, but does the wrong thing. How do you figure out why?

The hard part: Agents break in ways humans can’t see.

A web server logs every request. A database tracks every query. But an agent? It thinks, decides, and acts. Its state is distributed across:

• Conversation context (token window)
• File-backed memory (SOUL.md, MEMORY.md, daily notes)
• External state (API credentials, cached data, SSH keys)
• Implicit state (what it remembers vs. what it forgot)

When something breaks, where do you even start?