LangFuse for LLM Agents: How to Eat the Testing Cactus Without Getting Pricked

How to test autonomous multi-agent systems if their quality is not deterministic and errors can hide deep within the chain of agents?