AI Agents Fail Silently: How to Fix the Observability Gap

A team's customer support agent built on LangChain denied a valid refund because the model hallucinated a 14-day return policy instead of 30, logging success throughout. Multi-step LLM systems are stateful, non-deterministic, and prone to compounding errors that existing monitoring tools miss. Open-source gateway Ajah addresses this with per-step hallucination scoring, RAG verification, and session step tree visualization.