The Two Layer Observability Mistake I See in Most Multi Agent Systems
Agent systems need two layers of observability, not one. Layer one is infrastructure (did the job run?). Layer two is semantic (was the output correct?). Most teams ship with only layer one and discover the gap the first time they actually read what their agent produced. The semantic layer is its own tooling category now: Langfuse, LangSmith, Helicone, Braintrust, Arize Phoenix. Pick one early.