Reliability issues occuring in the hypergrowth stage of a company's lifetime has always been pretty common, way before agents were involved in the scaling. And that was for products for which the computational load for a single client was negligible. For inference though, even a single client's load is pretty terrifying.