I keep blaming the model for slow agent loops, but the real bottleneck is my architecture making every step wait its turn. I want to rework my agents so steps run in parallel and cheap fast models handle the filler steps, saving the big model for the hard calls. Latency is a design choice, not a fact.