inference speed

Three model calls, finished before you blink

Qwen 3.8 27B at 1500 tokens/s on Cerebras is an orchestration budget, not a quality upgrade - spend it on validation and decomposition, not more agents.