OpenAI's kill switch can take 35 seconds because authorization is cached
Control Room moves stop authority into the inference layer, while Hex's cost work shows why production agents need local evals and harness telemetry, not one global model choice.