“Training agentic models requires a clean decoupling of training and inference systems to maintain high throughput and prevent pipeline collapse.”