What actually stalls agentic AI projects is that even when the model knows what to do, the system can't reliably do it.