The boundary between tasks an agent can reliably solve and tasks it consistently fails, where performance is most uncertain.