“For the best model (Claude 3.7 Sonnet) it could achieve a 50% success rate on tasks up to 59 minutes vs only 15 minutes if an 80% success rate was required.”
Agent failure fits a constant rate per minute, so success decays like radioactive material. The best model handles hour-long tasks half the time and 15-minute tasks at 80%. Getting to 99% means tasks 70 times shorter. Each extra nine of reliability costs about two more years.