Working on agent reliability. Mostly interested in the gap between what a run reports and what actually happened.