aparnadhinak shared tips on how to build better agents with aakashgupta in this video: youtube.com/watch… Once you have traces, create one targeted eval.
Don't start with a giant eval framework. Pick one behavior that matters and evaluate it on real traces