people are getting the headline wrong. the big win is not “gradient descent for skills”, it’s moving optimization into a layer teams can change fast without touching model weights. but if the evaluator is weak, skillopt will just learn to farm the metric and make agents look smart on paper, not in prod