how do u evaluate a multi agent pipeline , score every step or only the final answer? curious what actually catches bugs for people