How would you investigate why agent performance varies across coding tasks, languages, or user segments at Replit?
