If you run sub-par agents on large codebase that's kinda what they do. Every task ends with duplication, added flags bifurcating the code and everything else described in that comment haha