Until benchmarks punish agents for writing dirty code, it will continue. Right now “completing the task” is more important than “don’t damage the codebase”.