Alibaba's Co-Design Bench finds AI coding agents still need human cleanup on most real-world tasks
Alibaba's new coding-agent benchmark tests AI agents on real user requests, finding that even the top model required human cleanup on nearly three out of four tasks.