AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test

Alibaba's new Co-Design Bench, which tests AI coding agents on real user requests that had already tripped them up in practice, found that even the top performer, GPT-6 Astra, fully solves only 26.4 percent of tasks. The same models score above 99 percent on older benchmarks, meaning human cleanup is still needed on nearly three out of four real-world tasks.