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

A new benchmark from Alibaba called Co-Design Bench reveals that top-tier AI coding agents fail to fully complete about 74 percent of real-world programming tasks. While these models often perform near perfectly on older, standardized tests, the results suggest a significant performance gap when handling authentic user requests. This disparity indicates that current automated coding tools still require substantial human oversight to address errors and incomplete solutions in practical development environments.