AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
Alibaba’s new coding-agent benchmark, Co-Design Bench, tests AI agents on real user requests that already tripped them up in practice. Even GPT-6 Astra,…
