Co-Design Bench Shows Top Coding Agents Fully Complete Only 26.4% of Real Tasks

Alibaba and university researchers released Co-Design Bench, built from 26,314 real coding sessions yielding 273 user change-request tasks scored against 2,331 hidden acceptance criteria. The best model fully completed only 26.4% of tasks.

2026-11-12 ~ 2026-11-12 · 2 related posts

• New Co-Design Bench: best coding agent fully completes only 26.4% of real-world tasks — GaryMarcus · 2026-11-12
• Alibaba's Co-Design Bench: top coding agent fully completes just 26.4% of real user requests — krishnan · 2026-11-12