AI 編程代理在 74% 的任務中需要人類清理

a team from alibaba and several asian universities developed co-design bench, a benchmark for coding agents that uses real user change requests. they tested ten models, including gpt-6 astra, claude opus 5, and qwen3.8-max. gpt-6 astra, the top performer, meets 70% of task requirements on average but fully completes only 26.4% of tasks. the models struggle with tasks that require multiple types of changes, such as layout, design, interaction, and content. the researchers found that the performance of the models drops significantly as the number of change types increases. the harness, or environment, in which the models are run also affects the results, highlighting the importance of testing both the model and the working environment together.