AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
The Decoder
Real-world testing reveals that top AI coding agents still require human cleanup on nearly 74 percent of tasks.
Alibaba has introduced a new coding-agent benchmark called Co-Design Bench to test AI models on complex real-world user requests. In evaluations, even the top-performing model, GPT-6 Astra, successfully completed only 26.4 percent of the tasks. This contrasts sharply with older benchmarks where the same models achieved scores exceeding 99 percent.
Why it matters: The performance gap between old benchmarks and real-world testing highlights that current AI coding agents still require substantial human oversight and cleanup for practical applications.

