AIThe Decoderabout 1 hour ago
AI Agents Fail on Three of Four Tasks
4 min read
Researchers from Alibaba and Asian universities built Co-Design Bench, revealing that AI coding agents fully complete only 26.4 percent of tasks. The best model, GPT-6 Astra, meets 70 percent of requirements but still requires human cleanup for nearly three out of four jobs. Performance drops significantly when tasks combine multiple change types, and extra compute offers little improvement.
Level
Hype check
Tap to vote and see what everyone thinks.
