Tag
LoopArena benchmarks models as runtime controllers for coding tasks, revealing that even GPT-5.5 only achieves a 24.69% success rate, emphasizing the need for better control mechanisms in agent systems.