Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard

A multi-agent system called Memdoor using five Qwen3.8-27B models achieved a mean score of 9.8 out of 10 on LiveCodeBench Hard problems, outperforming Claude Fable 5's score of 8.0 and matching the original authors' harness at 9.6. The system employs five specialized roles (planner, ideators, solver, and manager) that communicate only through files on disk, with success determined by shell script validation rather than model judgment. The 50 total runs across ten problems with five passes each cost $35.79 and processed 15.3 million output tokens.