Post

team-solve: a Memdoor workflow reproducing GVS5H

A file-coordinated multi-agent coding workflow reports solving 49 of 50 runs on a ten-problem LiveCodeBench Hard subset, using Qwen3.8-27B through hosted APIs. It recreates the previously covered GVS5H prompts and loop, with separate planner, ideation, solver and manager roles.

The comparison is narrow: five passes on ten problems, potential training contamination, and a single-call Fable 5 baseline from the original authors’ results—not the same multi-agent harness. The reported 50 runs cost $35.79. HN readers discussed whether local quantized models could reproduce the result, and cautioned that heavy quantization may reduce quality.