A team of five Qwen3.8-27B models, orchestrated via a Memdoor workflow, achieved a mean score of 9.8 out of 10 on the first ten LiveCodeBench Hard problems. This outperformed the authors' original harness (9.6) and Fable 5 alone (8.0). The system uses five distinct roles—planner, two ideators, solver, manager, and repair worker—each operating from a fresh context and communicating only through files on disk.
Success is determined by a shell script running sample tests, not by the model itself. Rebuilt using the authors' exact prompts and loop, the Memdoor version replicated their method with high fidelity, as verified by byte-for-byte request comparisons. Over five passes, 49 of 50 runs solved their problems, costing $35.79 in total.
The setup ran entirely on Open Router, utilizing hosts like Modal, Core Weave, and Wafer, with FP8 precision. Despite identical configurations, minor differences in host and timing led to slight performance variations. The method shows promise but highlights persistent challenges in multi-agent coordination and test coverage.
Source: Hacker News · Summarized by HeadlinesBriefing