A developer rebuilt a five-model multi-agent coding harness (planner, two ideators, solver, manager plus a repair worker) as a Memdoor workflow using Qwen3.8-27B via OpenRouter, scoring 49 of 50 solves on the first ten LiveCodeBench Hard problems across five passes, versus 8.0/10 for a single Fable 5 call. The 50 runs cost $35.79 (15.3M output tokens), with a median of 40 minutes and $0.39 per run, and agents coordinate only through files on disk with fresh contexts. For API developers, it shows multi-agent file-based orchestration can match much larger single-model results, though training-d…
Read the full story on github.com
What to doIf you want to replicate the setup, install Memdoor, replace task.md with your problem and paired test files, and budget roughly $0.39 per run at these token rates.
- Provider
- Qwen
- Source
- github.com
- Models
- Qwen3.8 27Bqwen3.8-27b