# Memdoor reports 49 successful coding runs in small multi-agent test

_Published Saturday, October 31, 2026 at 10:18 PM EDT · AI · Latest · Tier 2 — Notable_

![A lightweight coding and automation agent in one CLI binary. Run locally or over SSH, optimize token use, and automate CI and cron with durable workflows. - guregodevo/memdoor — Primary](https://opengraph.githubassets.com/1fe4806e400b248b948ddcfe288d7e0f788b85f4c94aa3b2a915e0926859f2e1/guregodevo/memdoor)

Memdoor reports that its workflow using Qwen3.8-27B solved 49 of 50 programming runs across ten LiveCodeBench Hard problems tested five times each. The workflow averaged 9.8 correct solutions out of ten, compared with 8.0 for Fable 5 alone in results taken from the original researchers' files.

Separate model roles plan, propose algorithms, write code and review test results, exchanging information through files. The 50 runs cost $35.79 and took a median of 40 minutes of model time per run.

The test covered ten problems rather than the original researchers' 100. Training contamination cannot be ruled out, and Fable 5 was not rerun in Memdoor's harness.

## Sources

- [github.com](https://github.com/guregodevo/memdoor/tree/main/examples/workflows/team-solve)

---
Canonical: https://techandbusiness.org/newswire/oFTHRdVQfZUTx2Oy2N8MpH
Published: 2026-11-01T02:18:05.208Z
Story chronology: 2026-10-31T22:46:12.000Z
Retrieved: 2026-11-01T04:33:23.917Z
Publisher: Tech & Business (techandbusiness.org)