In a technical release published to the guregodevo/memdoor repository on GitHub alongside version 1.773.0, developers detailed a concrete implementation of the GVS5H multi-agent programming architecture using standard Memdoor workflow configuration files. The release provides complete public reproducibility for five small language models collaborating to solve competitive programming tasks without custom Python orchestrator code.
The architecture coordinates five distinct agent roles operating over a shared directory structure: a planner, two parallel ideators, an implementation solver, a review manager, and a repair worker. Each agent executes in a clean context window capped at 128,000 tokens, communicating exclusively through discrete disk files rather than an accumulating multi-turn conversation thread. The workflow engine initiates the sequence by providing task.md and sample test cases to the planner, which produces plan.md and notes.md.
Once planning concludes, two ideator agents execute concurrently against the plan. One ideator is instructed to propose the most direct provable algorithmic approach, while the second explores an alternative algorithmic paradigm. The solver agent reads both proposals, commits to a chosen strategy recorded in notes.md, and writes the initial program to solution.py. A deterministic bash command then executes solution.py against paired input and output samples, piping raw execution stdout and error codes into test-output.log without terminating the pipeline.
Evaluation results conducted on the first ten problems of the LiveCodeBench Hard set across five independent passes demonstrated remarkable consistency. The Memdoor workflow achieved an average of 9.8 problems passed out of 10 across hidden unit tests, matching the original paper's harness score of 9.6 and outscoring single-call Fable 5, which registered an 8.0 mean score across the identical problem subset. Individual pass scores for the workflow registered 10, 10, 10, 10, and 9 out of 10.
Financial and computational overhead remained modest throughout the evaluation. Across the full 50 benchmark runs hosted on OpenRouter, the median execution duration was 40 minutes per problem, with a median inference expenditure of $0.39 per run. The cumulative cost for all fifty benchmark runs amounted to $35.79. The authors published all fifty raw input and output logs, intermediate artifacts, and grader outputs in bench/results/parity-5pass/.
The release notes highlighted key architectural limitations and observed edge cases. In one run, the manager agent hallucinated an invalid counterexample, prompting the repair worker to overwrite a previously functioning program with a flawed algorithm. In two other runs, correct implementations were temporarily overwritten before subsequent repair passes recovered the solution. Furthermore, the underlying competition problems date from spring 2025, leaving open the possibility of pretraining data contamination for both Qwen3.8-27B and Fable 5.