Minimal open-source implementation mirrors Meta agentic coding pipeline

Hot take: this is useful mostly as a reproducibility seed, not as a polished benchmark suite yet.

• –It targets the paper’s core idea directly: compact rollout summaries plus sequential refinement and recursive voting.
• –The choice to focus on Gemini 3.1 Pro makes it practical for experimentation, but also narrows portability.
• –SWE-bench support gives it immediate relevance for agentic coding evaluation.
• –Because it is described as minimal, expect gaps versus the paper’s full experimental matrix, especially model coverage and additional benchmarks.