DisCo turns research and code into reusable AI-agent skills, a post says
The post reports MLE-bench rising from 31.11% to 72.89% with GPT-5.5, Codex and the task-running budget unchanged. It says DisCo’s library contains 5,353 skills distilled from 1,000 machine-learning repositories.
TLDR
AI agents may waste effort rediscovering methods they could remember, a post about “Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills” argues. It describes DisCo as turning GitHub repositories, papers and task research into compact skills that tell agents what to use, when to use it and how to proceed. The post reports MLE-bench rising from 31.11% to 72.89% without changing GPT-5.5, Codex or the task-running budget, and says PaperBench, FrontierCS and PassNet improved too. Its central argument is that better agents may come from teaching them proven ways to work, not endlessly making the underlying model smarter.
Combined views
—
1 Source, first seen ago
