Very useful insight in this paper.

AI agents may be wasting huge amounts of effort rediscovering things they could simply remember as reusable skills.

Better agents may come from teaching them proven ways to work, not endlessly making the underlying model smarter.

The current problem is: a model may know what a tool or method does, but still waste time figuring out which one to use, how to set it up, and what to do when it fails.

DisCo turns GitHub repositories, papers, and task research into compact skills that tell the agent what to use, when to use it, and how to proceed.

With GPT-5.5, Codex, and the task-running budget unchanged, MLE-bench jumped from 31.11% to 72.89%. PaperBench, FrontierCS, and PassNet improved too.

Its library contains 5,353 skills distilled from 1,000 ML repositories.