Autonomous agents are increasingly capable of conducting machine learning (ML) research end-to-end, combining a model backbone with planning, execution, memory, and verification harnesses. However, these systems often lack what researchers call 'operational knowledge'—the practical know-how that turns a theoretical method into a working implementation. This knowledge is typically scattered across repositories and papers, written for human readers and too large to be loaded during a task.
To address this gap, researchers from Hugging Face and VectorSpaceLab introduce DisCo, a skill-powered research agent that distills operational knowledge into compact, verified skills. These skills can be reused across tasks, eliminating the need to rediscover domain-specific insights during each run. DisCo performs distillation in two complementary modes: task-agnostic, which condenses widely used repositories into general-purpose skills, and task-oriented, which generates skills tailored to a specific research problem.
The task-agnostic distillation applied to the open ecosystem produced the AREX-Skill Library, containing over 5,000 verified skills derived from 1,000 popular ML repositories, organized into 20 areas and 178 capability families. This library serves as a rich resource for autonomous research agents.
In experiments, the skill-equipped agent, using a GPT-5.5 backbone and a fixed research harness and execution budget, demonstrated substantial performance improvements over the same agent without skills: 134.3% higher on MLE-bench, 34.4% higher on PaperBench, 9.2% higher on FrontierCS, and 14.0% higher on PassNet. These gains are attributed solely to the addition of distilled operational context.
The research is detailed in the paper 'Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills', available on arXiv and Hugging Face. The team has also released the AREX-Skill library on GitHub and a DisCo CLI on npm for developers to explore.