Autonomous agents are now capable of conducting end‑to‑end machine‑learning (ML) research. They typically combine a model backbone with modules for planning, execution, memory and verification, yet they still lack domain‑specific operational knowledge—the know‑how that turns a method into a working solution. This knowledge resides in code repositories and papers, written for humans and too large to load during a task.
When such knowledge is distilled into compact, verified skills, it can be reused across tasks instead of being rediscovered each time. We introduce DisCo, a skill‑powered research agent that creates skills and invokes them during research. The distillation runs in two complementary modes: a task‑agnostic mode that condenses widely used repositories into reusable skills, and a task‑oriented mode that produces the skills required by a concrete task.
Applying the task‑agnostic mode across the open ecosystem yields the AREX‑Skill Library, containing over 5,000 verified skills distilled from 1,000 popular ML repositories, organized into 20 areas and 178 capability families. With a GPT‑5.5 backbone, a fixed research harness and execution budget, the skill‑equipped agent scores 134.3% higher on MLE‑bench, 34.4% higher on PaperBench, 9.2% higher on FrontierCS and 14.0% higher on PassNet compared to the same agent without skills. All gains stem from adding distilled operational context under the same setup.
Review: DisCo demonstrates that systematic extraction and reuse of operational knowledge can markedly boost the efficiency and output quality of autonomous research agents, providing a scalable skill foundation for the AI‑4‑AI ecosystem.