DisCo Agent Distills GitHub Repos into AI Skills

Autonomous agents are increasingly capable of conducting end-to-end machine learning (ML) research. While these agents typically combine a model backbone with components for planning, execution, memory, and verification, they often lack domain-specific operational knowledge—the practical know-how required to implement a method effectively. This crucial information is often buried in human-readable formats within repositories and papers, making it too extensive for direct use during a task. The DisCo framework addresses this gap by distilling this operational knowledge into compact, verified "skills" that can be reused across multiple tasks, eliminating the need for constant rediscovery. DisCo operates in two modes: task-agnostic distillation, which condenses widely used repositories into general reusable skills, and task-oriented distillation, which generates skills specific to a concrete research task. Applying the task-agnostic approach across the open ecosystem has resulted in the AREX-Skill Library, comprising over 5,000 verified skills derived from 1,000 popular ML repositories, categorized into 20 areas and 178 capability families. When equipped with these skills, a research agent using a GPT-5.5 backbone and a fixed research harness and execution budget showed remarkable performance improvements. It scored 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. These gains underscore the significant advantage of integrating distilled operational context into AI research agents.