AREX-Skill: Give Coding Agents Reusable ML Workflows

Overview

AREX-Skill packages practical operating knowledge from machine-learning repositories as skills for coding agents. Rather than just summarizing source code, each skill can describe when a capability applies, what steps to run, how to validate results, and how to respond when an experiment fails.

The project is aimed at researchers and developers using agents for ML engineering or autonomous research. Its library is routed as skill graphs, so an agent can load the relevant instructions incrementally instead of taking in an entire repository’s guidance at once.

Key Features

• Organizes skills around SKILL.md, with optional reference material and executable scripts.
• Routes tasks through repository skill graphs to focus the agent on relevant workflows.
• Includes skills distilled from machine-learning repositories and research sources.
• Provides DisCo, a CLI for installing and using the library, and for creating, verifying, refreshing, and exporting skills.
• Supports integration with coding agents including Codex, Claude Code, and Pi.
• Includes validation steps and recovery guidance intended to make agent-run workflows more reproducible.

Use Cases

• ML engineers can use repository skills to set up, run, and compare inference or training tools under controlled conditions.
• Researchers reproducing a paper can use relevant skills to identify workflows, dependencies, checks, and potential failure points.
• Developers adopting an unfamiliar ML repository can give a coding agent focused operational guidance without manually assembling every procedure.
• Teams building autonomous research workflows can create or refresh skills as upstream repositories change.

Getting Started

DisCo requires Node.js 22.19.0 or later. Install the CLI and library, then start Researcher mode:

npm install -g --ignore-scripts @arex-skill/disco
disco repo-skills install
disco

See the README for provider setup, agent imports, and further instructions.

Alternatives

• skills: AMD Skills focuses on AMD hardware and software workflows, while AREX-Skill packages executable workflows distilled from machine-learning repositories and research.

Considerations

• DisCo needs a configured model provider, such as OpenAI, Anthropic, or Gemini, to run agent workflows.
• Skills may have different licenses. Check each skill’s license metadata before using or redistributing it; the repository-level Apache-2.0 license does not override those terms.
• The README says the complete technical report for its benchmark results is forthcoming, so treat the published results as a summary rather than a full evaluation report.