AgentVM Position Page

This is the project page for Position: Digital Agents Require Unified Agent-Native Computers.

Overview

The page presents the Agent-Native Computer position paper and AgentVM project. It highlights the need for unified GUI and text interfaces, shared OS state, compositional environment views, OSWorld results, and case studies.

Key Results

• +9.76 OSWorld points over the GUI-only baseline
• Unified GUI, OS view, text tools, and code execution over shared state
• Case studies from GDPVal and GAIA showing why modality switching matters

Website Structure

agentvm-page/
├── index.html          # Main HTML page
├── style.css           # CSS styling
├── script.js           # Smooth scrolling and BibTeX copy
├── assets/
│   └── images/         # Figures and diagrams
│       ├── position_teaser.png
│       ├── agent_native_computer.png
│       └── position_case_study.png
└── README.md

Sections

1. Position - Explains why GUI-only and API-only environments are insufficient
2. Framework - Introduces Agent-Native Computers and environment views
3. Evidence - Shows OSWorld results from layered interaction surfaces
4. Case Studies - Summarizes workflows requiring both GUI and text actions

Deployment

GitHub Pages

1. Push this repository to GitHub
2. Go to repository Settings > Pages
3. Select "Deploy from a branch" and choose main branch
4. Your site will be available at https://<username>.github.io/<repo-name>/

Local Development

Simply open index.html in a web browser, or use a local server:

# Using Python
python -m http.server 8000

# Using Node.js
npx serve .

Citation

@inproceedings{wu2026agentnativecomputers,
  title={Position: Digital Agents Require Unified Agent-Native Computers},
  author={Wu, Yiran and Liu, Jiale and Zhang, Jieyu and Zhang, Yaolun and Liu, Shilong and Wang, Chi and Wang, Mengdi and Wang, Huazheng and Wu, Qingyun},
  booktitle={International Conference on Machine Learning},
  year={2026}
}

Links

• Project Repository
• AG2 AI

License

This project is part of AG2 AI research.