AgentNet: 22.6K Human-Annotated Computer-Use Tasks Across Three OSs

AgentNet contains 22.6K human-annotated computer-use trajectories across Windows, macOS, and Ubuntu operating systems. Developed by xlangai and released in early 2026, it serves as a foundation for training vision-language-action (VLA) models for desktop automation.

Use Cases

• Training VLA models to map visual desktop states to specific keyboard and mouse actions
• Benchmarking multi-modal agents on task completion across Windows, macOS, and Ubuntu
• Fine-tuning image-to-text models for UI element recognition and functional description

Strengths

• 22.6K human-annotated tasks
• Cross-platform coverage of Windows, macOS, and Ubuntu
• Supports vision-language-action (VLA) model training

Limitations

• Specific column names and data schemas are not provided in the metadata
• Trajectory data may be sensitive to specific software versions or UI updates

Provenance

Source

xlangai (Arxiv: 2508.09123)

Collection Method

Human-annotated

Freshness

Last updated January 2026.