Computer Use: Successful OSWorld Agent Trajectories

Comprising 160 successful computer-use agent trajectories comprising 1,378 total steps, generated by markov-ai using the Gemini 3 Flash Preview model on OSWorld tasks. Updated in February 2026, the collection focuses on fully completed tasks (score 1.0) across software domains including Chrome and GIMP.

Use Cases

• Training GUI agents using the linearized accessibility-tree grounding features
• Analyzing step-by-step interaction patterns for Chrome-based web browsing tasks
• Fine-tuning multimodal models for image-to-text action mapping in GIMP

Strengths

• 100% success rate across all 160 trajectories ensuring high-quality demonstrations
• Includes 1,378 discrete interaction steps with linearized accessibility-tree grounding
• Covers specific software domains including 21 Chrome tasks and 15 GIMP tasks

Limitations

• Small sample size of 160 trajectories limits statistical diversity
• Selection bias present as the data only contains successful outcomes (score = 1.0)
• Agent-specific bias from the Gemini 3 Flash Preview model used for collection

Provenance

Source

markov-ai via OSWorld benchmark tasks

Collection Method

Synthetic agent generation using Gemini 3 Flash Preview

Freshness

Last updated February 13, 2026.

Geography

United States