Recording more clicks is not enough to teach reliable desktop work. The central insight is to augment demonstrations with reasoning about state changes and recovery from mistakes, so training teaches how to correct a trajectory as well as imitate it.
Key Findings
- AgentNet includes 22,625 released human tasks across Windows, macOS, and Ubuntu; this makes cross-system transfer a concrete research question.
- Reflective reasoning and mixed training data improve scaling beyond raw state-action imitation, providing an inspectable training recipe.
- The paper reports 45.0% success for the 72B model on OSWorld-Verified at 100 steps. AgentNetBench offers cheaper offline iteration, while online execution still tests end-to-end completion. Paper
Who Should Read It
Great fit if you study desktop agents, demonstration collection, or recovery from GUI mistakes and need released data and models to examine the pipeline. Look elsewhere if you need dependable unattended automation: most benchmark tasks still fail, and small environment changes can alter outcomes. Treat these results as research evidence rather than a deployment guarantee. Project