Efficient Agent Training for Computer Use

Scaling up high-quality trajectory data has long been a critical bottleneckfor developing human-like computer use agents. We introduce PC Agent-E, anefficient agent training framework that significantly reduces reliance onlarge-scale human demonstrations. Starting with just 312 human-annotatedcomputer use trajectories, we further improved data quality by synthesizingdiverse action decisions with Claude 3.7 Sonnet. Trained on these enrichedtrajectories, our PC Agent-E model achieved a remarkable 141% relativeimprovement, surpassing the strong Claude 3.7 Sonnet with extended thinking onWindowsAgentArena-V2, an improved benchmark we also released. Furthermore, PCAgent-E demonstrates strong generalizability to different operating systems onOSWorld. Our findings suggest that strong computer use capabilities can bestimulated from a small amount of high-quality trajectory data.