Agent Skills Should Go Beyond Text: The Case for Visual Skills

The paper proposes a multimodal skill framework for visual-centric agents that combines textual logic with visual support.

It stores reusable knowledge as static priors, dynamic visual working memory, and interleaved visual skills to capture spatial and state-dependent context.

An automatic pipeline converts task trajectories into these multimodal skills.

The authors report that these skills outperform text-only skills on GUI and other visual-centric benchmarks.