SkillRise: Agents That Grow Skills Instead of Relearning

A new paper called SkillRise (arXiv 2607.26784) goes after one of the more honest weaknesses of today's agents: they relearn everything from scratch. Solve a task, throw away what you figured out, meet a similar task tomorrow, start over. SkillRise proposes an agentic reinforcement-learning setup where the agent accumulates skills and carries them across tasks, so competence compounds instead of resetting.

The bet underneath is worth stating plainly. The difference between a one-shot agent and one that gets better over time isn't a bigger model, it's a durable, transferable skill library and a training loop that knows how to grow it. Cross-task transfer is the hard part everyone hand-waves past, because it's where reuse either happens or doesn't.

This lands in the middle of a real wave. Anthropic pushed skills into Claude, there's been a run of skill-distillation and skill-discovery papers, and the whole field is circling the same intuition: the unit of agent progress might be the skill, not the weight update. SkillRise is one more shot at making that intuition trainable. Paper at arxiv.org/abs/2607.26784.