BREAKING: A new academic paper suggests that agents based on large language models can not only use tools, but also create and refine complex skills on their own.

The proposal, called EvoSkills, reports significant improvements over human curated skills and demonstrates transferability across multiple model families.
EvoSkills achieved a 71.1% success rate on SkillsBench, compared to 53.5% for human curated skills and 30.6% for no skills at all.