Do AI-generated skills actually help agents, and when do they fail?

Language agents increasingly learn by extracting and reusing "skills"—procedural recipes distilled from experience. This work maps the entire skill lifecycle: generation, extraction, and use. Across five task domains, the authors find model-generated skills help on average but frequently backfire, with no correlation between model scale and skill utility. They identify which experience patterns and skill properties actually matter, then propose a "meta-skill" that reduces negative transfer and boosts quality across domains.