Subtask-level skills transfer across tasks; task-level skills make agents worse

arXiv·high signalXBlueskyLinkedInCopy link

An August 20 arXiv study of skill induction in LLM agents finds task-level skill extraction often degrades performance below baseline, while subtask-level skills raise it, and text-format skills transfer better than code-format ones. The authors propose a skill utility score combining specificity (how closely a skill matches task requirements) and abstractness (how broadly it applies), which predicts transfer success without running the task. That makes it a cheap offline check on a growing skill library, which matters directly if you are accumulating agent skills or SKILL.md files run over run.