More Skills, Worse Agents? Skill Shadowing Degrades Performance When Expanding Skill Libraries
A new research paper from arXiv explores the performance degradation of LLM agents as their skill libraries expand. The study found that as libraries grow, agent performance can drop by as much as 21%, primarily due to "skill shadowing," where agents incorrectly select skills. While context overhead also plays a minor role, the research indicates that the failure in skill selection is the main bottleneck, not the increased context size. AI
IMPACT Identifies a key limitation in scaling LLM agents, suggesting a need for improved skill selection mechanisms.