Skill libraries degrade agent performance by 21% due to 'skill shadowing'

A new paper on arXiv (Song & Wei, May 2026) exposes a counterintuitive problem in LLM agent design: giving agents more skills can make them perform worse. The researchers tested agents using skill libraries—collections of task-specific instructions loaded on demand—and found that scaling from a small set of helpful skills to a 202-skill library caused a 21% drop in pass rates. They attribute this degradation to two effects: skill shadowing (the agent picking the wrong skill for a task) and context overhead (degraded reasoning from longer input).

Strikingly, the team’s empirical estimates showed skill shadowing accounts for almost all of the performance loss, while context overhead was negligible and statistically indistinguishable from zero. This asymmetry means the real bottleneck is skill selection, not context length. For developers building modular agent frameworks, this is a critical warning: larger skill libraries need better retrieval and selection mechanisms—or they’ll actively hurt the agent's capabilities.