Masquage de skills (skill shadowing)

Skill shadowing is what happens as an agent's skill library grows: the agent more often opens a skill that sits next to the right one, because several descriptions answer the same request. The right skill is installed, and another one steps in front of it. **Where the term comes from.** Two researchers, Hongwen Song and Song Wei, coin it in a study posted on arXiv on 21 May 2026 and revised on 23 June. They compare the pass rate of an agent given a small set of skills known to help with the pass rate of the same agent given the full library of 202 skills. Performance drops, by up to **21%**. They then split that loss into two effects: - **shadowing**: the agent picks the wrong skill more often as the library expands; - **context overhead**: a longer listing would hurt execution even when the pick is right. The finding is one-sided: shadowing grows with library size and accounts for a significant share of the loss, while context overhead stays small, statistically indistinguishable from zero. The bottleneck is which skill gets picked, not how much room it takes. **Why it happens.** An installed skill permanently leaves the agent only its name and its description. On every request the agent rereads that listing and decides which one to open. With ten descriptions that all mention code review or writing, it settles between neighbours on a handful of words. **What it changes for you.** Counting tokens or trimming descriptions fixes nothing: you need fewer skills that overlap. Before installing one, the right question is not "how heavy is it?" but "which of my skills already answers this request?". The rule of thumb going around, staying under a hundred skills, is a usage ballpark, not a measured threshold: 202 is the size of the library tested, nothing more. **Not to be confused** with the "tool shadowing" of security audits, where a malicious tool poses as a legitimate one. Nobody is attacking here: the library gets in its own way.

Official site

View on Coeurdar