Agent skills work by procedural anchoring (65.7%), not knowledge injection (4.5%) — and retrieval precision collapses from 29.6% to 3.3% as the skill pool grows

arXiv·high signalXBlueskyLinkedInCopy link

'Demystifying Agent Skills: Why They Work—Until They Don't' (arXiv:2608.14036, Aug 14, 2026, Zhiyuan Jiang et al.) normalized 8,135 trial records and open-coded 238 labels into a taxonomy of 12 skill-use modes. Skills stabilize execution paths rather than supply new facts: procedural anchoring accounts for 65.7% of the benefit versus 4.5% for explicit knowledge injection, and skills beat Workflow Memory by 6.06 points in matched comparisons. The failure mode is retrieval — expanding the skill pool from 5 to 100 items drops actual-use precision from 29.6% to 3.3%, which is a direct warning to anyone stockpiling hundreds of SKILL.md files.