Agent Skills: Where They Help and Where They Hit a Wall

Researchers from Princeton and UC San Diego ran more than 8,000 test runs to determine why AI agents need "skills" — pre-built instruction sets for specific tasks. A structured action sequence inside a skill helps in 65.7% of cases, while direct knowledge transfer yields results just 4.5% of the time.

The key limitation is scale: as the library grows from 5 to 100 skills, retrieval accuracy drops from 29.6% to 3.3%, and in one out of every ten runs the agent mechanically applies the wrong instruction. Skills work as rigid action playbooks, not knowledge bases — which is exactly why scaling such libraries keeps hitting a wall.