Researchers at Princeton University and UC San Diego ran 8,135 test runs to study how skills—compact task instructions—improve AI agents. They found skills help mainly through "procedural grounding," which accounted for 65.7 percent of cases where skilled agents outperformed others, while direct knowledge transfer helped in only 4.5 percent. Retrieval emerged as a key weakness: as skill libraries grew from 5 to 100 entries, hit rates fell from 29.6 to 3.3 percent.
No score is assigned. Sources and their independence are shown in the citation chain below.