Princeton's 8,000-agent study found skills work by process not facts

researchers from princeton and uc san diego analyzed over 8,000 test runs to determine why ai 'skills'—compact instruction sets for specific tasks—improve agent performance. results show that skills function best as a procedural playbook, steadying agent actions and reducing execution errors in 65.7 percent of cases, rather than acting as a knowledge base. the study identified a critical bottleneck in scaling: when a skill library expands from 5 to 100 entries, the agent's ability to retrieve the correct instruction plummets from 29.6 percent to 3.3 percent. additionally, agents mechanically applied skills inappropriately in 10 percent of cases. the authors conclude that future agents require better lifecycle management for creating, retrieving, and applying skills rather than simply storing more.