macOS AI agents rely on pre-written skills for benchmark success

macOS AI agents show high benchmark scores, but a new study reveals most gains come from pre-written skills, not advanced models. A study using MacAgentBench found that AI agents on macOS perform better due to pre-programmed "skills" rather than inherent model intelligence. When these skills are removed, performance significantly drops. The research highlights that while macOS offers a robust automation stack, the effectiveness of current AI agents relies heavily on pre-defined task solutions, raising questions about their adaptability to novel workflows.