Preprint reports gains from separating AI agent skills by function

Researchers report that separating AI agent skills into four functional components improved performance across six benchmarks, three agent models and two execution harnesses. Their preprint describes Cognitive Differentiation in Skill Evolution, which organizes skills into Meta, Experience, Tool and Script components. The method learns from successful and failed task histories, then updates and validates each component according to its role. It achieved the highest scores in all 36 tested settings, averaging gains of 7.6 percentage points over the Base agent and 2.6 points over the strongest competing method in each setting. These are benchmark results reported in a preprint.