Explicit skills tie context on 500 agent tasks

ContinualSkillBench tested agent learning across five domains with 100 interconnected subtasks each, and found explicit skill maintenance performed only comparably to in-context learning on average. The researchers report that reusable skills help selectively, especially when tasks demand repeatable procedures or exact outputs, while weaker models accumulate larger, more fragmented task-specific libraries. Builders extending the harness lessons in Octobench’s five-task performance swing should benchmark retained skills against an equal-context baseline before paying the complexity and token cost of a permanent skill store.