Curious how this integrates with LangGraph, specifically if these skills map to subgraphs. I've been managing context bloat by filtering tools manually, so I am wondering if the dynamic loading here actually lowers latency compared to that standard approach.