We build models and tools that help agents choose which model to use, how much reasoning to apply, and when to run locally.