Hybrid Reasoning Agents: 35–55% fewer tokens and up to 2.6× faster inference