A new Gemini capability: agentic video understanding. Process long-form video with more accuracy while using fewer tokens, by letting the model decide what to watch.