Google has announced a major upgrade to its Deep Research agent, making it available to developers through a new Interactions API, with consumer-facing rollouts coming soon to Google Search, NotebookLM, and the Gemini app. The upgraded agent runs on Gemini 3 Pro and is designed for iterative, multi-step research. It plans searches, reads results, identifies information gaps, and continues querying until it arrives at a well-supported answer. Google says this enables more rigorous and autonomous web research than previous approaches. A new Interactions API unifies access to Gemini models and agents, offering native support for MCP connections to external tools and data sources. Alongside the release, Google open-sourced DeepSearchQA, a 900-task benchmark created to evaluate multi-step web research more rigorously than existing tests. Google reports state-of-the-art results on Humanity’s Last Exam (46.4%) and the DeepSearchQA benchmark (66.1%), outperforming the base Gemini 3 Pro model. The update strengthens Google’s position in the increasingly competitive deep research agent space.