Google DeepMind launches Deep Research agents with 160 searches per task

Google DeepMind has unveiled two autonomous research agents—Deep Research and Deep Research Max—now in public preview through the Gemini API. Powered by Gemini 3.1 Pro, the standard Deep Research agent replaces the December preview and targets interactive, user-facing products with fast, cost-efficient performance. The heavier Deep Research Max is designed for asynchronous background workflows like overnight due diligence, using extended test-time compute to iteratively search and refine results, running approximately 160 search queries per task. Benchmark scores hit 93.3% on DeepSearchQA for web research and 85.9% on BrowseComp for hard-fact retrieval, as shared by Google DeepMind’s Philipp Schmid.

Both agents search the open web, user uploads, and connected data sources via Model Context Protocol (MCP) servers. Key partners including FactSet, S&P Global, and PitchBook are building MCP integrations to plug financial data feeds directly into workflows. The agents accept multimodal inputs (PDFs, CSVs, images, audio, video) and support collaborative planning—users review and edit the research plan before execution. Real-time streaming shows reasoning steps as they happen. Google Chief Strategist Neil Hoyne emphasized the distinction: “Regular Deep Research is fast and cheap. Max works through the night so you wake up to something exhaustive.” He cautioned it won’t replace judgment but can give professionals their evenings back. The infrastructure already powers research features in Gemini App, NotebookLM, and Google Search, and is now exposed to developers via the Interactions API and AI Studio.