Google released two new autonomous research agents in the Gemini API—Deep Research and Deep Research Max—both built on Gemini 3.1 Pro. The split mirrors the way teams actually use research: one tuned for speed and integration into interactive products, the other tuned for depth and the kind of nightly cron job that lands an exhaustive due-diligence report on an analyst's desk by morning.
Crucially, both ship with Model Context Protocol (MCP) support, native chart and infographic generation, file uploads, and a collaborative planning step—turning what was a "sophisticated summarization engine" in December's preview into something Google explicitly positions as the foundation for enterprise workflows in finance, life sciences, and market research.
Two Agents, One Underlying Capability
- Deep Research — optimized for speed and efficiency; replaces the December preview with significantly lower latency and cost at higher quality. The agent for interactive surfaces where users are waiting on a response.
- Deep Research Max — optimized for maximum comprehensiveness; uses extended test-time compute to iteratively reason, search, and refine the final report. The agent for asynchronous, background workflows.
Google says Deep Research Max represents a leap on industry-standard retrieval and reasoning benchmarks compared to the December release, and consults significantly more sources while identifying nuances the older agent missed.
MCP Support Is the Headline
The most consequential addition is full MCP support across arbitrary remote servers. Deep Research can now seamlessly connect to custom data and specialized professional data streams—financial providers, market data, internal knowledge bases—while still ranging across the open web, file uploads, and connected file stores. Or any subset of those. Or none of them, with web access entirely off, in which case the agent searches exclusively over your custom data.
Google is explicit about who this is aimed at: it's actively collaborating with FactSet, S&P Global, and PitchBook on their MCP server designs so shared customers can plug those data universes straight into Deep Research workflows. That's the practical bar for enterprise adoption—not "it's smart," but "it can read our actual data."
Native Charts and Infographics
For the first time in the Gemini API, Deep Research natively generates high-quality charts and infographics in-line with HTML or Nano Banana—dynamically visualizing complex data sets to enrich the analytical reports. The output is presentation-ready rather than something an analyst has to rebuild in another tool before sharing with stakeholders.
Control Over the Research Process
Google added the kind of agent controls that turn a one-shot tool into a workflow primitive:
- Collaborative planning — review, guide, and refine the agent's research plan before it begins execution; granular control over scope.
- Extended tooling — combine Google Search, remote MCP servers, URL Context, Code Execution, and File Search simultaneously.
- Multimodal grounding — provide PDFs, CSVs, images, audio, and video as input to ground the agent's research in custom context.
- Real-time streaming — track intermediate reasoning steps with live thought summaries and receive text and image outputs as they're generated.
Same Infrastructure as the Consumer Products
Building with Deep Research means tapping into the same autonomous research infrastructure that powers Gemini App, NotebookLM, Google Search, and Google Finance. That's a distribution and battle-testing argument: developers aren't getting a research-preview SDK, they're getting the API behind products Google has been iterating on with millions of users.
Availability
Deep Research and Deep Research Max are available starting today in public preview via paid tiers in the Gemini API, exposed through the Interactions API. Both agents will also soon be available to startups and enterprises in Google Cloud.
Why It Matters for Web Developers
Two reads. First, MCP is now table stakes—Google joining OpenAI and Anthropic in shipping MCP-native agent products means the protocol has crossed from "interesting standard" into "expected integration surface." Any tool you build that exposes data should ship an MCP server. Second, the speed-vs-depth split between Deep Research and Deep Research Max is a useful template: most agent products will need both modes within the next year, and the ones that ship both behind a single primitive—rather than forcing developers to pick at API time—will have a meaningful UX advantage. For developers building research-flavored products on top of Gemini, Deep Research replaces what used to require weeks of agent scaffolding with a single API call that returns fully-cited, chart-rich reports.