Google's Deep Research Max gets MCP superpowers

If you have ever asked an AI to write a deep industry report, you know the frustration: it spits out polished prose in seconds, but the content is often shallow, hallucinated, and completely blind to gated enterprise databases.

Google is tackling that wall head-on. The company just rolled out two new autonomous research agents built on Gemini 3.1 Pro: Deep Research and Deep Research Max. Both are available now in public preview via the paid tiers of the Gemini API Interactions API, with Google Cloud availability coming soon.

Fast lookups vs. overnight deep dives

Instead of forcing one model to handle every workflow, Google split the agent into two distinct tools tailored for different speeds.

Standard Deep Research replaces the preview agent Google dropped in December. It is tuned for low latency and lower cost, making it fit right into customer-facing apps where someone expects an answer without staring at a loading spinner for ten minutes.

Deep Research Max takes the opposite bet. It leans into extended test-time compute, allowing the model to spend serious time cycling through iterative reasoning, multiple search passes, and deep synthesis. Google frames this as an engine for overnight analyst work: you set a background cron job at night, and it hands you an exhaustive due diligence dossier by morning, complete with full citations from sources like SEC filings and peer-reviewed journals.

Developers also get collaborative planning, meaning you can review and guide the agent's research blueprint before it runs wild on a prompt.

Breaking into walled gardens with MCP

Open web search is only half the battle for professional researchers, and Google finally addressed the walled garden problem.

Both agents now support the Model Context Protocol, letting developers plug Deep Research straight into proprietary databases and paid financial feeds. Google is already collaborating with FactSet, S&P Global, and PitchBook on MCP server designs. If you need strict confidentiality, you can even switch off Google Search entirely and force the agent to run exclusively on your own data vaults.

Crucially, the output is no longer just walls of text. The agent now generates native charts and infographics inline using HTML or Nano Banana. You can also ground investigations across multimodal inputs by feeding it combinations of PDFs, spreadsheets, audio, video, and raw images, while watching its intermediate thought summaries stream in real time.

Why it matters

We think this launch pushes autonomous research agents past the demo phase and into serious production pipelines.

For analysts, researchers, and developers, the combination of MCP support and native visual generation solves two massive headaches: getting gated data into the prompt and turning raw numbers into presentation-ready visuals. You are no longer copying and pasting CSV snippets into a chat window; you can point an agent at a proprietary database, let it pull historical data, and get an illustrated briefing back.

Strategically, Google is playing to its infrastructure strengths. By bringing established data heavyweights like FactSet and PitchBook into the MCP loop, Google is building a moat around high-stakes enterprise use cases where accuracy actually matters.

That said, extended test-time compute is not cheap. Running Max on complex research loops will likely rack up substantial API costs, and businesses will need to be deliberate about when a workflow truly warrants a deep background run versus a quick search. Plus, when multiple authoritative sources contradict each other, human verification will still be required before handing these reports to executives.