Google has just expanded the capabilities of its autonomous research agents with Deep Research and Deep Research Max, two configurations based on Gemini 3.1 Pro. The proposal aims to go beyond summarizing web pages: developers can now build enterprise workflows capable of researching, cross-checking sources, and generating cited reports using public and private data.
What’s the difference compared with a traditional chatbot? Instead of responding only with what it finds in a conversation, the agent can plan an investigation, consult multiple sources, review the evidence, and produce a document ready for professional analysis.
Two agents for different needs
Google is replacing the preview version launched in December with two options designed for different scenarios:
- Deep Research: prioritizes speed, cost, and lower latency. It’s designed for interactive experiences where users expect fast results.
- Deep Research Max: uses more computing power during execution to reason, search, and iteratively refine the report. It’s better suited to asynchronous processes, such as generating a due diligence report overnight for a financial team.
The distinction matters. Not every investigation needs the maximum level of depth. For a query integrated into an application, speed may be decisive. When evaluating a company, reviewing a market, or studying scientific evidence, it may be more valuable to consult additional sources and detect nuances that would normally go unnoticed.
Deep Research Max doesn’t just look for information: it tries to build a more complete, cross-checked synthesis that’s useful for making decisions.
From the web to a company’s private data
One of the most important new features is that Deep Research is no longer limited to open web searches. The agent can work with websites, uploaded files, connected document repositories, and remote servers compatible with Model Context Protocol, known as MCP.
MCP makes it possible to connect the agent to specialized tools and repositories. For example, a company could integrate it with financial information providers, market databases, or internal systems. This allows the agent to consult sources that aren’t publicly available and combine them with information from the web.
Google mentions collaborations with FactSet, S&P Global, and PitchBook to design MCP servers focused on financial data. The goal is for shared customers to incorporate those information ecosystems into automated research workflows.
It’s also possible to completely restrict internet access and have the agent work only with proprietary data. This can be especially useful for organizations that need to keep an investigation within a controlled environment.
Reports with charts and multimodal content
Deep Research includes native chart and infographic generation. Instead of delivering text alone, the agent can represent datasets within the report using HTML or Nano Banana, according to Google’s description.
In addition, developers can provide context in different formats, including:
- PDF files.
- CSV spreadsheets.
- Images.
- Audio.
- Video.
This capability opens the door to broader use cases. A life sciences team could combine research papers, experimental tables, and images. A market analyst could cross-reference PDF reports with tabular data and earnings presentations. The goal is for research not to depend on converting everything into text beforehand.
More control over how the agent conducts research
Autonomy doesn’t mean losing oversight. Google has added several features so developers can observe and adjust the process:
- Collaborative planning: users can review and modify the research plan before execution begins.
- Combined tools: Deep Research can use Google Search, MCP servers,
URL Context, code execution, andFile Searchwithin the same workflow. - Real-time streaming: the application can receive summaries of reasoning steps and display text or images as they’re generated.
- Configurable sources: developers can choose which repositories the agent consults and which ones remain excluded.
For an interactive application, real-time streaming helps show that the system is still working. For an internal process, collaborative planning allows an analyst to correct the scope before consuming time and computing resources.
More complete reports for finance and life sciences
Google says Deep Research Max consults significantly more sources than the December version and pays greater attention to contradictory evidence. It’s also designed to assess the diversity and authority of references, such as SEC regulatory documents and open-access, peer-reviewed scientific papers.
The expected result is a report that doesn’t simply list findings. The tool aims to explain nuances, organize technical information, and turn it into a format that analysts, executives, and other decision-makers can use.
Even so, greater autonomy doesn’t eliminate the need for human review. In regulated fields, an AI-generated report must be verified, especially when it influences investments, diagnoses, regulatory compliance, or strategic decisions. The quality of connected sources and the clarity of instructions will remain decisive.
Available through the Gemini API
Deep Research and Deep Research Max have been available since April 21, 2026, in public preview, within the paid tiers of the Gemini API. Developers can access these features through the Interactions API.
Google also says both options will soon reach startups and enterprises through Google Cloud. The company states that the infrastructure used is related to the research capabilities found in products such as Gemini, NotebookLM, Google Search, and Google Finance.
This evolution is significant because it moves AI agents from isolated answers toward complete workflows. The real challenge won’t just be getting them to research more, but ensuring that their results are traceable, verifiable, and useful within the systems where organizations already make decisions.