An incredible new feature has been added to Google Gemini that will allow Google Gemini to create a mini-clone of itself and send it out on the web to find information for you based on your prompts. Then, when it returns, it can create a complex and detailed report with links to the information it finds.

This new feature will be available to Gemini's senior subscribers as a drop-down in a new model called “Gemini with deep research” In addition to the herd of AI agents, a new document feature similar to ChatGPT canvas will be announced simultaneously, allowing reports will be viewable and editable.

This is the first step in entering a whole new era of AI agents, where control is handed over to artificial intelligence. This will allow artificial intelligence to efficiently perform a series of tasks on our behalf; Deep Research is independent of the Google Jarvis agent that is rumored to be included in Chrome next year.

Aarush Selvan, a senior product manager on the Gemini team, spoke to me prior to the announcement and said it can infer the data it finds to generate reports. He explained, “[It's] a hungry analyst, but not a graduate PhD assistant.”

Examples of potential use cases for Gemini's deep research include planning a home renovation project, preparing a report prior to purchasing a new car, or even something more complex such as analyzing alternative business proposals prior to launching a new company.

“We really wanted to create a really useful feature for those thorny research topics where you have to open a lot of tabs and piece together information and don't know where to start,” Selvan told Tom's Guide.

“When [Gemini Deep Research] browses the web, it infers what it finds to figure out what it is going to look for next.

When Gemini Deep Research is done with its report, it breaks it down into headings and sections and presents it to you with citations and references. You can link to the sources from which Gemini Deep Research gathered the information, and you can also ask follow-up questions to improve the report.

The Deep Research demo allowed you to gather specific information on graduate school grants. Follow-up could refine the data already collected, add additional information, including available funds from each grant, and present it in a table.

Google Gemini Deep Research is not a new independent model, but a “1.5 pro model system. You give a prompt, and the first Gemini 1.5 Pro will ponder that request and formulate a plan of action.

It then creates its own version to implement that plan, including absorbing the content of potentially hundreds of websites and public sources.

It will not work with research papers or websites behind a paywall, or websites that require a sign-in to access, but it can give you your own data source.

This is not a quick process. It can take several minutes to complete the search and provide a final report (which can be exported to Docs).

During these few minutes, Gemini refines its analysis, browses the web just as a human would search, finds interesting parts, and runs new searches based on what it learns. Gemini can repeat this over and over until it is confident that it has enough information to produce a comprehensive report based on your instructions.

“I would describe this as an agent,” Selvan said, adding that “Gemini 1.5 Pro sends out a version of itself to perform these tasks.”

Google Gemini Deep Research is already an impressive new feature on its own. It introduces the concept of agents to mainstream chatbots, even if they are only available to those who have paid for a Google One account. Add in other features such as the inference feature introduced in Gemini 1.5 Pro and the document interface for reports, and you get a hint of what Google is trying to do.

Unlike OpenAI, which makes a big deal out of every new feature, Google is quietly cooking under the surface to add new elements to its existing products to make them easier to use.

Don't be surprised if deep research is added to the already impressive Google Gemini Gems. Also, the addition of a podcasting feature in NotebookLM would be a powerful creative tool.

The key here is the emergence of agents: by 2025, agents from every imaginable AI company will dominate AI, performing every conceivable task: writing code, booking airline tickets, browsing the web, etc.