NotebookLM Upgraded to Gemini 3.5: New Agentic Research Capabilities Explained
Google upgraded NotebookLM with Gemini 3.5, a cloud computer, 100+ skills, and new output formats. Here's what changed and how to use it for research.
What Actually Changed in NotebookLM’s Latest Upgrade
Google’s NotebookLM has been a standout tool since its launch — a research assistant that lets you upload documents, ask questions, and get grounded answers based only on your sources. But the latest upgrade goes well beyond that original design.
With the move to Gemini 2.5 Pro (referred to in Google’s roadmap communications as Gemini 3.5), NotebookLM now supports genuinely agentic research workflows. That means the tool doesn’t just answer questions about files you’ve uploaded — it can now browse the web, use a cloud computer to interact with external tools, apply over 100 specialized skills, and generate research outputs in formats far beyond basic chat.
This post breaks down exactly what changed, what the new features do, and how to actually use them for research work.
Why the Gemini Upgrade Matters
Previous versions of NotebookLM used earlier Gemini models with strong reasoning but limited context and no external action capabilities. The core loop was: upload sources → ask questions → get cited answers. Useful, but passive.
Gemini 2.5 Pro changes the underlying architecture in a few meaningful ways:
- Longer context window: The model can now hold significantly more information in memory at once, which matters when you’re working across dozens of sources or long-form documents.
- Better multi-step reasoning: The model can chain reasoning steps together, which is the foundation of agentic behavior.
- Tool use and function calling: The model can now invoke external tools — not just read text, but take actions based on what it finds.
Together, these capabilities let NotebookLM go from a passive document reader to something closer to an active research assistant.
The Cloud Computer: What It Is and How It Works
The most significant new feature is what Google is calling the cloud computer integration. This is not a virtual machine you interact with directly — it’s a sandboxed compute environment that NotebookLM can use on your behalf.
What the cloud computer enables
When you ask NotebookLM to research something that goes beyond your uploaded sources, it can now:
- Open a browser and visit live web pages
- Search for and read current information
- Interact with web-based tools and forms
- Pull structured data from external sources
- Cross-reference what it finds with your existing sources
Think of it as NotebookLM having hands. Before, it could only reason about what you gave it. Now it can go get things.
How this works in practice
Say you’re researching a market competitor. You’ve uploaded analyst reports and your own internal notes. In previous versions, you were limited to questions about those files. With the cloud computer, you can now ask NotebookLM to check the competitor’s current pricing page, pull their latest press releases, or scan their job postings for signals about product direction — and then synthesize that real-time data with your existing sources.
The cloud computer runs in isolation, which means it doesn’t have access to your local files or system. It operates within defined task scopes set by your prompt or workflow.
Limitations to know
The cloud computer isn’t unlimited. It works within the context of a session and is subject to rate limits. It can also make mistakes when navigating complex or JavaScript-heavy websites. Treat its web-gathered outputs as a starting point, not a final source — especially for anything that requires precision.
100+ Skills: What They Are and How to Use Them
The “100+ skills” refers to a library of pre-built capabilities that NotebookLM can now invoke as part of a research workflow. These aren’t just prompting tricks — they’re structured functions the model can call to perform specific tasks.
Categories of skills available
The skills break down roughly into these areas:
Research and discovery
- Web search with source filtering (news, academic, regional)
- Citation lookup and verification
- Topic clustering across multiple documents
- Gap analysis (identifying what’s missing from your sources)
Writing and synthesis
- Executive summary generation
- Comparative analysis across sources
- Q&A document creation
- Structured outline building
Data and analysis
- Table extraction from unstructured text
- Trend identification across dated sources
- Entity extraction (people, organizations, locations, dates)
- Numerical fact-checking
Output formatting
- Converting findings into different document formats
- Creating structured briefs, memos, and reports
- Generating slide-ready talking points
- Producing annotated bibliographies
How skills get triggered
You don’t need to manually invoke skills by name in most cases. NotebookLM’s upgraded model is designed to recognize when a skill is appropriate based on what you’re asking. If you ask for a “comparison of these three approaches across my sources,” the model will apply the comparative analysis skill automatically.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
For more control, you can be explicit: “Use gap analysis to identify what my sources don’t cover about [topic].” This tends to produce cleaner, more targeted outputs.
New Output Formats
The original NotebookLM was primarily a chat interface. The upgraded version adds structured output types that are actually useful for sharing research findings with others.
Audio Overviews (now improved)
This feature existed before the upgrade but has been substantially improved. NotebookLM can generate a conversational audio summary of your sources — two AI voices walking through the key ideas as if co-hosting a podcast.
With Gemini 2.5 Pro, these overviews are longer, better at handling nuance, and more accurate at representing the balance of evidence across sources. You can also now direct the overview toward specific questions or audiences rather than getting a generic summary.
Study Guides
NotebookLM can now generate full study guides from your source material. These include:
- Key concept definitions
- Practice questions with answers
- Flashcard-style summaries
- Suggested reading order for complex topics
This is particularly useful for anyone trying to get up to speed on an unfamiliar domain quickly.
Briefing Documents
New in this upgrade: the ability to generate structured briefing documents. These follow a more formal format — executive summary, key findings, supporting evidence, open questions, recommended next steps — and are designed to be shared with stakeholders who haven’t read your source material.
Interactive FAQ Documents
NotebookLM can now generate a structured FAQ based on your sources, anticipating the questions someone would ask when reading your research area. Each answer is cited back to the relevant source.
Research Timelines
For sources that include dates and historical context, NotebookLM can extract and organize events into a chronological timeline. This works well for competitive history, regulatory developments, or tracking how a scientific field has evolved.
How to Set Up an Agentic Research Workflow in NotebookLM
The term “agentic” gets used loosely, but here it means NotebookLM can take multiple steps autonomously to complete a research task — rather than waiting for you to manually direct each step.
Step 1: Define your research goal clearly
Agentic workflows work best when you give NotebookLM a clear, scoped objective rather than an open-ended question. Instead of “tell me about electric vehicles,” try: “Summarize the key arguments for and against solid-state batteries from my uploaded sources, identify any gaps, and find two recent web sources that address those gaps.”
That kind of prompt gives the model a goal, a scope, and a follow-up action — all the elements it needs to run an agentic sequence.
Step 2: Organize your source library
Before running complex queries, organize your notebook sources by relevance and recency. NotebookLM’s performance improves when you tag or group sources. Irrelevant sources add noise and can dilute the quality of synthesis.
Remove sources that are clearly outdated or redundant before asking for a comprehensive analysis.
Step 3: Use the cloud computer for live research
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
When you need NotebookLM to supplement your sources with current information, explicitly tell it to use the web. Something like: “Check for any developments on this topic published in the last 90 days and add them to your analysis” will trigger the cloud computer to search and retrieve.
Review what it brings back before including it in any final output. The model will cite its sources, which makes verification straightforward.
Step 4: Choose your output format intentionally
Before generating output, think about who will use it and how. A briefing document for an executive is different from a study guide for a junior team member, which is different from a structured FAQ for a website.
NotebookLM’s new format options are genuinely distinct — don’t default to chat when a structured document format would serve your actual use case better.
Step 5: Iterate and refine
Agentic doesn’t mean hands-off. The best results come from reviewing initial outputs, identifying where reasoning went wrong or sources were misused, and prompting for corrections. Treat the first output as a draft, not a final product.
Where MindStudio Fits for Research Automation
NotebookLM’s new agentic features are impressive, but they’re still contained within Google’s product. If your research workflow needs to connect to other systems — pulling data from a CRM, pushing summaries to Slack, triggering follow-up tasks in project management tools — you hit a wall quickly.
This is where MindStudio comes in. MindStudio is a no-code platform for building AI agents that connect across tools and systems. Where NotebookLM excels at deep research within a document library, MindStudio lets you build the surrounding workflow that feeds into and flows out of that research.
A practical example: you could build a MindStudio agent that monitors your email for incoming research requests, routes them to a structured research prompt using Gemini (one of 200+ models available on the platform), synthesizes the output, and delivers a formatted briefing to a Notion page or Slack channel — all automatically.
MindStudio supports autonomous background agents that run on a schedule, webhook-triggered agents that respond to events in other tools, and direct integrations with Google Workspace, Slack, Notion, Airtable, and hundreds of other business tools.
If you’re doing recurring research work — competitive intelligence, literature reviews, market monitoring — building a MindStudio workflow around an AI research model can save substantial manual time. The average build takes under an hour, no coding required. You can try it free at mindstudio.ai.
FAQ
What model does the upgraded NotebookLM use?
The upgraded NotebookLM runs on Gemini 2.5 Pro, Google’s most capable multimodal reasoning model as of mid-2025. This model brings a larger context window, stronger multi-step reasoning, and the ability to invoke external tools — which enables the agentic research features described in this post.
What is the cloud computer feature in NotebookLM?
The cloud computer is a sandboxed environment that NotebookLM can use to browse the web and interact with external sources on your behalf. When you ask NotebookLM to supplement your uploaded documents with current information, it uses this environment to search and retrieve content, then integrates those findings with your existing sources.
Can NotebookLM now search the internet automatically?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Yes, but with some caveats. NotebookLM can search the web using the cloud computer when given a prompt that requires external information. It doesn’t browse the web by default for every query — you need to either explicitly direct it to do so or set up a workflow that includes web research as a step. Results should be reviewed for accuracy before being used in final outputs.
What are NotebookLM skills and how do I use them?
Skills are pre-built capabilities — web search, entity extraction, gap analysis, comparative synthesis, and more — that the model can invoke as part of a research workflow. In most cases, the model selects the appropriate skill based on your prompt. For more precise control, you can name the skill explicitly in your request (e.g., “run a gap analysis on my sources about X”).
What output formats does NotebookLM support now?
The upgraded NotebookLM supports audio overviews, study guides, briefing documents, interactive FAQ documents, and research timelines — in addition to the standard chat interface. These structured formats are designed to make research outputs more shareable and usable by people who haven’t engaged with the source material directly.
Is NotebookLM suitable for professional or enterprise research?
It’s increasingly capable for professional use, particularly for synthesizing large document libraries and generating structured research outputs. For enterprise workflows that require integration with external systems, compliance controls, or automated routing of research outputs to other tools, you’d typically need to combine NotebookLM with a workflow platform or build a custom agent. MindStudio’s agent-building platform is one option for teams that need that layer of automation.
Key Takeaways
- The Gemini 2.5 Pro upgrade gives NotebookLM a longer context window, stronger reasoning, and the ability to take external actions — not just answer questions about uploaded files.
- The cloud computer feature lets NotebookLM browse the web and retrieve live information, which it then integrates with your existing sources.
- 100+ skills cover the full research stack: discovery, synthesis, data extraction, gap analysis, and structured output generation.
- New output formats — briefing documents, study guides, timelines, audio overviews — make it easier to share research findings with people who haven’t read your source material.
- For teams that need research outputs to flow automatically into other tools or trigger downstream actions, building an AI agent in MindStudio around a Gemini model is a practical next step. Start building free at mindstudio.ai.
