NotebookLM just became an agent, not a reader

NotebookLM moved from a tool that reads documents you give it to one that finds sources, runs code, and builds deliverables on its own. The shift is driven by three engine-level changes — a new model, a new agent runtime, and a new output layer — that together turn a notebook into an active research workspace rather than a passive Q&A box.

The takeaway: the four upgrades below are not separate features bolted on — they are one capability (an agent that can act) expressed across discovery, analysis, and output.

Gemini 3.5 is the model now answering your questions; Antigravity is Google's agent-first, AI-powered IDE — the runtime that lets NotebookLM write and execute code behind the scenes. Together Google says they deliver more accurate, more reliable answers with better visibility into how the system reaches them.

For everyday use, the practical change is twofold. First, large-document and multi-source reasoning is more dependable, which matters most when a notebook holds dozens of long PDFs. Second, NotebookLM now exposes its thinking steps directly in chat, so you can see the chain of reasoning rather than trusting a black-box answer — useful for catching where a synthesis went sideways before you cite it.

Antigravity also ships a catalog of more than 100 curated software skills — the discrete capabilities (data parsing, charting, format conversion, computation) the agent draws on when a task needs code rather than prose.

Every notebook now includes a secure, isolated cloud computer that NotebookLM uses to write and run code for deeper research and more complex analysis. In plain terms: it can do real data work on your sources instead of only describing them.

Where this earns its keep:

• Messy data, unified. Point it at datasets with conflicting international formats — dates, decimal separators, currencies — and it reconciles them programmatically rather than guessing in prose.
• Math on your actual files. It performs accurate calculations directly against source data, so a financial model or statistical summary is computed, not approximated.
• Charts and reports. It can generate visualizations and assemble them into a finished document in one pass.

The mental model: treat a notebook less like a chatbot and more like a junior analyst who has a sandboxed laptop, your sources, and the ability to show their work.

Yes. You can now open a blank notebook with nothing but a question or a loose idea, and NotebookLM's chat will guide you through building a source repository — using Google Search to find relevant, high-quality material and add it for you. This removes the old precondition that you gather and upload everything before the tool was useful.

Two scenarios Google highlights are exactly the ones researchers hit constantly: finding primary sources in other languages to bring in perspectives outside your default search bubble, and surfacing related works by an author you've just discovered. The chat proposes sources; you decide what makes the cut.

NotebookLM can now create downloadable outputs in more than a dozen formats, and — crucially — you can give detailed instructions up front (e.g. "a PDF report with charts and tables" or "a detailed budget spreadsheet") and then request edits after generation. This is the upgrade most likely to change your daily workflow.

Native PPTX and XLSX are the headline: a notebook can now hand you an editable PowerPoint deck or a working Excel model, not a screenshot of one. Google says more formats are coming. For deck-specific tactics see our Slide Deck guide; for structured extraction, the Data Table guide.

Google's launch examples (a data analyst, a program manager, a gym owner) are deliberately generic. Here are workflows that only became possible on June 8 — written for the people who actually live in NotebookLM.

The unifying move across all four: stop treating NotebookLM as a summarizer and start handing it the whole task — discover, compute, deliver — then verify its work.

A capability this large arrives with real friction — the stuff Google's announcement doesn't lead with, and the reason not to point this at high-stakes work unsupervised yet.

The rollout started June 8, 2026 on the web for Google AI Ultra subscribers and Workspace business customers with AI Ultra access. Google says the capabilities will reach other tiers over time, but has not published a timeline or pricing for free and Plus users, and mobile timing is unconfirmed.

If you don't see the new features yet, that's expected — confirm your plan qualifies, check you're on the web app, and watch our update log, which we revise as access widens. To compare what each plan and competitor offers, see the AI tool comparison and our system limits & benchmarks.