NotebookLM Is Now an Agentic Research Workstation

NotebookLM is now an agentic research workstation — on June 8, 2026, Google overhauled the tool from a source-grounded document assistant into a system that plans multi-step research, writes and runs code in a sandboxed cloud computer, and exports finished artifacts. It is the clearest sign yet that the "chat with your documents" era is giving way to "hand a researcher a brief."

The shift matters because it changes what a marketing team can credibly delegate. A tool that only chats produces talking points; a tool that runs code, normalizes a dataset, and generates a formatted deck produces drafts. That is a different category of help — and a different category of risk, because a confident draft full of unverified numbers is more dangerous than a hedged answer in a chat window.

This guide covers exactly what shipped, the cloud-computer-plus-skills architecture underneath it, how to read Google's vendor-stated win-rate numbers honestly, which of the 11 output formats matter for agency work, and the governance line every agency principal should draw before pointing this at client material. Everything below is sourced from Google's June 8 announcement and corroborating independent coverage.

On June 8, 2026, Google announced a major NotebookLM overhaul in a post authored by Trond Wuellner (Director of Product Management) and Usama Bin Shafqat (Software Engineer, Google Labs). The headline: NotebookLM now runs on Gemini 3.5 Flash and Antigravity, replacing the prior Gemini 3-based stack, and Google says the upgrade delivers "even more accurate and reliable information along with better visibility into the thinking process."

Two points of precision matter before going further. First, the model is Gemini 3.5 Flash (API ID gemini-3.5-flash), which became generally available June 4, 2026 — not a "Gemini 3.5 Pro," which does not exist, and not Gemini 3.1 Pro. Second, Antigravity in this context is the reasoning and orchestration backbone inside NotebookLM that plans multi-step tasks and invokes skills — distinct from Antigravity the standalone, VS Code-based agentic IDE. Throughout this guide, Antigravity means the orchestration layer, not the coding environment.

The functional changes group into four shifts. You can now start from a vague question and have NotebookLM help build the source repository in chat — suggesting sources via Google Search, locating primary sources in other languages, and surfacing related authors, with the human accepting or rejecting each one. Reasoning steps are now visible in chat, giving a verification path where the model previously operated as a black box. You can provide detailed formatting instructions before generating, and edit outputs after. And underneath all of it, each notebook gets a cloud computer.

The real story is not a model swap — it is that NotebookLM became a lightweight agent platform with pre-provisioned tools. Two pieces make that true. Each notebook is now equipped with a secure cloud computer: an isolated virtual machine that lets the tool write and run code in the background during a research session. That is what moves data work from guessing to computing — calculations, dataset normalization, and statistical analysis happen in code rather than by token prediction.

Layered on top are more than 100 curated software skills. Each skill is a predefined technical blueprint for a class of task — data transformation, statistical analysis, document generation. Crucially, users do not configure or invoke skills manually; the model selects them automatically based on the request. The cloud computer plus the skills library is a software layer, not just a smarter chatbot — and that framing is what puts NotebookLM in the same conversation as dedicated agent platforms.

This is why the "workstation" framing earns its keep. Passive RAG could only reorganize what you uploaded. An agent with a sandbox and a skills library can take raw inputs and produce something new — a normalized dataset, a chart, an analyzed spreadsheet — which is precisely the kind of work that used to sit between a junior analyst and a finished deliverable. For agencies already building agentic competitive intelligence workflows, this collapses several manual steps into one prompt.

Google ran a side-by-side internal evaluation of the new system against the prior one and reported a 65% average win rate across five core evaluation dimensions — a 15-percentage-point margin above parity. Two dimensions stood out: a 69.9% win rate on large-document analysis and a 78.2% win rate on advanced web research and source discovery. Independent outlets including PPC Land, Chrome Unboxed, and 9to5Google all report the same figures.

Read those numbers as what they are. They come from Google's own internal side-by-side evaluation; they have not been independently replicated in academic or third-party benchmarks, and a "win rate versus the prior system" measures relative preference, not absolute accuracy. They are a credible directional signal that the agentic system is preferred to its predecessor on Google's own tasks — not a claim of independently verified superiority over any other vendor's tool.

NotebookLM now supports 11 downloadable output formats, up from a limited prior export set: data visualizations (PNG, SVG); documents (PDF, DOCX, Markdown, plain text); images via Nano Banana (PNG, JPG, GIF); structured data (CSV, JSON); Microsoft Excel (XLSX); and Microsoft PowerPoint (PPTX). The format count is independently confirmed by TechCrunch and Google's own post.

For agencies, the formats that change the workflow are XLSX, PPTX, and PDF. An analyzed Excel from raw spend data, a first-draft client deck from research documents, and a charted PDF report are the artifacts that used to consume junior hours. The table below maps each format to its agency use case. Note one important constraint: the agentic features — code execution, the cloud computer, and full output generation — went to premium tiers at launch, not the free tier.

Nano Banana is Google's image-generation model integrated into NotebookLM for the PNG, JPG, and GIF outputs; its Pro variant was introduced in November 2025 for the Infographics and Slide Deck features. For most agency reporting, the generated-image formats are a nice-to-have — the structured-data and document formats are where the time savings live, because those feed directly into AI-powered content strategy for agencies and existing reporting pipelines.

This is the section no mainstream coverage has built, and the one agency principals actually need. NotebookLM's agentic outputs are drafts, not final deliverables. The tool cannot exercise strategic judgment on whether a competitive insight matters for a specific client's positioning, cannot make qualitative brand claims without sourced data, and cannot replace the senior analyst's interpretation layer. The verification responsibility stays with the practitioner.

The matrix below splits ten common agency tasks into automate, partial, and keep-human, with the reason and the required human review step for each. It is author analysis, drawn from the feature documentation and marketing-specific use-case coverage — treat it as a governance starting point, not a fixed rule set.

The pattern is consistent: NotebookLM is strong at gathering, converting, and drafting, and weak at deciding what matters and standing behind a claim. The tasks safe to fully automate are the deterministic ones — format conversion, structured data collection — where a wrong output is obvious and cheap to catch. Everything that requires a judgment about a specific client sits in the partial or keep-human columns, because that judgment is the service you sell.

The capability story has a compliance counterpart most coverage skips. On the free Standard tier, human reviewers may review queries, uploads, and model responses when a user submits feedback. Free-tier notebooks have no audit trail and no data-residency control, and they can be shared publicly via link — including every source document inside. That combination makes the free tier inappropriate for confidential client materials, PII, or commercially sensitive documents.

Paid Workspace and Enterprise tiers change the calculus. Google contractually commits that documents and outputs in paid tiers are never used to train foundational models, and enterprise tiers integrate with Google Cloud VPC Service Controls and Identity and Access Management. For any agency handling client data under an NDA, the tier is not a budget decision — it is a contractual one. Confirm the current terms before pointing a notebook at a client's confidential material.

NotebookLM cannot be purchased standalone — it is bundled with Google AI subscription plans. At the June 8 launch, the agentic update went to Google AI Ultra subscribers and Workspace business customers with AI Ultra Access and AI Expanded Access. Free, Plus, and Pro tier users were not included at launch, and Google said the update will come to other tiers "over time." The stated reason for the gate is infrastructure cost: running sandboxed cloud-computer instances for millions of simultaneous data-analysis queries carries compute overhead that makes free-tier delivery economically unviable at launch.

On pricing, treat specifics with care. Per-tier limits and monthly prices reported by third-party aggregators do not always match Google's official plans page, so the dependable move is to read current pricing directly from Google's AI plans page before committing a client budget. What is well established: Google AI Ultra became substantially more accessible after a price cut at Google I/O 2026 (May 19, 2026), and the June 8 NotebookLM update was the first major feature wave to land on that tier afterward — Google made the agentic tier cheaper, then immediately gave it powerful new capabilities.

For agencies, the practical sequence is to pilot on the free tier with public data, confirm the agentic outputs are worth the subscription against your real workflows, then move to a paid plan sized to your client-confidentiality needs — not just your usage volume. The governance tier almost always dictates the choice before the feature tier does.

Step back and the trend is clearer than any single feature. NotebookLM is the latest mainstream tool to cross from "answers your questions" to "does the task," and it is doing so inside a source-grounded box rather than the open web. That source-groundedness is the differentiator worth understanding: because NotebookLM works from files you upload plus sources it discovers with your permission, it reduces hallucination risk in citation-dependent work, at the cost of a narrower knowledge surface. That trade is exactly right for competitive intelligence and client reporting, where a sourced answer beats a comprehensive one. We are careful not to call it the first or only agentic research tool — ChatGPT and Claude both ship Deep Research modes; the difference is the grounding discipline, not the novelty.

Projecting forward, the second-order effect is what to watch. Once first-draft decks, analyzed spreadsheets, and competitive teardowns are a prompt away, the scarce skill stops being production and becomes judgment — knowing which insight matters, which claim is defensible, and which draft to throw away. Agencies that treat tools like this as a junior-analyst multiplier, with senior review as the non-negotiable last mile, will compound the time savings. Agencies that ship the drafts unedited will discover that a confident wrong number costs more than the hours it saved. The winning posture is to wire the gathering and drafting into your agentic SEO and research workflows while keeping a human firmly on the interpretation layer.