Google turns NotebookLM into an autonomous research agent

For the past few years, NotebookLM has been that reliable study buddy who happily reads your 50-page PDFs and turns them into tidy summaries or surprisingly fun podcasts. But at the end of the day, it was still just an assistant answering questions strictly within the fence you built.

Google took down those fences with a massive upgrade that transforms NotebookLM into an active research agent, now powered by Gemini 3.5 and Antigravity.

A secure cloud computer in every notebook

The biggest leap under the hood is that Google now equips every notebook with its own secure cloud environment.

In plain English, NotebookLM can now write and execute real code right inside your workspace. Instead of guessing math or hallucinating statistical trends, it spins up scripts to crunch numbers, backed by over 100 curated software skills.

Google claims the new setup beat the prior baseline across top core evaluation benchmarks, pulling off an average win rate of over 65%. In large document analysis, it clocked a 69.9% win rate, and jumped to 78.2% in advanced web research and source discovery. Notably, aggregator Future Tools initially reported the update as running on Gemini 2.5, though Google's official announcement explicitly points to Gemini 3.5.

Crucially, you no longer need a pristine folder of documents to start. You can begin with a rough question, and the agent will use Google Search to pull relevant web sources into your project, keeping every citation attributed and under your control.

Turning notes into finished deliverables

The other major pain point Google tackled is the final export. Previously, getting insights out of an AI tool meant copy-pasting raw text into three different desktop apps.

NotebookLM can now assemble your source data directly into download-ready deliverables. The roster includes Microsoft PowerPoint decks, Microsoft Excel spreadsheets, structured CSV and JSON files, PNG and SVG data charts, and styled PDF reports complete with tables. It even supports image generation via Nano Banana.

Imagine feeding an unformatted mess of sales figures and ad spends into the system. NotebookLM can research market benchmarks on the web, execute code to calculate ROI, and hand you a clean presentation slide and a budget sheet ready for leadership.

Right now, the upgrade is rolling out globally on the web to Google AI Ultra subscribers and Workspace business accounts with AI Expanded Access, with broader availability planned down the line.

Why it matters

We think this is one of the smartest product pivots Google has made with its AI portfolio.

Most AI document tools treat the chat box as the final destination. But anyone doing serious research knows that reading is only twenty percent of the job; the rest is verifying data, running calculations, and building slide decks to convince someone else. By coupling a code-execution sandbox with direct Office-ready exports, Google is turning NotebookLM from a passive summarizer into a capable junior analyst.

If you regularly drown in fragmented industry reports, this is definitely worth testing, provided you have access through an Ultra or business Workspace tier. The real value is not just smarter text generation, but skipping the tedious grunt work of formatting spreadsheets and slides.

For the broader AI landscape, this raises the bar for vertical agents. When frontier reasoning models start performing similarly on text, competitive advantage shifts to workflow integration. Google is playing to its unique strengths here: combining top-tier web search, a secure execution backend, and native document workflows into a single tab.