Google has completed the transformation of its research service NotebookLM—used by over 30 million people and 600 thousand organizations—into an autonomous agent called Gemini Notebook. The reason to take a fresh look at the platform is a new bundle of ten features that automation practitioner Julian Goldie tested on real work tasks. Beneath the rebrand and new export buttons, it is easy to miss the main engineering detail of the update: every notebook now gets its own isolated cloud computer capable of writing and running code.
Until now, these notebooks worked like a neat research assistant: you uploaded a stack of files, asked a question, and got a summary with footnotes pointing to the right pages. But what changes in day-to-day work with reports and spreadsheets when a knowledge base stops merely hunting for quotes and starts crunching data with scripts and assembling finished documents on its own?
Isolated Python Sandbox: Computing with Code Instead of Guessing Numbers
Classic Retrieval-Augmented Generation handles meaning very well, yet stumbles on basic bookkeeping. A text model is built like a well-read humanities scholar with a sharp memory: it instantly finds the right clause in a contract, but when asked to sum hundreds of rows from a table, it predicts digits by probability—like multiplying large numbers in your head by ear. That is where hallucinations come from, when the text sounds convincing while the final total is off from reality.
What caught my eye in the updated Gemini Notebook architecture is how directly Google engineers closed this vulnerability. A notebook holds up to 50 sources of up to 500,000 words each inside a 1 million-token context window. As soon as your question involves calculations, metric trends, or a messy export, however, the neural network stops guessing with words and launches a compute loop inside a secure container:
- End-to-end indexing. The agent links text excerpts from documents with numeric grids from attached CSV and Excel tables.
- Script generation. The system writes Python code itself using the Pandas and Matplotlib analytics libraries.
- Self-check loop. If the script trips on a broken cell, a crooked date format, or a shifted column, the agent reads the error, rewrites the data-cleaning step, and reruns the calculation until it reaches a mathematically exact answer.
Take a real-world task for a financial analyst: on the desk sit three unstructured quarterly company PDF reports and a raw transaction export transactions.csv. You need to reconcile operating expenses and find what caused a spike in logistics costs. In real corporate spreadsheets, column names in CSV rarely match the wording in PDFs, and dates are formatted inconsistently. Instead of wrestling with formulas by hand, you state the task in plain text in the chat, and in the cloud sandbox the agent builds and executes a calculation like this:
import pandas as pd
import matplotlib.pyplot as plt
# Read transactions and normalize mixed date formats
df = pd.read_csv("transactions.csv")
df["date"] = pd.to_datetime(df["raw_date"], format="mixed", dayfirst=True)
df["quarter"] = df["date"].dt.to_period("Q")
# Map item codes to categories from PDF reports
category_map = {"ship_cost": "Logistics", "delivery_fee": "Logistics", "ops_misc": "Operations"}
df["category_clean"] = df["dept_col"].replace(category_map)
# Compute exact quarterly expense trends and save chart
summary = df.groupby(["quarter", "category_clean"])["amount"].sum().unstack(fill_value=0)
summary.plot(kind="bar", figsize=(8, 4), title="Operating and Logistics Cost Trends")
plt.tight_layout()
plt.savefig("expenses_trend.png")
For an analyst, this changes the entire pace of Friday's routine: instead of hours spent cleaning cells and chasing broken references, they get a verified report in 3 minutes. Without any guesswork, the Python interpreter normalizes dates, merges shipping line items, and pinpoints the exact quarter where logistics broke the budget.
That step removes the main reason people distrust AI in analytics. The language model handles understanding your task and reading documents, while all arithmetic is handled by deterministic code that is physically incapable of making up numbers.
Artifact Studio Panel: Building Slides and Workflows from Raw Files
Getting exact totals is only half the job, since you cannot send raw script console output to a manager or client. The second pillar of the update is the right-hand Artifact Studio panel, which turns calculations into finished work files.
In our expense example, the agent does not leave numbers buried in the chat feed: it places a clickable summary table in the side panel, attaches the trend chart, and builds a slide highlighting the top three operational risks with links back to the source PDFs. A finished artifact is easy to pin, tweak, and export to Word, Excel, PowerPoint, PDF, or CSV files, or move directly into Google Docs, Google Sheets, and Google Slides. To avoid redoing finished layouts, it helps to pair this export with the two-stage outline approval mode in Gemini Notebook, which lets you sign off on the structural plan before final slide assembly.
The same pairing works far beyond finance: once a knowledge base gains executable code and file output, a folder of documents turns into a repeatable pipeline. From the accompanying release features, Julian Goldie built six practical systems—ranging from content plans, competitor analysis tables, and SEO briefs to scripted video studios and team training. The agent plans Deep Research web searches across dozens of sites on its own and supports custom roles in chat settings, while Audio Overview podcast hosts discuss findings from the Python sandbox alongside new stylized video overviews and Nano Banana illustrations. Study flashcards—familiar from connecting the notebook with Google Docs and interactive formats—also got an upgrade: they now support images on the front, a stopwatch for practicing FAQ answers in under 5 seconds, and morning reminders on Android and iPhone.
Network Isolation, Locked Libraries, and Container Boundaries
Giving a neural network permission to run code over internal documents is a sensitive step, so Google wrapped the sandbox in a strict security perimeter. Those boundaries immediately show where the tool works seamlessly and where it hits a ceiling.
First, the serverless container is completely cut off from the internet. A Python script running inside cannot pull live stock quotes from an external API, but it is also physically unable to send rows from your spreadsheet to an outside server. That is a huge win for data security, though all needed reference tables must be uploaded to the notebook in advance.
In addition, script execution is capped by a safety timeout of a few dozen seconds and a closed list of standard Python analytics packages—installing third-party libraries is blocked. In the current version, the code itself mostly runs "behind the scenes" of the agent: you see the finished chart and table in the studio, without getting an interactive terminal for manual line-by-line debugging. So for training custom models, parsing multi-gigabyte logs, or using non-standard libraries, familiar Google Colab and local Jupyter remain irreplaceable.
Agentic features are rolling out in waves, starting with Google AI Ultra subscriptions and Workspace business plans. Before uploading real financial summaries, contracts, or customer databases, be sure to check data retention policies in your work Google Workspace console—a personal account must not be used for confidential trade secrets.
Pairing an isolated Python sandbox with the Artifact Studio panel moves Gemini Notebook into an entirely different role: instead of a passive quote summarizer, you have an executive digital analyst at hand. It takes on the stickiest grunt work—cleaning mismatched tables with code without hallucinations, cross-referencing them against text PDFs, and assembling a clean summary or presentation in minutes while you keep container limits and privacy rules in mind. Looking at your own work folder right now, which recurring report or perpetually messy spreadsheet are you still reconciling by hand every week?


