Gemini gains autonomous actions to call, buy, and book
Google launches Tasks in Gemini, an agent mode capable of acting in the background, calling businesses, and paying within user-set limits. The service is free in six countries, with clear limits on reliability and availability.
On November 17, 2026, Google began rolling out Tasks in Gemini, an agent mode that allows Gemini to perform multi-step tasks, call businesses, and complete certain purchases with Google Pay, under a spending limit set by the user. The service is free in six countries — United States, Canada, Australia, India, Japan, and Korea — for users aged 18 and over, while some Pro and Ultra subscribers also have access in other markets. The goal is simple: to evolve Gemini from a responding assistant to an executing assistant, although Google warns that automation is not yet complete.
Un agent qui agit, mais pas un pilote automatique
The change is visible both in the interface and in usage. The user switches from Chat to Tasks on Gemini’s home screen, then the agent can work in the background and notify when the result is ready. Google says the tool can receive instructions via a forwarded email or SMS, access Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps, then chain multiple steps without continuous intervention.
In the example provided by Google, Gemini locates dealerships with Maps, calls two establishments in parallel, checks slots in Calendar, and books a maintenance appointment. The agent can also compare products and proceed to payment, but only within a defined framework. On some sites, it can perform much of the work before handing control back to the user for the final click. The message is clear: the agent advances, but the human hand remains in the loop. Otherwise, the web reminds who really is in charge.
Paiement, appels et messagerie sous contrôle
The most sensitive functions arrive first in the United States. Gemini can call businesses, use Google Password Manager to access approved accounts, compare products, and pay with Google Pay. Purchases go through the Agent Payments Protocol, supporting digital tokens on certain compatible Mastercard and Visa cards. Google says supported cards retain their rewards and purchase protections, that American Express will follow, and that a link with the Universal Commerce Protocol is planned to connect discovery and payment.
The user can set a firm spending limit that Gemini is supposed to respect. Receipts and tracking land in Wallet. On the communication side, Google also allows forwarding emails to gemini@google.com or verifying a US number on an eligible carrier to activate SMS. The company cites, to illustrate usage, turning a school email into a calendar event or adding grocery items to an Instacart order from text or photo, before handing control back to Instacart for the rest.
La sécurité repose sur une architecture séparée
To perform these tasks, Google relies on Agent Compute Engine, which runs the agent in a hardware-isolated virtual machine dedicated to the user’s account. The architecture separates the agent’s “harness” — the layer deciding the next action — from the execution sandbox. Persistent credentials and cloud keys remain outside the guest environment, while outgoing traffic passes through mediated infrastructure.
Google adds filters against prompt injections on search results and web pages, as well as Safe Browsing to limit links Gemini can open. Third-party connections are disabled by default, and the user can disconnect apps, modify task settings, or delete conversation history. Instructions can also be added while a task runs. For critical steps, like a purchase, Gemini must request confirmation or pre-approval; routine actions can be set to “always allow.”
Les limites restent assumées, ce qui est plutôt sain
Google emphasizes a less marketable but more useful point: Tasks can fail, slow down, or miss its target. Some tasks nearly complete but then require a final click from the user. The rollout to “hundreds of millions of people,” according to Google’s presentation, mainly expands the testing base of a still imperfect consumer agent.
This version continues Gemini Spark, introduced May 19, 2026, as a cloud personal agent based on Gemini 3.5, with Workspace integration, background operation, and recurring tasks. It also follows the arrival of Hydrogen, announced November 12, 2026, to accelerate agent workflows. Google says this model improves instruction tracking and tool selection in the Gemini app, with fewer back-and-forths to get a reliable answer. In practice, the promise remains the same: fewer clicks, more execution. Time will tell if the product maintains this line without turning the assistant into a very diligent but sometimes distracted intern.
By the numbers
- November 17, 2026 — start of Tasks in Gemini deployment.
- 6 countries — United States, Canada, Australia, India, Japan, Korea.
- 18 years and older — access condition for the free product.
- 2 cards — Mastercard and Visa, via digital tokens on compatible cards.
- 5 US states — Indiana, Louisiana, Minnesota, Montana, and Nebraska excluded from calling.