What happened
Google announced on November 17, 2026 that Tasks is rolling out free of charge to hundreds of millions of people in the United States, Japan, South Korea, Canada, Australia, and India. The feature runs on Gemini 4 Hydrogen. In the Gemini app, users switch from Chat to Task and describe the work they want done. Google presents Tasks as suited to longer projects and recurring schedules rather than quick exchanges. In the United States, users can also forward emails to gemini@google.com or text Gemini after configuring the feature in the app. Texting requires a US phone number with an eligible carrier.
The examples in Google's announcement range from calendar management to shopping. In one, Gemini called two nearby car dealerships, waited on hold, found an early Friday appointment, and added it to the user's calendar in under five minutes. Google says the assistant disclosed that it was an AI to the businesses, and that businesses can opt out of calls. Another example describes granting access to a utility portal through Google Password Manager, then asking Gemini to analyze past bills in email; the assistant identified a heat pump running during peak electricity pricing. These are examples from Google, not independently measured results.
For shopping, Google says Gemini can compare products using its Shopping Graph, ask follow-up questions, and help complete checkout through Google Pay. Purchases use the Agent Payments Protocol (AP2). Google says it is working with American Express, Mastercard, and Visa on digital tokens for supported cards; American Express support is described as coming soon. Users can set a hard spending cap, and receipts and tracking information go to Google Wallet. Google says customers retain card points, cashback, and purchase protections. In a grocery example, Gemini added items to an Instacart order from typed instructions and a photo, then handed the user off to Instacart to complete the order. Google says Gemini connects with more than 60 popular apps.
Google describes the execution environment as a hardware-isolated virtual machine dedicated to an account and not shared with other users, using Google Cloud's Agent Compute Engine. It says the VM is sandboxed, regularly tested by security red teams, and supplemented by prompt-injection filters and Safe Browsing. Users can intervene during a task; Gemini is designed to ask before critical actions such as clicking Buy, and to ask clarifying questions when instructions are ambiguous. For routine actions, users can set “Always allow.” Google also says users can disconnect apps, adjust task settings, and delete chat history. It acknowledges that tasks can fail or miss the mark, and that on some websites Gemini may stop short and leave the final click to the user.
What the archive adds
New Runtime covered Gemini 4 Hydrogen's global rollout in Search on November 14, 2026. This is a new event on a known model story: Gemini 4 Hydrogen now powers a consumer task feature with connected services, payment controls, and delegated actions.
Trend signal
This supports agent-ready software exposing capabilities: Gemini's usefulness depends on actions exposed through connected apps, with user permissions and confirmations shaping what the agent can do. The announcement also makes clear that execution boundaries matter, but it does not establish how well those controls perform in real-world use.
Why now
Google says Tasks is rolling out now across six countries and connects the launch to a broader expansion of Gemini's ability to complete everyday tasks. The timing is therefore a product rollout, not a response to a disclosed incident or external deadline.
Connections to other events
New Runtime's earlier coverage described Gemini 4 Hydrogen becoming available globally in Google Search on November 14, 2026. The new announcement extends the same model into an assistant workflow that can interact with apps and services, rather than only providing Search results.
Why it matters
For teams building agents, the announcement makes the consumer product boundary concrete: useful work depends on a combination of service access, scoped permissions, execution isolation, and human approval at consequential steps. AP2 and the spending cap illustrate how payment actions can be bounded instead of delegated as an unrestricted instruction. The described sandbox and the ability to intervene are relevant design choices for systems that handle private data or external actions.
The caveat is equally important. Google’s examples show a product direction, not evidence that the workflow is dependable across businesses, websites, or unusual cases. Builders should treat handoffs and confirmation points as part of the task design, rather than assume an agent will complete every action end to end. If adopting similar workflows, start with reversible tasks and require approval before actions that spend money or commit users.
Limits and uncertainty
Google says it has tested Tasks extensively but has not eliminated mistakes; it gives no independent success rate, failure rate, or measure of how often users must intervene. It does not quantify latency, task limits, or the scope of supported apps beyond saying there are more than 60. The announcement does not describe how long task data is retained in the isolated environment or provide an independent security assessment. Availability of particular actions varies: calling is unavailable in five named US states, and some websites may require the user to finish the final step.
Skill to build now
For any agent that can act on a user's behalf, separate low-risk preparation from consequential execution. Let the agent gather options, fill a cart, or draft a schedule, then require explicit confirmation before purchase or booking. Where a task involves spending, set a hard cap before handoff. Google’s examples make these controls concrete: routine trusted steps can be pre-approved, while purchases are held for confirmation.
Reader takeaway
Tasks is a meaningful shift toward assistants that act across connected services, not just answer questions. Its design pairs that access with isolation, user intervention, and approval for consequential steps. The practical lesson for builders is to make permissions and confirmation part of the workflow itself—and to judge the product by measured reliability, not the launch examples.