Tasks in Gemini (formerly Gemini Spark)

A task can work across Gmail, Calendar, Docs, Drive, Sheets, Slides, Keep and Google Tasks, along with Contacts, Google Photos, YouTube, Gemini Notebook, Search, Maps, Flights, Hotels and Finance. Since July it edits private spreadsheets and presentations, reads comments in Docs, Sheets and Slides, and edits shared files. Connected Apps add Canva, Instacart, OpenTable, Dropbox and Zillow, and Custom Connected Apps link any personal or third party app through an MCP server URL. Signing in to a new service with Google waits for the user's confirmation.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Given a multistep goal, Gemini plans the steps and carries them out in the background as a cloud agent, so the work goes on after the laptop closes or the phone locks. One request can pull meeting notes from email and chat, write a Google Doc and draft the follow up email, and on complex work Gemini retrieves and reviews several sources in parallel. Up to 15 tasks run at the same time, and a scheduled run is skipped while 15 are already going.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Personal Intelligence lets a task draw on memory of past chats, Instructions for Gemini, the Connected Apps that work with it, the sites a user is signed into and location. Files come in from the device, Google Drive or a Gemini Notebook, and Search, Gmail and Drive supply live material as a task needs it. Each task assembles its context from those personal sources, and there is no knowledge base that a team curates or indexes for the agent.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Gemini asks before it sends a message, changes or deletes data, makes a purchase or financial transaction, submits a web form or signs in with Google, and safety checks still pause before sharing health, financial or government ID details even when an action is set to Always allow. In a remote browser the user can watch live, take over the task and hand it back. Denying an action ends that approach with no workaround, and stopping from the text box cancels the whole task. Bulk actions on private Google Tasks can run without a prompt.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Tasks run only for personal Google Accounts held by adults 18 and over, with work and school accounts excluded for now. Chrome asks once before auto browse may connect and Gemini confirms each browsing task, passwords and payment details go in on the site itself after the user takes control, and the user links each Connected App. Task mode requires Keep Activity, so chats can help train Google's models and a sample goes to human reviewers who keep it for up to three years. Tasks carry no compliance certification and sit outside any admin console.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

A progress chip on each task opens a work panel listing the steps Gemini took and the files it read or updated, and remote browser work can be watched live. Tasks and chats share one history in the sidebar, the Scheduled page lists every schedule, and conversations stay in Gemini Apps Activity, where they can be reviewed and deleted. That record lives inside the app for one person, with no exportable log of each action a task took.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Gemini keeps saved info about a user's life, work and preferences under Personal context, where each item can be switched off, edited or deleted, and it can reference past chats. A task's remote browser keeps sign in cookies between sessions and its remote computer keeps data between runs, both deletable in settings, and turning off task mode erases them. Saved info carries no retention period of its own, while chats in Activity auto delete after 18 months by default.

Sourcesupport.google.com/gemini/answer/16413516read 2026-10-24

Tasks run as a cloud agent on Google's infrastructure, with a remote browser and a remote computer that Google hosts, and on desktop they can also drive the user's own Chrome. People reach them through the Gemini mobile app, the Mac app and the web, in every country where Gemini Apps is offered except the European Economic Area, Nigeria, Switzerland and the United Kingdom. A user cannot choose where task data is processed or stored, and there is no private or self hosted option.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Skills are reusable instructions with added context that a user writes, then picks with @ or lets Gemini apply when relevant, and they also work in ordinary chats without a subscription. Saying set up, get started or interview me opens a guided conversation in which Gemini helps build a first skill and task. Every skill starts with the user, and there is no catalog of ready made skills, agents or task templates to adopt.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Schedules start a task automatically at a set date and time or in response to an event, and since June 30 they watch topics such as news, finance, sports and local events. Each schedule pauses or resumes from its task thread or the Scheduled page, and schedules stop running while task mode is off. People start and follow tasks in the Gemini mobile app, the Mac app and at gemini.google.com, notifications say what input a task needs, and Mac remote control lets the Gemini app act on a Mac from another device.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Tasks run on Gemini 3.5, Google's own model family, and no option brings in a model from another maker or routes parts of a task to one. Usage is metered in compute, rising from Google AI Pro to the two Ultra tiers, which carry 5 and 20 times the Pro limits, and Gemini warns when a user nears the limit. A person shapes a task through its goal, skills and connected apps, while the model behind it stays Google's.

Sourceblog.google/innovation-and-ai/products/gemini-app/next-evolution-gemini-appread 2026-10-24

Tasks are started and managed only inside the Gemini apps on mobile, Mac and the web. Custom Connected Apps let a task call out to a personal or third party MCP server, and developers build on Gemini models through Google's separate developer products, but no API, SDK or MCP server lets outside software create a task, send it work or read its results.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

Google warns that Gemini can make mistakes and do unexpected things during a task, and that prompt injection on the web is a real risk. Checking a result means reading the work panel's steps and files or watching the remote browser live, then correcting the task in its thread. A user has no test cases, scored evaluations or comparison of a task's output across changes to its instructions or skills.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24

On desktop, a task drives the user's own Chrome through auto browse, using signed in sites and, with permission, saved passwords, while the device and Chrome stay awake. A remote browser in Google's cloud keeps working with the device closed and pauses at steps such as sign in until the user takes over, and a remote computer runs code and keeps its data between sessions. Payments go back to the user, and Mac remote control lets the Gemini app act on a Mac from another device.

Sourcesupport.google.com/gemini/answer/17094507read 2026-10-24