You can ask Gemini to handle multi-step tasks for you with your supervision. In the Gemini web app, you can use the experimental feature called Gemini Agent ( Labs) to:
- Categorize your emails and draft replies
- Send you a summary of the day ahead
- Rebuild your calendar to help you build habits toward your goals
- Conduct research using the live web browser, including sites you sign in to
- Make restaurant reservations and book accommodations
What you need
To use Gemini Agent, you need to:
- Be 18 or over and in the US.
- Be signed in with a personal Google Account. This feature isn’t available if you sign in to a work, school, or supervised Google Account. Learn how to sign in to Gemini Apps.
- Have a Google AI Ultra subscription on your personal account.
- Have Keep Activity on.
Important:
- For now, this feature is only available in English.
- For now, this feature isn’t available in the Gemini mobile app, Gemini in Google Messages, Gemini in Chrome, or Gemini on Android XR.
Before you use Gemini Agent
Important: If you ask Gemini Agent to use a service like Gmail or Google Docs in the live web browser, it may perform the action even when the related app in Gemini is disconnected.
Do not enter sensitive info, like passwords and payment info, directly in chatDo not enter sign-in info, payment details, or other sensitive info directly into the chat. Instead, if necessary to complete your action on a webpage, take control of the live web browser and enter the details directly on the webpage.
You can ask Gemini to schedule actions for recurring tasks, like daily briefings. Gemini is still learning and can make mistakes. To help keep your information safe and prevent unintended actions, avoid scheduling actions for important or sensitive tasks. If a scheduled action runs when you are offline, you may not be able to stop Gemini from completing an unintended task. Some tasks that use the live web browser may require your confirmation before Gemini completes them. Learn more about scheduled actions.
Use Gemini Agent to complete complex tasks
With Agent, you can ask Gemini to help with multi-step tasks.
- On your phone or tablet, go to gemini.google.com.
- In the text box at the bottom, tap Tools
Agent
.
- In the text box, enter details about the tasks you want Gemini to complete or try an example.
- Tap Submit
.
- If your task requires Gemini to use an app, like Google Workspace, that app must be connected. If it isn’t connected, Gemini will ask you to connect it. Learn how to use and manage apps in Gemini.
- Gemini may ask for more details before it creates an action plan for your request.
- If you want to make changes to the plan, enter a prompt that explains what you want updated and submit.
- Review each task and tap Decline
or Confirm
. You can also tap Decline all or Confirm all.
- It usually takes a few minutes to complete the request. For more complex tasks, it may take longer.
Tips:
- Gemini Agent is still in early development. If Gemini does something you don't like, you can stop it by tapping Stop response
in the chat or under "Using browser," tapping Open
.
- If you use Gemini Agent to create a recurring task, the task is saved in your scheduled actions. Learn how to manage scheduled actions.
Examples
- Stay organized: Every morning at 7 AM, take a look at my unread emails and summarize what I need to know, create tasks for things I need to follow up on, and archive unimportant emails.
- Get briefed: Every weekday at 8 AM, send me a briefing document to help me prepare for my day ahead.
- Place an order: Order my usual tall iced vanilla latte with oat milk from the closest drive-thru. Pre-fill the pickup order so all I need to do is confirm.
Take control of Gemini's browser
With Agent, Gemini can complete some of your tasks in a live web browser. For example, it can navigate to a website and interact with the page, like selecting an item from a menu for you. You can take control of the browser at any time. For certain tasks, Gemini will ask you to take control.
- In the chat under “Using browser,” tap Open
.
- Complete your steps.
- To give back control to Gemini, at the bottom, tap End control.
- Tap Resume.
Share a chat where you used Gemini Agent
Important:
- If you share a chat where Gemini interacts with a remote browser, screenshots of what it did in the browser are included. Before you share a chat, review the public link and make sure it doesn’t include any private or sensitive info.
- When you share a chat, a public link is created and anyone with the link can view it. Learn more about sharing chats from Gemini Apps.
You can share chats where you use Agent to complete tasks. Learn how to share your chats.
Delete remote browser data
When Gemini uses its live web browser to complete tasks, browser data, like cookies that contain your website authentication info (your sign-in details), is saved for future sessions for convenience. You can delete all saved browser data from past sessions, at any time. To help protect your privacy, Gemini may automatically delete saved browser data on your behalf, and you may need to sign back in to websites.
- On your phone or tablet, go to gemini.google.com.
- At the top, tap Menu
- For "Remote browser data,” tap Delete.
Gemini Agent limits
In Gemini Apps, there are limits for the number of:
- Requests you make using Gemini Agent each day
- Tasks you can run using Gemini Agent at the same time
If you’re close to your limit, Gemini lets you know how many requests are left for the month. Learn more about limits in Gemini Apps.
How Gemini Agent works with you to keep you & your info safe
Gemini Agent is an experimental feature in early development. As we continuously work to improve Gemini Agent, your supervision is important to help prevent unintended and potentially harmful actions.
AI agents, like Gemini Agent, are helpful and powerful tools, but it's important to be aware of some possible risks. Remember that Gemini can make mistakes and do unexpected things.
About prompt injection
Prompt injection is an attempt to elicit an unintended or harmful response from generative AI tools like Gemini Agent. This happens when content, like a website, an email, a document, or multimedia, has malicious instructions that might be hidden from you but visible to the AI agent.
If an AI agent reads these instructions while doing a task, it might be tricked into doing unintended things. For example, a malicious instruction could tell the agent to:
- Take your private info from your emails or documents and post it on a public website.
- Send your emails in Gmail to an external service without you knowing.
- Expose insights about you based on your data in connected apps.
Gemini can share your info with websites
When you use Agent, Gemini can share information from your chat with websites while using the live web browser. This can create privacy risks, for example when Gemini Agent uses connected apps like Google Workspace which allow Gemini to access your private info.
To help reduce possible risks associated with the use of AI agents, Gemini Agent includes features designed to support safer use and mitigate the likelihood of unintended or potentially harmful actions. While these do not guarantee protection against all risks, these features include:
- Prompts for user confirmation: Gemini is designed to ask for your review and confirmation before it completes certain actions, such as:
- Sensitive actions, like sending communications, modifying your data, making purchases, and submitting web forms.
- Other actions that require your confirmation, like scheduling events.
- Prohibited task recognition: Gemini has safeguards designed to help it recognize and not act on certain types of requests that could be harmful or are outside of its intended purpose, including certain tasks that may violate the Generative AI Prohibited Use Policy.
- “Take control” mode: Gemini may pause and ask you to “take control” of the browser to complete specific actions, including entering specific types of sensitive information (such as passwords or payment details).
- Planning info: Gemini is designed to show you what it’s planning to do after you submit a prompt, including what information it plans to share with the live web browser.
Gemini’s safeguards for Agent described above don’t guarantee protection against all risks.
Supervise important & sensitive tasks
Your active supervision is the most important way to protect against risk while using Agent. Be extra watchful when Gemini:
- Handles an important task
- Is signed in to websites with sensitive access or info
- Does anything where a mistake could be a problem
- Uses connected apps, like Google Workspace
You’re in control
If Gemini is doing something you don't like, you can stop it by selecting Stop response in the chat or Take control
over Gemini’s browser. Gemini’s safeguards aren’t intended to replace your active supervision. You can also help keep your info safe by deleting your remote browser info after sensitive sessions.