Google tests Remy, a Gemini AI personal agent

Google is running a staff-only trial of Remy, a new Gemini AI personal agent that can act on users’ behalf, according to internal documents and interviews with two employees. The trial is limited to employees testing the tool before any public release. A Google spokesperson declined to comment. The company has not disclosed a timeline for wider availability or which services are included in the employee test.

Internal materials describe Remy as a ’24/7 personal agent’ built to perform tasks across Google services, monitor items relevant to users and learn individual preferences. The project is presented as a further step beyond Gemini’s chat responses and existing agent features, such as Agent Mode, by allowing the assistant to take actions on behalf of a user rather than only offering suggestions.

Google’s public documentation for Gemini lists the apps and services an agent can connect to. That connected-app surface includes Gmail, Calendar, Docs, Drive, Keep and Tasks in Google Workspace, plus GitHub, Spotify, YouTube Music, Google Photos, WhatsApp, Google Home and Android utilities. Those connections are the mechanism through which an agent like Remy would access information and carry out actions.

Gemini’s Privacy Hub provides tools for reviewing and deleting Gemini Apps Activity, setting auto-delete preferences and choosing whether data can be used to improve Google AI. It also includes controls for which apps and data Gemini can access and for managing information users ask Gemini to save. Remy’s reported ability to learn preferences and store memories focuses attention on how those controls will be presented and enforced for actions taken automatically.

Google’s documentation describes different levels of agent actions, from reading information in Workspace apps to creating calendar events, sending messages, launching apps and controlling smart-home devices. The internal reports do not specify whether Remy will require user confirmation for each action, what approval workflows will look like, or how completed actions will be logged and audited. Technical details such as the model version behind Remy and the degree of autonomy being tested were not disclosed.

Guidance from Google Research and Google Cloud emphasizes that AI agents should operate under defined human controllers, use limited privileges, keep actions observable and maintain transparent, auditable logs. Those recommended principles relate to the open questions around permissions, oversight and logging for an agent that can act across connected services.

The project’s dog-fooding status indicates initial testing is limited to Google employees. Google has not provided further details about data handling, memory controls, approval mechanisms or the scope of Remy’s autonomy.