Google is internally testing an AI agent codenamed Remy that runs inside the Gemini app and can take autonomous actions on behalf of users, according to Business Insider, which obtained internal documents and spoke to two people familiar with the project. The agent is being dogfooded by employees in a staff-only Gemini build.
Internal documentation describes Remy as “your 24/7 personal agent for work, school, and daily life, powered by Gemini.” The key line: “It elevates the Gemini app into a true assistant that can take actions on your behalf, not just answer questions or generate content.”
What Remy Can Do
An APK teardown of Google app version 17.20 beta by 9to5Google revealed specific capability strings: Remy can communicate with others, share documents, and make purchases. To power these actions, it draws on user chats, Connected Apps, Personal context, Personal Intelligence, Agent files, and location data.
The internal description also states Remy can “monitor for things that matter to you, handle complex tasks proactively, and learn your preferences over time,” according to Business Insider. That combination of proactive monitoring and preference learning puts Remy closer to a persistent agent than the current Gemini Agent mode available to AI Ultra subscribers.
The APK strings uncovered by 9to5Google also describe a task management interface with four categories: completed, in progress, needs input, and scheduled. Users can pin, rename, and continue tasks, suggesting Remy maintains persistent state across sessions rather than treating each interaction as a one-shot query.
Safety Guardrails and User Controls
Google appears to be building oversight mechanisms into Remy from the start. The APK strings include explicit warnings that the feature is “experimental” and “can make mistakes and expose data unintentionally,” per 9to5Google. Users are told to “supervise its tasks and actions in your dashboard” and cautioned against using it for legal, medical, or financial decisions.
On data controls, the strings describe options to clear browser and cookie data (including saved login credentials), turn off Personal Intelligence and Connected Apps, and manage Personal context through settings. The agent saves browsing history and login credentials by default, a design choice that will likely draw scrutiny given Remy’s ability to make purchases.
The Distribution Advantage
Google has no timeline for a public Remy launch. A Google spokesperson declined to comment, and the project is described as a “dogfooding” effort. But Google I/O is scheduled for later this month, and agents are expected to be a centerpiece of the event. Google DeepMind CEO Demis Hassabis has spoken repeatedly about building a digital assistant as a core company goal.
The competitive context matters. OpenClaw, Hermes Agent, Perplexity’s Personal Computer, and Anthropic’s Claude Code all ship some form of autonomous agent today. But none of them sit inside a product with Gemini’s installed base. Google’s Gemini app runs on Android devices by default. If Remy ships as a Gemini feature rather than a standalone product, it arrives pre-installed on over two billion devices, according to Droid Life.
Where Remy Fits in the Agent Market
The current Gemini Agent mode, available to Google AI Ultra subscribers, handles multi-step tasks but lacks the persistent monitoring, preference learning, and purchase capabilities described in the Remy strings. The upgrade would represent a shift from reactive AI assistance (answer questions, generate content) to proactive agent behavior (monitor, act, learn).
The name Remy itself may be intentional. As Business Insider notes, it derives from the Latin “Remigius” (meaning oarsman or rower) and is also the name of the rat chef in Pixar’s “Ratatouille.” Both references fit an agent designed to do the heavy lifting behind the scenes.