Google’s Gemini Spark Leak Hints at Always-On AI Agents Ahead of I/O

Google appears ready to shift its AI assistant from a reactive chatbot to a proactive background operator. Leaked details over the past two days reveal an experimental feature called Gemini Spark, designed to handle email triage, meeting prep, and custom news updates without requiring constant user prompts.

First spotted in an onboarding screen within the Gemini web app, the feature was confirmed by beta code in the Google app version 17.23. According to Android Authority, Spark can clean junk from Gmail inboxes, assemble notes before key meetings, and generate a tailored news digest that tracks stories as they develop. These tasks go far beyond what current Gemini versions deliver.

But the trade-off is significant. Spark requires deep access to personal data—connected apps, chat histories, scheduled tasks, location data, and more—and can share that information with third parties. The disclosures appear in screenshots shared by developer Fandu on X and analyzed across tech sites.

TestingCatalog captured the full onboarding message: “Let Gemini do more as your everyday AI agent, ready 24/7 to help with your inbox, online tasks, and more.” The feature builds on an earlier internal codename called Remy and ties to an existing Gemini Agent option limited to Google AI Ultra subscribers.

Privacy warnings come attached. The text states Spark is experimental, designed to ask permission before sensitive actions, yet it may share information or complete purchases without asking. Users must supervise it. The notice explicitly advises against relying on the agent for medical, legal, financial, or other professional advice.

9to5Google found Spark in a redesigned navigation drawer within the Android app beta. A two-tab layout splits Chat and Agent views. Users create new tasks in the Agent view and see lists of active work and scheduled items. The icon resembles a comet streaking across the interface.

Capabilities described in the leaks paint a picture of practical productivity gains: decluttering inboxes by summarizing newsletters and offering unsubscribe options, preparing concise briefs with relevant documents before meetings, tracking news stories in depth, and creating reusable skills for recurring tasks.

Some limitations remain visible. Current screenshots show integration mainly with Google’s own apps. Third-party support appears possible but unconfirmed. The agent stops short of full computer control, unable to match projects like Anthropic’s Claude Computer Use or OpenAI’s Operator in raw system-level actions. Still, the direction feels clear: Google wants agents that operate independently once given goals.

And the timing adds weight. Google I/O opens next week on May 19. Multiple outlets suggest Spark could feature prominently in the keynote. The leaks arrive just days before the event—a pattern familiar from past product rollouts.

Concerns about data sharing have already surfaced in online discussions. Reddit threads linked to the 9to5Google story show users reacting to the list of data sources and the explicit warning about purchases without confirmation. Google has provided settings to clear remote browser data, disable connected apps, and manage activity controls.

The emergence of related model names fuels speculation about underlying technology. Leaks mention Spark Robin, described as offering rich visual responses, and a new MCP Tool Testing category in the model selector. These elements hint at expanded multimodal abilities and tighter integration with external tools.

Google has not commented publicly on the leaks. That silence is standard before major events. Yet the volume of coordinated details across web app, Android beta, and third-party testing suggests internal testing has reached an advanced stage. Power users who already pay for premium Gemini access stand to benefit first.

Industry watchers will dissect the I/O presentation for clues about broader strategy. How deeply will Spark integrate with Android, Workspace, and Chrome? Can the company balance powerful capabilities with the transparency users demand? The leaks provide a preview. They do not deliver final answers.

What feels certain is the trajectory. Chat interfaces alone no longer satisfy. Users want systems that act—that monitor, prioritize, and execute on their behalf. Gemini Spark, if delivered as described, positions Google to compete directly with agent-focused offerings from Anthropic, OpenAI, and others. The race has moved from who answers questions fastest to who finishes the work while users focus elsewhere.