An agent that could share data with third parties or make purchases without asking
In May 2026, traces of Google's Gemini Spark surfaced inside Android beta code — a quiet signal that one of the world's most powerful technology companies is racing to reclaim ground in the emerging contest over autonomous AI agents. Spark promises to act on a user's behalf across email, calendars, and the web, collapsing the distance between intention and execution. Yet embedded within the same code are the company's own warnings: that this agent may share private data or spend money without asking, a candid admission that the power being built has not yet been fully tamed.
- Google's Gemini Spark was discovered in Android beta before any official announcement, signaling a company moving urgently to close the gap with rivals Claude Cowork and OpenClaw.
- Spark can autonomously organize inboxes, summarize meetings, and compile news by reaching across Gmail, chat logs, location data, and connected services — all without step-by-step human approval.
- Internal code warnings flag the agent as experimental and capable of sharing sensitive data with third parties or making purchases without user confirmation, raising immediate privacy and security alarms.
- Google's own guidance instructs users to monitor Spark's actions and avoid trusting it with medical or financial decisions, revealing the company's awareness that the tool can cause real harm unsupervised.
- Spark currently lacks full operating system control, leaving it technically behind competitors, and its public release date remains unannounced as development continues.
In May 2026, code buried inside version 17.23 of Google's Android app revealed a project the company had not yet announced: an AI agent called Gemini Spark, formerly known internally as "Gemini Agent." The discovery placed Google squarely in a competitive race with Claude Cowork and OpenClaw, tools that have already begun automating knowledge work at meaningful scale.
Spark is designed to live inside the Gemini interface as a dual-tab tool — one side for conversation, the other for action. The distinction is deliberate. Where chat responds, the agent executes. Spark can organize a Gmail inbox, pull together summaries for upcoming meetings, and assemble a daily news digest, all by moving autonomously across Google Workspace, chat history, connected websites, and location data. Users will also be able to build custom skills for recurring tasks, echoing the "Projects" feature Claude already offers.
The ambition, however, comes with explicit internal warnings. Google's own code classifies Spark as experimental and flags two serious risks: the agent could share sensitive user data with third parties, or make purchases without asking for confirmation. The company's guidance tells users to supervise what the agent does and to avoid relying on it for medical or financial decisions — caveats that suggest Google is aware the tool can cause harm when left unattended.
Technically, Spark can control the Chrome browser and access local files, but it cannot take full command of the operating system. That boundary leaves it trailing Claude Cowork and OpenClaw, both of which operate more deeply within desktop environments. What the leak ultimately reveals is a three-way contest still in motion: Google moving fast to prove it belongs, competitors already demonstrating what's possible, and the harder question — how much autonomy is safe to hand over — still unanswered.
Google is building an AI agent called Gemini Spark that can work through your digital life without asking permission for each step. The tool appeared in beta versions of Google's Android app in May 2026, a signal that the company is moving to compete directly with Claude Cowork and other advanced AI assistants that have begun automating knowledge work at scale.
Code analysis from version 17.23 of the Google app shows the system has shed its working name—"Gemini Agent"—in favor of Spark. It will live inside Gemini as a dual-interface tool with separate tabs for chat and for agent operations. The distinction matters: chat is conversation; agent is action.
The core capability is straightforward in concept but complex in execution. Spark will handle workflows that require touching multiple services at once. It can organize your Gmail inbox, pull together summaries for upcoming meetings, and compile a daily news digest—all by reaching into Google Workspace apps and pulling data from connected services, chat history, logged-in websites, and your location. Users will be able to build custom skills to automate recurring tasks, a feature that mirrors the "Projects" function Claude offers to its users.
But autonomous operation at this scale carries real risks. Internal warnings buried in the app's code classify Spark as experimental and flag serious concerns: the agent could share sensitive data with third parties or make purchases without asking first. Google's own guidance tells users to watch what the agent does and not to rely on it for medical or financial advice. These are not minor caveats. They suggest the company knows the tool can cause harm if left unsupervised.
The technical boundaries also matter. Spark can control the Chrome browser and access files on your computer, but it cannot take full control of the operating system itself. That limitation puts it behind competitors. Claude Cowork and OpenClaw, another automation tool, can operate more deeply within the desktop environment. For now, Spark is constrained—powerful enough to be useful, not powerful enough to be dangerous in the way a fully autonomous agent might be.
What's emerging is a three-way race. Google is moving fast to catch up after competitors proved the market exists. Claude has already shown what an agent can do. OpenClaw is pushing the boundaries of what's possible on the desktop. Spark's appearance in beta suggests Google believes it can compete, but the privacy warnings and technical limits hint that the company is still figuring out how to build something powerful without building something that escapes human oversight. The public version, when it comes, will reveal how much Google is willing to trust its own creation.
Citas Notables
Users should supervise the agent's actions and avoid relying on it for medical or financial advice— Google's internal guidance on Gemini Spark
La Conversación del Hearth Otra perspectiva de la historia
Why does Google need another AI tool when Gemini already exists?
Gemini is a chatbot—you ask it questions and it answers. Spark is different. It's an agent that can act. It can open your email, read your messages, and reorganize things without you telling it to do each step. That's a different product category entirely.
So it's like hiring an assistant who knows your whole life?
Exactly. It has access to your Gmail, your calendar, your location, your chat history. It can see what you're logged into. The question is whether you trust it to make decisions on your behalf.
The source mentions it could make purchases without asking. That seems like a major problem.
It is. That's why Google put warnings in the code. They're saying: this thing is experimental, don't let it loose on anything important. It's a way of saying we built this, but we're not sure we've solved the safety problem yet.
Why release it in beta at all if it's that risky?
Because Claude Cowork is already out there doing this. Google can't afford to wait until it's perfect. They need to learn from real use, see what breaks, understand what users actually want. Beta is how you do that without fully committing.
What's the difference between Spark and Claude Cowork?
Claude can control your entire desktop. Spark can't—not yet anyway. It's limited to apps and the browser. That's a real gap. But Google has the advantage of owning Android and Chrome, so it can integrate deeper over time. This is the opening move, not the final product.