Google I/O 2026: 'Remy' Is Google's 24/7 AI Agent, and It Lives Inside Your Phone

Google I/O kicked off yesterday and the keynote was two hours of AI. Not "AI features." Not "AI-powered." The entire thing was AI. The one thing people will actually remember: Remy.

Notably, Google showed nothing about AI in Search beyond "we'll share more later this year." The Search silence is telling. Google's cash cow is under real threat from AI search products like Perplexity and ChatGPT Search, and the company offered no plan for defending it. Whether that means they have nothing ready or are saving it for a separate event is unclear.

What Remy actually does

Remy is a 24/7 AI agent built on top of Gemini 2.5 Pro. It lives on your Android phone (iPhone later, Google says). Unlike Google Assistant which waits for you to ask it something, Remy does things on its own timeline:

• It reads your incoming emails, figures out which ones matter, drafts replies for the ones that need them, and archives the ones that don't
• It monitors your calendar for conflicts, proposes reschedules before anyone notices the double-booking, and preps briefing docs from your email and Drive before meetings
• It watches real-time traffic on your commute route and moves your first meeting if you're going to be late, sending the "running 10 minutes behind" message without you touching anything
• At end of day, it generates a summary of what happened, what decisions got made, and what's waiting for you tomorrow

Under the hood, Remy runs a hybrid architecture. A smaller Gemini Nano model handles on-device processing for privacy-sensitive tasks. When a task requires more reasoning, it calls out to Gemini 2.5 Pro in the cloud. Google calls this "tiered inference," the same pattern Apple described for Apple Intelligence but never fully shipped.

Google CEO Sundar Pichai called it "the first AI that doesn't wait to be asked." That's the right description. And it's also what makes me uncomfortable.

How Remy compares to the agent field

The AI agent space is getting crowded fast. Anthropic's Claude with Dreaming focuses on computer use: controlling browsers, filling forms, managing files for knowledge workers. OpenAI's agent platform targets enterprise workflows. Meta's Hatch is consumer-focused inside Meta's platform. Remy is positioned differently: a personal productivity agent operating across services people already use, Gmail, Calendar, Maps, Drive, Messages. The strategy is classic Google: own the integration layer.

The distribution numbers make this strategy hard to bet against. Android has 3 billion active devices. Gmail has 2 billion users. Google Maps has over a billion monthly users. No other AI company can plug an agent into that many existing touchpoints. Anthropic has better writing quality. OpenAI has stronger brand recognition. Neither has distribution. Google also has data nobody else has: commute patterns from Maps, communication habits from Gmail, schedule from Calendar, document workflow from Drive. Combined, they let Remy do things competitors simply cannot replicate. An Anthropic agent can write a brilliant email. It cannot know you are about to be late because I-280 is backed up.

The privacy math

Remy needs access to everything. Email. Calendar. Messages. Location. Drive. Call logs. Probably your thermostat and your doorbell, eventually. Google says all processing happens on-device via the Tensor G6 chip, that "your data never leaves your phone," that Remy's memory is encrypted and local.

Google has said versions of this before. The company's track record on privacy promises is, let's say, not perfect. The technical claim about on-device processing might be true for the Gemini Nano model running locally. But Remy connects to the internet to send emails, check traffic, and pull meeting briefs. Some data leaves the phone. The question is what else goes with it.

The demo showed Remy declining to forward a sensitive-looking email without asking first. It flagged a calendar conflict and suggested a fix rather than auto-applying it. These are good signs. They're also carefully scripted demos at a developer conference. Production is different.

There is a deeper tension Google did not address. Remy's value comes from acting autonomously. The more autonomy you give it, the more useful it becomes. But every autonomous action is an opportunity to do something wrong. Google's solution is a permission gradient: low-stakes actions happen automatically, medium-stakes get suggested, high-stakes require explicit confirmation.

The question is who defines the stakes. Google's demo classified email forwarding as high-stakes and calendar rescheduling as medium-stakes. For a journalist with an anonymous source, forwarding any email from that source is catastrophic. For a salesperson closing a deal, a wrongly rescheduled meeting could cost a commission. The risk profile is personal, and a one-size-fits-all permission system will get it wrong for someone.

What this means if you use Android

If you have a Pixel or recent Samsung phone, Remy arrives in beta next month. U.S. only initially. Google says iPhone support comes "later this year," though Apple's restrictions on background app activity and system-level access mean the iOS version will almost certainly be less capable than the Android version, even if Google does not say so directly.

For the average Android user, Remy represents something genuinely new: an AI that does not require you to open an app and type a prompt. It just handles things. The psychological shift from "I use AI" to "AI handles things for me" is bigger than any individual feature. Google is betting that once people experience that shift, they will not want to go back. Based on how sticky Google's previous integration-first products have been (Gmail, Maps, Photos), the bet is probably right.

No pricing has been announced, but the betting is that Remy lands inside the Google One subscription tier, likely the Premium plan at $10 per month. If Google includes it in the base tier or makes it free, the strategic goal is clear: lock users into the Android-Google platform so deeply that switching to iPhone or Outlook becomes unthinkable. If it is a paid add-on, the goal is revenue diversification beyond advertising. Either way, Remy is not a side project. It is Google's biggest platform bet since Android itself.

The agent race just got faster

Anthropic shipped Dreaming for Claude Managed Agents last week. OpenAI has its enterprise agent platform. Meta launched Hatch for consumers in April. Now Google has Remy.

What makes this moment different from previous AI feature races is that agents are sticky in a way chatbots are not. Switching from ChatGPT to Claude is a matter of opening a different website. Switching from Remy to a competitor's agent means unplugging the thing that reads your email, manages your calendar, and routes your notifications. The switching costs are enormous. Google knows this. Getting Remy onto phones first, before Apple ships anything comparable, before Meta expands Hatch beyond its own apps, is a land grab for the default agent on every device.

I will try Remy when it ships. I already use Gmail, Calendar, Maps, and Drive every day, so I am exactly the user Google is designing for. I expect it will be genuinely useful, probably more useful than any AI product I have used so far. The question that will determine whether I keep using it is not about features. It is about trust. Does Remy make mistakes I can live with, or mistakes that create problems I then have to spend time cleaning up? Does it respect the boundaries I set, or does it gradually nudge me toward giving it more access than I am comfortable with? Google's track record with gradually expanding data collection, think about how Maps went from "your location while using the app" to "your location always," does not inspire confidence. I am going in with my eyes open and my privacy settings dialed as tight as they will go. I recommend you do the same.