Google's agents work while you sleep

The headline from Google I/O wasn't a smarter answer box. It was agents that stop waiting for you to ask. Google's new information agents run in the background, around the clock, scanning the web, reasoning across sources, and pushing you an update when something you care about changes — no new query required. Tell it "keep me posted on apartments matching these exact requirements," and it just… watches, and taps you on the shoulder when a listing fits.

And it doesn't only watch. Google is expanding agentic booking to local services, and for some categories — home repair, beauty, pet care — you can ask Google to call businesses on your behalf. Search went from a thing you pull — type a query, get answers — to a thing that pushes, and increasingly one that acts. That's a bigger deal than the demo makes it look, and it changes the bar for anyone building anything with AI in it.

The chatbot is dissolving into the background

For two years the mental model of "AI product" was a box you type into. This is the start of that box disappearing. The interesting agents of 2026 are ambient: always on, but not in the way, reacting to your situation instead of waiting for a prompt. When the biggest search engine on earth makes "we watch for you" the flagship feature, that becomes the new baseline expectation. Users will start to ask why your product makes them check it, when Google's just tells them.

That's the opportunity and the trap. Proactive is genuinely more useful — but it introduces a hard new design problem that reactive never had: when is it worth interrupting someone? A reactive tool only speaks when spoken to, so it can never annoy. A proactive agent has to earn every notification. Get the relevance and timing right and it feels like magic; get it wrong and it's a spam machine you mute in a day. The skill stops being "answer well" and becomes "know what's worth surfacing, and when."

Watching 24/7 has a meter, and acting has a cost of its own

Two quieter consequences hide under the magic.

The first is money. An agent that runs around the clock is doing inference around the clock. That's the token meter running while you sleep — not per user request, but continuously, per user, forever. The economics of "always on" are completely different from "responds when asked," and anyone copying this pattern needs to do that math before shipping, not after the bill arrives. The freedom to watch everything is only free if you've designed the watching to be cheap.

The second is grounding. A background agent monitoring prices or listings is only as good as the live, real data it checks against — it can't be guessing. This is grounding as a hard constraint again: a proactive agent floating free of a real source of truth doesn't just give a bad answer, it pings you with a confidently wrong one at 2 a.m. The push model punishes ungrounded agents far harder than the pull model did, because nobody asked, so every wrong interruption is purely your fault.

When the agent acts, accountability gets real

The part that should give every builder pause is the calling-a-business feature. An agent placing a booking — or literally phoning a salon as you — is no longer returning information. It's taking an action in the world, with your name on it, that's often hard to undo. That raises exactly the accountability question I wrote about this week: when the agent books the wrong thing, or talks to a business as if it's you, who owns that?

The answers aren't mysterious, but they have to be designed in, not bolted on:

• Confirm before anything irreversible. Watching and suggesting can be autonomous. Booking, buying, sending, calling — anything you can't take back — should pass through a human yes, or be tightly bounded by rules you set.
• Give the acting agent a budget and a leash. An agent empowered to transact needs hard limits — spend caps, allowed categories, off-limits actions — the same way you'd scope any tool that can touch the real world.
• Be honest that it's an agent. An agent calling a business on your behalf edges toward impersonation. Disclosure isn't just polite; it's where the next round of rules is heading.

The takeaway

"Search that works while you sleep" is a great line, and the shift under it is real: agents are moving from answering when asked to watching continuously and acting on their own. If you build products, the expectation bar just moved with it — toward proactive, ambient, and action-taking.

Just don't copy the magic without copying the discipline underneath. Proactive means you have to earn every interruption. Always-on means a meter you can't ignore. And acting on someone's behalf means owning the consequences when the agent gets it wrong. The future Google demoed is one where the agent does things while you sleep — which is wonderful right up until the morning you wake to something it did that you'd never have approved. Build so that morning never comes.