이미지: Google
You close your laptop. But the work keeps going.
On the 30th of last month, Google expanded its personal AI agent 'Gemini Spark' to more than 160 countries, including Korea. Starting the same day, Google Korea has been rolling out Korean- and English-language service to Google AI Pro and Ultra subscribers.
Spark's defining trait fits in one sentence: it runs on Google Cloud, not on your device. That means the task you assigned keeps going even after you close your laptop or turn off your phone screen.
From Assistant to Agent
Until now, AI was close to asking a taxi driver for directions. Even with a good answer, you still had to drive. Spark, by contrast, takes the keys. Google describes this as a shift from an assistant that answered questions to "an active partner that actually does the work under my direction."
Technically, it runs on Gemini 3.5 using the 'Antigravity' harness. It is deeply integrated with Workspace tools like Gmail, Docs, and Slides, so the place where data lives and the place where output is produced sit within the same ecosystem.
Google cited three example use cases: a recurring task that automatically scans monthly credit card statements to spot newly added subscription fees; a learning task that continuously checks emails from a child's school and sends a once-daily summary of just the deadlines; and a bundled task that gathers meeting notes scattered across emails and chats, organizes them into a document, and even drafts a project kickoff email.
Browsing the Web While Logged In
The real substance of this update is Chrome integration. With user permission, Spark handles web errands using already logged-in accounts and saved passwords as-is. For example, it can pick a saved home listing to schedule a viewing, or research flights and start the booking process.
Google has disclosed two layers of defense here. One is prompt injection blocking — preventing attacks where hidden text on a webpage feeds the AI rogue instructions. The other is pausing at sensitive steps, such as payments, and handing control back to a human.
However, automatic Chrome browsing is launching in the US first. Google has only mentioned plans to expand to other regions, and according to Digital Daily, no timeline for a Korea launch has been disclosed.
Two Years to Get Here
The starting point for browser-manipulating AI was Anthropic's 'Computer Use' in October 2024. It worked by viewing the screen as an image and moving a cursor to click buttons — a slow experiment that often clicked the wrong spot.
The pace picked up quickly after that. OpenAI released 'Operator,' which handled bookings and orders, in January 2025, and by the second half of that year the field had moved to putting AI directly inside a real browser — Anthropic's 'Claude in Chrome' and OpenAI's own browser, 'ChatGPT Atlas.'
The three companies' strategies now diverge. OpenAI will shut down Atlas on the 9th and fold its browser features into ChatGPT and Codex — consolidating everything into ChatGPT rather than distributing a new browser. Anthropic is focusing on long, specialized work with 'Claude Cowork.' Google is the only one of the three that chose to turn an entire platform people already use into the execution environment.
What Only Google Has
Google's strength lies not in its model but in its deployment. Search for finding information, Workspace for handling email and documents, Chrome for opening the web, and Android in people's hands — Google owns all of it. The entire process of an agent finding, judging, and executing stays within one company's walls.
Google is also expanding external connections. Following Canva, Instacart, and OpenTable, Dropbox and the US real estate platform Zillow have joined as connected apps, and the system also supports linking services built by outside developers via MCP (Model Context Protocol).
The scale of the Gemini app itself is worth noting. At I/O this past May, Google said more than 900 million people use Gemini every month across 230 countries in more than 70 languages. Spark is a feature built on top of that base.
Editor's Take
Try several browser agents and you run into the same wall every time. What looked smooth in a demo grinds to a halt at a login screen on your own machine. Being 'already logged in' — nothing at all for a human — turned out to be a hard threshold for AI to cross.
That's exactly where Spark's Chrome integration stands out. Instead of making you install a new browser, it uses your already-logged-in Chrome as-is. It's less a technical leap than a deployment fix that removes the threshold — but in real-world use, that kind of fix ends up mattering more.
At the same time, this is also the most delicate part. Handing the web over to an AI that uses saved passwords means the scope of what can go wrong widens from 'getting a wrong answer' to 'something happening on your own account.' That's likely why Google went out of its way to mention prompt injection defenses and payment-step handoffs in its announcement. It also reads as the reason this feature is launching in the US first.
For domestic businesses, a more important shift lies elsewhere. Until now, users compared search results and picked sites themselves. Going forward, AI will narrow down candidates that fit the criteria and even decide which service to use for a booking or purchase. The bar for competition is shifting from 'ranking at the top of search' to 'being chosen by the agent.'
There's one thing worth checking right now: is our service in a state that software, not just people, can read and call? If information is locked inside images, or buried behind a complicated post-login process, an agent will simply skip past it. This is also why connection standards like MCP have suddenly become important.
This time, Korea only received access; automatic Chrome browsing hasn't arrived yet. If that time gap is time to prepare, it may be worth looking first at how our service appears to agents at the user touchpoint, rather than racing to build a matching model.
