Google has launched its proprietary AI agent ‘Gemini Spark’ in South Korea, igniting a full-fledged ‘browser agent’ competition. This service, which performs tasks on behalf of users in the cloud even when they turn off their devices or engage in other activities, is shifting the axis of AI competition from simple question-and-answer models to platforms that access user data and execute real-world actions.
Since the 30th of last month, Google Korea has been rolling out Gemini Spark to Google AI Pro and Ultra subscribers in both Korean and English. Spark is a personal AI agent that continues executing assigned tasks on Google Cloud even when users close their laptops or turn off their smartphone screens. Users can instruct it to find flight booking emails and organize travel itineraries, or check subscription services in their inbox and compile monthly fees into a table. It works by navigating across multiple services—Gmail, Google Drive, Docs, Sheets, and Slides—to gather information and produce deliverables.
For example, with user permission, it can use login credentials and saved passwords to schedule apartment viewings or research flights and initiate the booking process. At sensitive steps such as payment, it pauses and returns control to the user.
Google is expanding Spark to more than 160 additional countries, including South Korea, while also unveiling direct Chrome manipulation capabilities in the United States. However, Chrome’s autonomous browsing feature is initially available only in the U.S. The timeline for bringing Chrome manipulation features to South Korea has not been announced.
From Screen-Watching AI to ’24-Hour Task Delegation’ in Two Years
The starting point for browser-manipulating agents was Anthropic’s ‘Computer Use,’ unveiled in October 2024. This feature allows Claude to recognize computer screens as images, move the mouse cursor, click buttons, and type text. It drew attention because the AI could operate a computer by viewing the screen like a human, without needing dedicated integrations for each website. However, it was largely experimental at the time, with slow processing speeds and occasional misclicks.
OpenAI released ‘Operator’ in January 2025, which uses its own browser to handle reservations, orders, and online form submissions. While Anthropic’s Computer Use was closer to a developer-oriented feature, Operator was a consumer service that let everyday users delegate web tasks through natural language.
Competition escalated in the second half of the year as AI moved directly into actual browsers. In August, Anthropic piloted ‘Claude in Chrome,’ which reads webpages, clicks buttons, and fills out forms within the Chrome browser users were already using. In October, OpenAI launched ‘ChatGPT Atlas,’ a proprietary web browser with ChatGPT built in. Atlas understood the pages users were viewing and their browsing history, and could directly manipulate websites in agent mode.
In January of this year, Anthropic unveiled ‘Claude Cowork,’ designed for long-duration tasks such as data collection, analysis, file organization, and document creation. Cowork plans the entire workflow and leverages Claude in Chrome when web manipulation is required.
OpenAI, by contrast, has revised its standalone browser strategy. It is retiring Atlas on the 9th of this month, migrating related features—multi-tab navigation, file downloads, and login credential usage—into ChatGPT and Codex. Rather than pushing a standalone browser, the company aims to make ChatGPT a unified workspace where conversation, data analysis, coding, and web execution converge.
Google unveiled Gemini Spark at its annual developer conference, Google I/O 2026, in May. While early browser agents performed a single web task under user supervision, Spark continues working even when devices are off and executes recurring tasks on a set schedule or based on specific conditions.
OpenAI Pushes ‘Integration,’ Anthropic Focuses on ‘Specialized Work,’ Google Leverages ‘Platform’
All three companies are equipping their AI with the ability to manipulate the web and applications, but they differ in how they secure the execution environment.
OpenAI has chosen a ChatGPT-centric integration strategy. ‘ChatGPT Work’ aggregates data from multiple apps to produce finished deliverables like documents, tables, and presentations, while Codex handles code writing and web development. Browser manipulation technology developed for Atlas is being absorbed into the ChatGPT desktop app and Chrome extensions. The strategy is to keep users within the existing ChatGPT environment for all tasks rather than pushing them to adopt a new browser.
Anthropic concentrates on complex, long-duration professional tasks. Cowork handles data collection, analysis, and document production, while Claude Code manages development work. When website access is needed, Claude in Chrome uses the user’s login state to read pages and perform clicks and inputs. The company is extending its competitive edge in reasoning and coding into the practical workflows of developers, professionals, and office workers.
Google, rather than promoting a new workspace, is transforming the entire platform users already rely on into the agent’s execution environment. Spark operates on Google Cloud and understands work context across Gmail, Docs, and Drive. It can search for information, manipulate websites through Chrome, and send progress updates and approval requests to Android devices.
External service integrations are also expanding. Since last month, Google has been sequentially adding Dropbox and U.S. real estate platform Zillow as connected apps for Spark, following Canva, Instacart, and OpenTable. It also supports connecting third-party developer services via the Model Context Protocol (MCP).
Google’s strength extends beyond a single AI model. The company directly owns the search engine users rely on for information, the Workspace suite for email and documents, Chrome for website access, and Android as the mobile touchpoint. This allows the entire process—from finding and evaluating information to taking action—to be connected within its own platform.
The Era of AI Choosing Services… How Will South Korea Respond?
The proliferation of browser agents will fundamentally change how users select services. Until now, users compared search results and chose websites themselves. Going forward, AI will filter information and products based on criteria, and even select which services to use for bookings or purchases.
The competitive benchmark may shift from appearing at the top of search results to being selected by AI. If an agent chooses a particular shopping mall or booking platform, users are unlikely to compare alternatives directly. AI is becoming the new gateway for commerce.
Google directly owns user touchpoints including Search, Chrome, Gmail, Docs, and Maps. If Spark prioritizes connecting to these services, information discovery and task execution could be completed entirely within Google’s ecosystem. South Korean search, content, and commerce companies risk losing direct access to users and being relegated to suppliers of information and functionality to overseas AI platforms.
An industry insider noted, “Personal data and work context form the foundation of agents, and that information resides on the platforms people already use. South Korea’s AI strategy must also expand beyond model development to securing user touchpoints and service selection power.”
