Google introduces Gemini-powered "auto browse" agentic capabilities to Chrome for enterprise users.
Moving AI agents directly into the browser DOM is a massive shift from isolated copilots to integrated execution engines. By leveraging Gemini to read live tab context and interact with web apps, Google is effectively turning Chrome into a universal RPA tool. The strict "human in the loop" requirement acknowledges the current unreliability of autonomous web agents while still drastically reducing friction for repetitive enterprise workflows.
What happened
At Google Cloud Next, Google announced new "auto browse" agentic capabilities for Chrome Enterprise users. Powered by Gemini, this feature allows the browser to understand the live context of open tabs and autonomously execute web-based tasks. Use cases highlighted include cross-tab data entry (e.g., moving data from a Google Doc to a CRM), comparing vendor pricing, and scheduling. Crucially, the system operates with a "human in the loop" constraint, requiring manual user confirmation before finalizing actions.
Technical details
Unlike traditional API-based integrations or standalone desktop apps, this implementation embeds the LLM directly into the browser's execution environment. Gemini is granted access to the live Document Object Model (DOM) across multiple tabs, allowing it to parse unstructured web data, maintain state across different web applications, and simulate user interactions (clicking, typing, navigating). This effectively bypasses the need for explicit API integrations with third-party SaaS tools, relying instead on visual and structural DOM comprehension to execute tasks like a standard user would.
Why it matters
This is a significant evolution in Robotic Process Automation (RPA). Historically, web automation required brittle CSS selectors, complex Selenium/Playwright scripts, or expensive API orchestration. By injecting a multimodal LLM into the browser, Google is democratizing automation, turning every web app into an AI-accessible endpoint without requiring developer intervention. For engineers, this signals a shift where the browser itself becomes the primary agentic runtime. However, the explicit "human in the loop" requirement highlights the ongoing challenges with LLM hallucination and action reliability in dynamic DOM environments.
What to watch next
Monitor how Google handles local versus cloud processing for these DOM-scraping tasks, especially regarding enterprise data privacy and latency. Additionally, watch for how web developers might need to adapt—or defend—their applications against agentic interactions. If "auto browse" gains traction, we may see the emergence of new web standards designed specifically to optimize or restrict LLM-driven DOM traversal.