Exploring Agent Browser: AI agents on the web

Agent Browser, developed by Vercel Labs, offers a novel approach to browser automation by leveraging an accessibility tree instead of the raw DOM, significantly reducing AI token usage by up to 97% per request compared to traditional methods like Playwright. This headless browser automation CLI is specifically designed for AI agents, providing stable identifiers for elements which remain robust even if underlying CSS classes change. By condensing interactive elements into deterministic references, Agent Browser overcomes the common issues of token bloat and fragile selectors, making it highly efficient for dynamic web applications. It supports multi-session operations and integrates seamlessly with AI coding tools like GitHub Copilot, allowing agents to perform tasks such as automated testing and exploratory testing with semantic locators. While Playwright remains a standard for traditional end-to-end testing, Agent Browser excels in scenarios where AI agents need to process dynamic page structures efficiently.