New Paradigm for Web Control: Alibaba Open-Sources Page Agent, Allowing Large Models to Understand DOM

In the field of browser automation, developers have long had to rely on external tools such as Selenium, Playwright, or Puppeteer, using complex screenshots or low-level protocols to "force" web page operations. Recently, Alibaba open-sourced a JavaScript client library called Page Agent, bringing a fresh approach to this process: it transforms browser automation from external control into direct operation within the page.

The core technology of Page Agent is "DOM Dehydration." Unlike traditional solutions that convert complex web pages into multimodal images for AI recognition, Page Agent runs directly within the web page, compressing the real-time DOM structure into a lightweight "FlatDomTree" plain text mapping. This technological innovation allows models to accurately locate and execute commands such as clicking buttons or filling out forms, without dealing with heavy visual information, relying only on simplified structural text.

For developers, this solution offers clear advantages. Since Page Agent runs embedded within the page, it can seamlessly inherit user cookies and session information, eliminating the need for cumbersome backend integration and authentication work. The project adopts a model-agnostic design, supporting any large language model compatible with the OpenAI interface. In practical application scenarios, such as building an AI co-pilot within SaaS products, automating form processing, or enhancing the accessibility of applications, Page Agent provides a highly cost-effective implementation solution.

Although Page Agent performs well in terms of usability, the development team also emphasized its technical boundaries: currently, the library focuses on interactions within a single page. Additionally, due to security considerations, permission control based on prompts (such as "prohibiting automatic payments") is a guiding restriction rather than a hard logic isolation. Therefore, for high-risk operations involving fund transfers or data modification, developers still need to retain strict verification mechanisms on the server side.

Currently, Page Agent is open-sourced on GitHub under the MIT license. For teams looking to quickly embed AI operation capabilities into their own applications without investing in high costs of multimodal models, this undoubtedly provides an efficient and practical engineering choice.