Page Agent puts GUI agents in-page

This is a smart twist on the browser-agent wave: instead of driving the web from outside, Page Agent moves the agent into the DOM and cuts a lot of the overhead that makes GUI automation brittle. If it holds up in real apps, it could become a practical building block for AI copilots inside enterprise web software.

• –The core pitch is unusually developer-friendly: one-line script injection or an npm package, plus bring-your-own-LLM support instead of locking teams into a hosted agent stack
• –Its text-based DOM approach is the real differentiator, because avoiding screenshots and OCR should reduce token costs and latency versus multimodal computer-use systems
• –Alibaba is aiming it at concrete workflows like form filling, SaaS copilots, accessibility, and admin tooling rather than generic “AI agents,” which makes the project easier to reason about
• –The optional Chrome extension for multi-page tasks shows the team knows in-page control alone is not enough for many real workflows
• –The README explicitly credits browser-use, so this looks less like a greenfield invention and more like a developer-focused repackaging of proven browser-agent ideas for client-side deployment