Alibaba PageAgent brings browser-native GUI agents

This is a smart twist on browser agents because it moves the agent into the app instead of remote-controlling the browser from the outside. That makes PageAgent feel less like a flashy demo and more like a shippable copilot layer for real web products.

• –The biggest differentiator is architecture: PageAgent runs as in-page JavaScript, which cuts out a lot of the friction that makes browser automation stacks brittle
• –Alibaba positions it as bring-your-own-LLM infrastructure, so teams can plug in their own models instead of being locked to one hosted agent backend
• –The project explicitly avoids screenshot-first interaction and leans on text-based DOM manipulation, which should be cheaper and easier to debug than multimodal browser agents
• –Optional human-in-the-loop UI and a Chrome extension for multi-page tasks make it more usable for real workflows than a bare research repo
• –It builds on ideas from browser-use, but repackages them for client-side web enhancement, which opens up interesting SaaS copilot and accessibility use cases