The browser as an agent boundary

A browser agent has to navigate pages, inspect content, and interact with controls. The main assistant needs enough of that information to make decisions, but a running transcript of every browser detail can become a distraction.

BrowserHelm explores a gateway between the assistant and browser automation.

Give each part a job

The main assistant holds the task and makes decisions. Browser automation performs actions on websites. A gateway can manage how those two parts communicate: compacting browser context, routing routine navigation, and bringing relevant information back.

That boundary is the interesting engineering problem. Too much detail creates overhead. Too little detail makes the assistant guess.

Make reuse deliberate

BrowserHelm also explores reusing task experiences through memory. A previous workflow is useful context, but a website can change. Reuse still needs to account for the current page and the current task.

An ongoing experiment

This note introduces the design direction. It does not make a performance claim. For implementation details and current progress, see the BrowserHelm repository.