Interaction-Driven Browsing
A human-in-the-loop browser-agent framework and prototypes that infer context, ask for decisions, and execute personalized tasks.
About this project
With Jinkyu Jang, I proposed a conceptual framework for browser-using agents informed by theories of human web browsing. The framework addresses tasks in which users’ goals evolve through exploration, decisions, and changing context. The agent runs actions and asks for user input at detected decision points, distinguishing exploration from exploitation to support control over the breadth and depth of browsing.
Browser prototypes
At Companoid Labs and Herbert Computer, I led research and prototyping of an agentic browser as a TypeScript Chromium extension built on a customized fork of the Eko agent framework.
I implemented context inference from browsing and interaction history, storing inferred preferences and intentions in a vector database. This context supports personalized task recommendations, executed in parallel by agents in background tabs.
The related Generative UI Browser prototype explores how these results can become interactive pages.
Research framework
The 2025 preprint presents the conceptual framework and hypothetical use cases. The prototypes explore its interaction principles in working browser systems.