Autonomous agents / 2024—
Agent-E.
A hierarchical web agent for complex, multi-step browser tasks. Agent-E combines planning and execution, filters noisy page structure, and tracks changes between actions; its paper reports 10–30% stronger performance than prior state-of-the-art agents across most WebVoyager categories.
The work
The paper identifies hierarchical planning, DOM distillation, and change observation as useful design principles for web agents. Agent-E achieved 73.2% on WebVoyager, state of the art at its 2024 release.
My role: Co-author of the Agent-E architecture and evaluation paper.
Tools & themes
Python · Playwright · LLM · GPT-4