Web Agents

Train language models to navigate and interact with web interfaces (browsers, forms, search) autonomously, treating web navigation as a partially-observed sequential decision problem. Web agents treat browser interaction as a sequential decision task. The model observes the web page state (HTML, accessibility tree, or screenshots), selects actions (click, type, scroll, navigate), observes the resulting page state, and continues until the task is complete. Training uses a combination of behavior cloning from human web demonstrations and RL from task completion rewards. Modern web agents use vision-language models to process screenshots directly, eliminating the need for HTML parsing.