IDEA: Autonomous Web Interaction and Full-Stack Synthesis Using Large Language Models
Year: 2026
Type: article
Abstract: This paper presents IDEA (WebPilot AI), a framework combining autonomous web interaction with full-stack software synthesis using large language models. It addresses challenges in accessing up-to-date information from dynamic websites and generating multi-file applications reliably. The paper describes methods for reducing web-page context, verifying generated code in isolated environments, and evaluating web agents across benchmarks and platforms. It also examines security risks, including indirect prompt injection and server-side request forgery, and outlines a hybrid model-routing approach.
Institution: MIT Art, Design and Technology University
Language: English
Topic: Web Data Mining and Analysis
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236104312
