Reproduce “FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zimu Lu, Houxing Ren, Yunqiao Yang, Ke Wang, Zhuofan Zong, Mingjie Zhan, Hongsheng Li

FullStack-Agent enables large language model (LLM) agents to generate fully functional full-stack web applications from natural language, moving beyond superficial frontend interfaces. It achieves this through a multi-agent development framework, a self-improving data generation pipeline, and a comprehensive evaluation benchmark, demonstrating superior performance in frontend, backend, and database functionality.

Key insights

  • The multi-agent architecture and specialized debugging tools are crucial for efficiently handling the complexity of full-stack web development, improving error localization and fixing.
  • Learning from high-quality, real-world repositories via back-translation is essential for improving LLM coding capabilities, as directly generated trajectories did not yield similar gains.
  • FullStack-Bench provides a reliable and highly human-aligned evaluation metric for full-stack web applications, moving beyond superficial UI assessments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.