MASON - A Multi-Agent LLM Framework for No-Code Development

Abstract

The rise of No-Code Development (NCD) has enabled citizen developers to build applications without traditional programming expertise. However, as these platforms scale to handle complex, interdependent tasks, their limitations become apparent. Large Language Models (LLMs) offer a potential solution, yet single-agent systems often struggle to manage full stack development reliably. This study introduces MASON—a Multi-Agent System (MAS) for Open No-code development framework—that coordinates specialized LLM agents via a YAML-based workflow to automate NCD tasks. The system was evaluated across four proprietary models—Claude 3.5 Sonnet, GPT-4o Mini, Gemini 1.5 Flash, and DeepSeek-Chat—using HumanEval and MBPP benchmarks to assess accuracy, execution time, and stability. MASON configurations showed improved task reliability in simpler workflows but introduced latency on more complex tasks. Additional testing with small, locally hosted LLMs revealed significant limitations, emphasizing the need for architectural redesign or model fine-tuning to support deployment in resource-constrained environments.