Automating Bug Fixes with Multi-LLM Agent Clusters: Cheaper Than You Think

This post details a novel approach to automated bug fixing using multiple large language models (LLMs). By integrating Asana, the Aider coding agent, and a Sublayer agent, the system automatically triggers three LLMs (GPT-4o, Claude 3.5 Sonnet, and Gemini 2.0 Flash) to attempt fixing the same bug. Each attempt runs in a separate Git branch, resulting in multiple pull requests. This 'wasteful inference' approach proves surprisingly cheap and efficient, offering redundancy and diverse solutions. Even if one model fails, others might succeed, providing alternative approaches. This experiment showcases the potential of this multi-model, automated, low-cost bug fixing, hinting at a paradigm shift in future development.