code refactoring

Comparing Fable and 10 other LLMs on refactoring a LangGraph god node

a month ago

The article describes an experiment where 11 LLMs (including Fable-5, GPT models, and others) were tasked with proposing refactoring solutions for a 'god node' in a LangGraph agent, followed by cross-evaluations of each other's proposals.
The 'god node' (plan node) contained about 350 lines of hidden logic, making the graph hard to debug, test, and change. The goal was to lift this logic to the graph level for clarity.
In stage one, each model generated a proposal to split the plan node. Proposals varied in granularity, from balanced pipelines (e.g., Fable-5's 5-stage split) to more coarse or detailed approaches.
In stage two, models evaluated all proposals. Reviews differed in thoroughness: Fable-5's review was meticulous with bug findings, while others like Gemini-3.1-pro were minimalistic.