Why we chose LangGraph to build our coding agent

Qodo has been developing AI coding assistants since the era of GPT-3, initially using structured workflows for tasks such as test generation and code reviews, which were effective with older models despite their limitations. The release of more advanced models like Claude Sonnet 3.5 enabled Qodo to create a more dynamic and flexible system while maintaining high code quality standards. To achieve this, they chose LangGraph, a framework that allowed for adaptable and opinionated workflows through a graph-based approach, which balances flexibility with structure by defining nodes and edges representing discrete steps and transitions. LangGraph's flexibility permits easy recalibration for new models, while its API simplifies the coding process by making workflows almost self-documenting. The framework's node-based architecture promotes reusability across different workflows, and its built-in state management enhances persistence and functionality. However, challenges remain, such as incomplete documentation and the difficulty of testing LLM-driven systems, which has led to a reliance on manual testing. Despite these issues, LangGraph has been a valuable tool for Qodo in developing adaptive coding assistants.