Is Agent Code Less Maintainable Than Human Code?

The paper presents CodeThread, a framework for evaluating the maintainability of code produced by coding agents compared to human-written code. Experiments with four coding agents revealed that agent-generated code resulted in a task resolve rate decrease of up to 13.1% when future agents attempted to build upon it, highlighting significant behavioral differences in input validation and error handling. This underscores the importance of assessing maintainability in AI-generated code, as traditional metrics may not capture the underlying issues that lead to increased downstream errors.