An empirical study of self-admitted technical debt in AI agents

Abstract

AI agents, characterized by their ability to make dynamic decisions and interact with complex environments (e.g., AutoGPT, graphrag), are reshaping human–AI collaboration paradigms. However, due to tight delivery deadlines and varying levels of developer expertise, developers often introduce technical debt (TD) by taking shortcuts during development. Developers typically annotate such suboptimal design and implementation choices through comments, known as self-admitted technical debt (SATD). If left unaddressed, technical debts can significantly impact the quality of AI agents in terms of reasoning capabilities, execution efficiency, and overall performance. While substantial empirical resea..