A Taxonomy of Agentic Errors: A systematic review of how agentic AI systems can fail

Agentic AI extends generative AI by enabling AI systems to select and take actions in the physical or digital world. With new capabilities and greater autonomy, new errors can occur in agentic AI systems. Addressing these errors requires a more comprehensive understanding of the types of possible errors. To address this gap, we performed a systematic literature review of agentic AI errors. We surveyed 1,379 agentic AI research papers. Of those, we identified 123 papers that discussed errors that occurred in their agentic AI systems. Then, we extracted 652 errors which we qualitatively classified into a formal taxonomy. Our goal is to provide a foundational understanding of agentic AI errors to enable a more systematic approach to identifying and addressing these errors.