AI agents write and execute code, manipulate user interfaces, and transact on their users' behalf. We ask what forms failure takes in human-agent interaction, and how those forms change as agents act with greater autonomy. From 695 documented incidents in two public repositories, we developed 12 failure modes bottom-up within three interaction stages adapted from Amershi et al.'s guidelines, coding the autonomy at which each arose. What a person can do when an agent fails depends on how much the agent completes between points of user involvement--a property fixed at deployment. Therefore, the same model on the same task can fail differently under different configurations. All 12 modes appear at both low and high autonomy, and rising autonomy changes the form each failure takes. Users need controls between full delegation and full manual work, and neither incident records nor agent benchmarks reach the interaction where these failures occur.

翻译：暂无翻译