The author shares their evolving mental model for understanding the risks associated with AI agents, drawing from two influential frameworks and their own experience over the past two years. This post aims to contribute to ongoing discussions in the AI safety community.