Safety Signals as First-Class Architecture: Embodied Fear and Social Modulation in a Persistent AI Agent

Agent safety requires detecting both threats and safety indicators. A robust homeostatic architecture must also ask: what makes this okay? What reduces uncertainty? What signals safety? This paper presents Potato, an AI agent implementing a fear state based on GPS distance, movement speed, accelerometer data, and idle time, with a social modulation layer where trusted faces reduce computed fear. The architecture extends to mobile robotics in ROS2, demonstrating that fear-based reinforcement learning signals generate emergent terrain preferences without explicit programming.