Large Language Model-based agents are increasingly capable of autonomously interacting with external environments and completing complex tasks. However, their growing autonomy also introduces safety risks that may not be fully addressed by model-level safeguards or external monitoring alone. In this project, we will explore isolation-first safety management for AI agents, using system-level isolation to limit the impact of compromised or misbehaving agents. Students will investigate the design of isolation mechanisms, build a prototype framework, and evaluate its effectiveness and practical trade-offs in representative agent tasks.