In this article, you will learn how to sandbox the code an AI agent writes by running it in a limited subprocess, a hardened Docker container, and a Firecracker microVM, and how to compare their safety and latency.