From fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAI

Securely executing untrusted code requires robust sandbox infrastructure to protect host systems from potential exploits while maintaining high performance. AI agents, when granted code execution capabilities, effectively solve verifiable problems like math and programming, but this necessitates isolated environments. While containers and gVisor offer varying degrees of protection, hardware-level virtualization via microVMs provides the most secure boundary by isolating the guest kernel from the host. Beyond runtime security, persistent storage is a critical unlock; incremental disk snapshotting allows agents to maintain state across long-running tasks, enabling complex operations like Monte Carlo tree searches and rapid recovery from infrastructure failures. Efficient orchestration further relies on leveraging snapshot lineages and memory-based restoration to minimize latency, ensuring that agents remain reliable and scalable in production environments.