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

The proposed sandbox cloud uses **microVM isolation**, persistent disk state, and **incremental snapshots** that store changes between checkpoints. A control plane and scheduler place or restore sandboxes across clustered nodes and regions.

For long agent tasks, checkpoint the workspace so node failure, maintenance, or experimentation does not erase hours of work. Make scheduling snapshot-aware: restoring on a node that already holds the required layers reduces transfer and startup work.

Persistence adds storage, lineage, and orchestration complexity. Warm pools consume idle resources, while memory-snapshot startup and hybrid pools trade operational simplicity for lower creation latency; the talk offers design intuition rather than measured fleet results.