Running AI agents safely requires separating where the agent lives from where it executes code. Using disposable sandboxes (called Sprites on Fly.io) for command execution protects the agent's host environment from destructive operations. Two real-world examples illustrate this: SpriteDoc spins up a fresh sandbox per user session with ephemeral credential injection so tokens are never stored at rest, while Hermes Agent keeps a persistent sandbox per task to preserve installed state between runs. Both approaches allow agents to skip manual approval prompts for dangerous commands since the sandbox acts as the security boundary. Checkpointing sandbox state before risky operations enables fast rollback in seconds using copy-on-write snapshots, making unattended agent execution practical.
