Building Agents That Don't Break Themselves

5 hours ago

Building agents is enjoyable, but rebuilding self-destructive ones is less fun. Many Fly users create agents that avoid self-destruction by executing risky actions in Sprites.
Agents should separate their living environment (where they run loops calling APIs) from their execution environment (where they run potentially dangerous code).
SpriteDoc demonstrates a multi-user agent architecture where each session runs in an isolated Sprite, ensuring user commands are sandboxed and credentials are injected only temporarily.
Hermes Agent uses a Sprite backend to run shell commands in a sandbox, allowing dangerous commands without approval prompts, as the sandbox provides security.
Even if an agent runs in a sandbox, its commands should be executed in a separate, disposable sandbox to prevent self-harm and maintain isolation.
Using Sprites allows for cheap, fast, and disposable sandboxes that idle when unused, reducing costs and enabling easy session revival.
Checkpointing in Sprites enables undo functionality, allowing agents to roll back changes after mistakes, such as accidental file deletions, in seconds.
This pattern saves time and enhances security by letting agents operate freely in isolated environments, where worst-case scenarios involve simple restores rather than full recoveries.

Hasty Briefsbeta