Fly.io engineers on building agents that don't break themselves
Building Agents That Don't Break Themselves
Fly.io's Daniel Botha argues for splitting an agent's 'brain' from its 'hands': keep the agent loop on a normal server, and run every risky shell command inside a disposable Sprite sandbox. Each user session gets its own Sprite, torn down when done, so a bad rm -rf can't hurt the agent itself. They also inject user tokens only for the duration of a single command—the sandbox never stores the credential at rest. The post walks through SpriteDoc, an internal troubleshooting agent, and a terminal backend for Nous Research's Hermes Agent as real examples.
Why it matters: A Fly.io engineer separates the agent loop from the command sandbox, so a bad agent action can't wreck the environment it runs in.