1 min read
Gemini Managed Agents on Sprites
Alex Ezell
Google launched Gemini Managed Agents yesterday and you can run those agents on a Sprite via the self-hosted environment setup.
Why would I do this?
This lets Gemini run the loop and model leave the shell and files and network and all that other “computer stuff” for the sprite to deal with. Let the loop do the thinking and the computer do whatever the loop tells it to.
Doing that on your work laptop? My grandmother used to say that’s “too much sugar for a nickel” meaning you might have to deal with a lot more than you reckoned for.
Why sprites are a good fit
Nothing listens. The worker dials out. No need for any inbound port juggling.
The VM is the sandbox. The worker confines the agent’s file tools to --root-dir. It does not confine execute_bash. That’s fine if the thing around the shell is a real boundary.
The worker is a service. Sprite services are supervised. If the worker crashes, it restarts.
It sleeps when you want it to. Sprites pause when idle and outbound polling doesn’t count as activity. The script holds a Task open with a five-minute expiry and refreshes it every minute. While the worker runs, the Sprite stays up. If the worst happens and the worker dies, the failure mode is “it goes to sleep,” not “it runs all month.”
The disk is just a disk. Clone the repo. apt install the thing. It’s all still there on the next interaction, and the one after that, because it’s a persistent Linux computer and not a snapshot being rehydrated.
Keep your keys in your pocket. The Sprite only ever holds the environment key in a 0600 file. Your Gemini API key never leaves your laptop.
How do I get started?
Follow our guide here to set it up: https://docs.fly.io/sprites/integrations/gemini-managed-agents or by checking out Google’s own example using Sprites.