AI agent sandboxes

Give an AI agent a real Linux machine to work in. The agent can install packages, run code it wrote and change files, all inside a microVM with its own kernel. If something goes wrong, delete the microVM and take a fresh one.

Running agent-generated code in a container on a shared host means trusting a shared kernel. A microVM puts a hardware-backed virtualization boundary between the agent’s code and your host.

This is a good fit when you want to:

• Let agents run untrusted or generated code safely.
• Give every agent session, or every task, its own clean machine.
• Keep sandboxes on your own hardware, near your data.

How it works

• Battery keeps a warm pool of pre-booted microVMs on your Flintlock hosts. Your agent platform claims a sandbox from the pool, uses it, and releases it. Battery then replaces it with a fresh one.
• guest-agent runs inside each microVM. The host uses it to run commands, copy files in and out, and open SSH sessions over vsock. The sandbox does not need a network connection to be controlled.
• Each sandbox’s kernel and root file system come from container images, so you can build sandbox images with the tools and libraries your agents need.