Stopping your laptop from killing your AI agents mid-task is a

The core problem with most current agent setups is that they're too ephemeral. If you're running a coding agent on a massive codebase or doing some RL training, you can't just rely on a local session. You run into RAM saturation, security risks (running --yolo on your own machine is basically asking for a credential leak), and the obvious issue where the process dies the second your WiFi blips.

What's interesting here is that they aren't doing containers or sandboxes—they're using full KVM virtual machines. This means you get actual kernel-level access and real GPU drivers without some weird syscall-interception layer slowing things down.

The technical breakdown

If you're looking for a practical tutorial on how this fits into an AI workflow, it's basically a one-command setup. You can spin up a box with:

machine0 new mybox

This gives the agent a static IP and an HTTPS endpoint. From there, the agent can essentially self-serve. If you give a tool like Claude Code or an MCP server access to this CLI, the agent can manage its own infrastructure—spinning up a build box, snapshotting the state, and tearing it down when the PR is ready.

• Hardware Range: Starts at 1 vCPU / 1 GB RAM ($0.013/hr) and scales up to 60 vCPU / 240 GB RAM.
• GPU Support: Everything from RTX 4000 Ada up to 8×H200s.
• Persistence: Block storage from 10 GB to 16 TB.
• Uptime: 99.99% VM-level reliability.

Why this matters for LLM agents

The real value here is the "Profiles" feature. You can bundle credentials, MCP connections, and env vars into a profile that gets injected at creation. This creates a hard security boundary; the agent gets exactly what it needs to finish the job and nothing more.

I've seen some wild use cases for this already. Some teams are using a "pilot agent" to scope out a project, which then delegates sub-tasks to a fleet of worker agents, each living on its own dedicated VM. Others are using it for model optimization, keeping a suspended H100 on standby and letting an agent grind on inference speeds overnight.

For anyone building a real-world AI workflow, moving away from "local-only" execution is the only way to scale. Using NixOS flakes or Ansible playbooks on these VMs ensures that the environment is reproducible, which is usually the biggest headache when moving agents from a dev environment to production.