Where we share what we learn building smol machines, and how we think about computers, AI agents and the infrastructure they run on.
Harbor, the harness behind Terminal-Bench 2.0, can run every trial in its own VM embedded in Harbor's process. We measured it against Apple's container tool on one MacBook: 1.3 seconds per trial against 21.
AI agents work on real computers, and computers have no undo. smol machines checkpoints an agent's whole machine, so it can undo a bad step, go back to any earlier point or try several paths at once.
Recent merges give a smolvm guest a GPU without passing one through: the guest carries no driver, only drop-in CUDA shim libraries that forward calls over vsock to a host that owns the GPU. Here is how API remoting works and how it stacks up against passthrough, vGPU, and container device injection.
The newest release runs k3d inside a single machine. Kubernetes is one of the most demanding things you can ask a Linux environment to host. Here is why a microVM can and a container cannot, and the one device node the kubelet refused to start without.