https://smolmachines.com/blog Where the smol machines team shares what we learn building it, and how we think about computers, AI agents and the infrastructure they run on. en Mon, 19 Oct 2026 00:00:00 GMT https://smolmachines.com/blog/agent-evals-on-a-laptop https://smolmachines.com/blog/agent-evals-on-a-laptop Mon, 19 Oct 2026 00:00:00 GMT Engineering 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. https://smolmachines.com/blog/undo-button-for-ai https://smolmachines.com/blog/undo-button-for-ai Thu, 24 Sep 2026 00:00:00 GMT Engineering AI agents work on real computers, and computers have no undo. smol machines checkpoints an agent's whole machine, so you can undo a bad step, go back to any earlier point, or branch to try several paths at once. https://smolmachines.com/blog/gpu-over-vsock https://smolmachines.com/blog/gpu-over-vsock Tue, 14 Jul 2026 00:00:00 GMT Engineering A smolvm guest runs CUDA workloads without a passed-through GPU or an NVIDIA driver: it carries drop-in shim libraries that forward CUDA calls over vsock to a host that owns the GPU. How API remoting works, and how it compares to passthrough, vGPU, and container device injection. https://smolmachines.com/blog/kubernetes-in-a-microvm https://smolmachines.com/blog/kubernetes-in-a-microvm Tue, 14 Jul 2026 00:00:00 GMT Engineering The newest release runs k3d inside a single machine. Kubernetes wants its own kernel, cgroup hierarchy, container runtime, and a specific set of device nodes. Here is why a microVM can host it where a container cannot, and the /dev/kmsg node the kubelet refused to start without.