smolvm lets agents checkpoint a running VM and branch it

smolvm does branchable computing for virtual machines. Agent sandboxes typically retry a failed action by rebuilding a container, losing whatever was in memory. Forking a live VM keeps that state, so an agent can run several attempts from the identical starting point and keep only the one that worked.

A branch takes an in-memory checkpoint of a running machine and continues from it. A full checkpoint writes that state to a durable .smolcheckpoint file you can restore elsewhere, per smolvm's GitHub repo. Built on libkrun and released under Apache-2.0, the project has picked up 6,155 stars.

Containers approximate this by tearing one down and spinning up a fresh one between attempts. That resets the filesystem, not the state sitting in memory when the agent was mid-task. A partially loaded model, an open database connection, or a half-written buffer doesn't survive a restart.

I haven't run smolvm against a real agent loop yet. I can't say what checkpoint overhead looks like at the frequency an agent would want it, maybe once per tool call. The open question before this replaces container resets as the default pattern is whether snapshotting memory is cheap enough to do constantly. The GitHub page doesn't give latency numbers, so it may fit occasional expensive retries better than routine branching.

Two releases three days apart, v1.16.1 on September 15 and v1.16.2 on September 18, suggests active iteration rather than a one-off drop.