All posts
Engineering
8 min read
Snapshots as a safety net for autonomous code
Letting a model run code it wrote is only reckless if you can’t undo it. We made forking a running machine a sub‑second operation so that every risky step has a checkpoint.
Fork first, ask questions later
Every Kernal machine can be forked in under a second. The fork shares blocks with its parent until one of them writes, so the cost is close to zero. We use this everywhere: before a migration, before a dependency upgrade, before an agent runs a script it authored thirty seconds ago.
If the step goes wrong you discard the fork. If it goes right you promote it. The agent never has to be careful, which is the point: careful agents are slow agents.
Implementation notes
Snapshots are consistent at the block layer and include memory when the machine is running. Restores are lazy, so a 40 GB disk is usable within milliseconds and pages in on demand.
