Yeah, I wasn’t expecting it to be a perfect example, and was mostly just really happy to be able to take one that someone else had spent the time/effort on so I didn’t need to do that myself. I do think I probably ought to have given it a bit more of a run-through upfront, and made some tweaks myself as a result.
Yes, I used Multipass VMs and had concierge set up Juju upfront. I had Claude write me a little script that would get all that set up as well as installing the copilot CLI and pushing in the starting prompt and so on.
Presumably this would also work with just plain lxd, and probably actually I could have used spread. But my normal development process is to kick off agents in Multipass VMs so I was just following that pattern.
I did have some memory pressure as a result of this - Copilot can suck up quite a bit of memory while it’s working, and that’s in addition to what Juju and the charms needed. I had originally tried to do a couple in parallel but was getting OOM’d often enough that I gave that up. I also had to bump up the memory allowance for the VM from 8 to 12 GB at some point (Copilot was running inside the VM, for safety).
(At the time, I was using my laptop with a sad 16G of memory – since then my corporate laptop has arrived, and doing this would be much simpler now.)