Same instinct locally with git worktrees: one agent per worktree so they do not thrash the same checkout. Cloud VMs buy isolation and scale; worktrees buy cheap isolation when you already have a machine. I end up using both patterns depending on whether the bottleneck is compute or just not stepping on each other.

Agreed. Per-thread VMs are quite similar to how local agents use worktrees to avoid cross contamination, but with the added benefit of being able to easily scale up/down compute requirements on demand.