Interesting direction. One adjacent workflow we've been looking at is cross-environment execution where the agent / dev loop stays local, but GPU access lives elsewhere. In our case the recurring pain isn't only orchestration, it's making an existing remote GPU easy to attach to from a laptop or lab machine without shifting the whole workflow into a remote VM mindset. I'm involved with GPUGo / TensorFusion, so biased, but I think local-first + remote capability is going to matter a lot for small teams and labs. Curious whether you expect most users to want symmetric peer-style composition, or whether local-first control over remote resources ends up being the dominant pattern.

[dead]