The test I would apply when a VM goes quiet and shuts down is where the attempt lives. The orchestrator already keeps the state machine, and the workers are ephemeral boxes with a proxy around them, but a resume across machines only works if the coordinator's scratch pad is not inside the box that just went idle. Three subworkers had to stay busy for 48 hours with the laptop closed, and the PM still had the complaints and the park state on Monday. I would keep that scratch pad where a replacement box can read it, not only in the thread that dies with the VM. Otherwise the next box starts from whatever the agent says it built, and the clipped coasters are a bug you cannot point at.
The test I would apply when a VM goes quiet and shuts down is where the attempt lives. The orchestrator already keeps the state machine, and the workers are ephemeral boxes with a proxy around them, but a resume across machines only works if the coordinator's scratch pad is not inside the box that just went idle. Three subworkers had to stay busy for 48 hours with the laptop closed, and the PM still had the complaints and the park state on Monday. I would keep that scratch pad where a replacement box can read it, not only in the thread that dies with the VM. Otherwise the next box starts from whatever the agent says it built, and the clipped coasters are a bug you cannot point at.
Did you use goal or loop for this?
Our project coordinator is like a smart loop. It's a loop with an agent in it
Yea, Claude is not nearly as good at pixel art or sprite sheets and it would've taken significantly longer besides