Agent Substrate avoids Kubernetes GPU snapshots

The author argues that GPU memory should be kept outside Agent Substrate snapshots on Kubernetes because GPU state cannot be serialized by runtimes like gVisor without breaking processes or losing density. Instead, agent sandboxes should be paired with a separate inference pool featuring a queue to handle model calls. Additionally, the author warns that during wake-up storms, the bottleneck is often the model generator queue rather than the worker pool, risking ghost occupancy and delayed responses. Open article →