Agent Substrate suspend and resume on Kubernetes

The author details a suspend and resume pattern for AI agents using Agent Substrate on Kubernetes, which checkpoints actor memory and filesystems to object storage to free up worker slots. This approach achieves significant resource density improvements by time-multiplexing workers and avoiding round-the-clock resource reservations for idle agents. The piece outlines key architectural tradeoffs, including restore latency, resume races, and constraints around secrets, GPUs, and network connections. Open article →