Running AI agents on Kubernetes raises a fundamental question about whether the Pod is the right deployment, identity, and lifecycle unit for agents. The kagent project initially gave each agent its own Pod for isolation and identity, but agents behave differently from microservices — they're bursty, short-lived, can spawn subagents, and may pause waiting for human approval, making dedicated Pods wasteful. Agent Substrate, introduced by Google alongside Agent Sandbox, proposes a control plane above Kubernetes that schedules logical 'Actors' onto long-lived worker Pods. This lets a fixed pool of Pods support far more logical agents, but shifts identity, access control, observability, and billing away from the Pod level to the Actor and ActorTemplate level — requiring rethinking of how platform engineers reason about agent lifecycle management.
