CNCF: Is the Kubernetes Pod the right unit for AI agents?

A Pod is not designed for ephemeral agents

A Pod — the smallest deployable unit in Kubernetes, containing one or more containers that share a network and storage — was designed for long-running services, not for short-lived tasks. Lin Sun from Solo.io explores this limitation in a new CNCF blog post published on July 14, 2026.

AI agents behave completely differently from web services: they wake up only when they receive a task, do the work in a timespan ranging from a second to a few minutes, and then go idle again. A dedicated Pod running between tasks consumes CPU and memory to no purpose, which quickly becomes expensive in large clusters.

What does the “agent-substrate” model offer?

Sun proposes a new control plane called “agent-substrate” that uses Actors deployed on Workers organized into a WorkerPool — a functional parallel to NodePool for nodes — instead of Pods. The key advantage: a single Worker can host a larger number of suspended Actors simultaneously.

Demonstrated in practice: 3 Workers are sufficient to manage significantly more logical agents than a strict one Pod = one agent model allows.

Open questions for the community

The proposal opens up a number of unresolved challenges: how to handle multi-tenancy when multiple clients share the same WorkerPool, how to set quotas and billing per agent, and how to ensure observability and auditing when an agent’s identity is no longer tied 1:1 to a Pod. The CNCF community is invited to discuss and contribute.