Generative AI is evolving from stateless function calls to a system of continuously running, coordinated AI agents. These agents require persistent context, tool usage, code execution, and inter-agent communication over extended periods. Kubernetes is the ideal infrastructure for these workloads, but traditional primitives don't perfectly fit the needs of these stateful, singleton agents. The new Kubernetes Agent Sandbox project, under development by SIG Apps, aims to bridge this gap. It introduces a custom resource definition (CRD) for managing AI agent runtimes. The Sandbox CRD provides strong isolation for untrusted code execution using runtimes like gVisor or Kata Containers. It also offers robust lifecycle management, allowing agents to scale down to zero when idle and resume instantly. Furthermore, each Sandbox is assigned a stable identity for seamless inter-agent communication. To accelerate development in the fast-moving AI space, an Extensions API layer is available. The SandboxWarmPool extension eliminates cold starts by maintaining a pool of pre-provisioned Sandbox pods ready for immediate use. Users can install the core and extension components into their Kubernetes clusters. The Agent Sandbox project is open source and invites community involvement for building the future of cloud-native AI agents.