Kubernetes Agent Sandbox is a Kubernetes SIGs project that gives AI coding agents kernel-isolated, disposable environments as first-class Kubernetes resources using a `Sandbox` CRD backed by gVisor or Kata Containers. The post explains why standard containers aren't sufficient security boundaries for AI agents (shared kernel, CVE-2019-5736), how gVisor's userspace kernel mitigates this, and how Agent Sandbox addresses the cold-start latency problem with warm pools and memory snapshots. A step-by-step Pulumi deployment on GKE is provided covering: creating a gVisor node pool, installing the Agent Sandbox controller, programmatically creating per-developer sandboxes, setting egress NetworkPolicies to block private IPs and metadata servers, and enabling private access via Tailscale with per-user ACLs.
