GKE Agent Sandbox is built on the open-source Kubernetes SIGs Agent Sandbox project and provides secure, disposable execution environments designed for AI agents that generate and run untrusted code on Kubernetes. By pairing gVisor or microkvm kernel-level isolation with warm pools (SandboxWarmPool and SandboxClaim) for low-latency startup, it offers a hardened runtime platform for single-replica agent workloads. The example demonstrates end-to-end setup, managing secrets with Secret Manager CSI and Projected Volumes, and configuring SPIFFE-based mTLS using Managed Workload Identity (MWLID).
See the full guide and example manifests for GKE Agent Sandbox on GitHub.