GKE Agent Sandbox

https://docs.cloud.google.com/kubernetes-engine/docs/how-to/agent-sandbox

GKE Agent Sandbox is a managed GKE add-on, built on the open-source kubernetes-sigs/agent-sandbox project, that runs untrusted AI-generated code in isolated, stateful, single-replica Pods with persistent storage and kernel-level isolation via gVisor. It is used by AI agent runtimes, code interpreters, and development environments that need to execute code safely.

Pricing model No additional charge for Agent Sandbox; standard GKE resource pricing (per-second, vCPU/memory/storage) applies
Pricing The headline compute price as the provider states it. Sorting ranks prices stated per vCPU over time, converted to hourly. Other pricing models sort after, unranked. No Agent Sandbox surcharge; standard GKE vCPU/memory/storage rates apply (see GKE pricing)
Free while idle Whether a stopped, paused, or sleeping sandbox costs nothing.
Elastic Resources flex with what the sandbox actually uses, and the bill follows: yes when the provider bills on active or observed use, no when capacity is reserved or allocated and billed while it runs.
Memory snapshots Whether RAM state survives a pause and resume, not just disk.
Wake on request Whether a stopped sandbox wakes automatically on inbound traffic.
Start / resume Typical time from create or resume to running. Sorting ranks stated times, converted to seconds; claims with no figure sort after, unranked. ~200 ms typical (90% of allocations under 200 ms; up to 300 sandboxes/sec/cluster; pre-warmed pods enable sub-second creation)
Max runtime The longest a sandbox may run. Configurable TTL (no fixed maximum documented)
Isolation The isolation technology between sandboxes. gVisor (also supports Kata Containers and other OCI-compatible runtimes; runtimeClassName: gvisor)
GPUs Whether GPU instances are available for sandboxes.
Docker Whether Docker containers can run inside the sandbox.
SDKs Python, Go

✓ yes · ✕ no · – no cited public fact.

GKE Agent Sandbox implements a Default Deny network security posture for all sandboxed environments.

Sources

Last verified 2026-08-09. Corrections land through the update log.