Chinese
Uncategorized

GKE Agent Sandbox and Agent Gateway: The Infrastructure Split I Would Make

GKE’s new Agent Sandbox and Agent Gateway features expose a useful split between isolated execution and traffic policy.

Jacky Wang 4 min read 45 views

When I run a long-lived workload on Kubernetes, the dangerous shortcut is to put everything in one Pod: session state, untrusted tool execution, and inbound traffic. It works during a demo. Then the agent waits for a model turn, a human approval, or an external API, while the Pod keeps its resources. Google Cloud’s recent GKE updates make a more useful split visible: Agent Sandbox for execution and Agent Gateway for traffic policy.

The GKE release notes say Agent Sandbox on GKE became generally available on November 12, 2026. It is aimed at isolated, stateful, single-replica workloads such as AI agents. The same update lists Agent Gateway for GKE as a Preview feature in the Rapid channel, providing a central management layer for traffic to, from, and between agents. Google’s announcement about the open agent infrastructure stack also says Agent Substrate 1.0 is generally available on GKE. Its design snapshots idle agent sessions to storage and restores them when computation is needed. The GKE release notes and Google Cloud announcement are the primary references.

Execution and traffic are different layers

Agent Sandbox answers: where does this agent run, how is it isolated, and how does it resume? Agent Gateway answers: who can reach it, what can it reach, and how do we observe or control that traffic? Putting both responsibilities inside an application container makes incidents harder to understand. A stuck session, a denied network request, and a failed restore become the same vague “agent error.”

I would draw the boundary like this:

client -> Agent Gateway -> agent session
                         -> sandboxed tool execution
                         -> state snapshot / restore

Before moving a real service, I would test model calls, tool calls, and resume behavior separately. An agent may compute for a few seconds and then wait for minutes. Treating it like an always-busy web Pod creates a cost problem and hides the real state machine. Every failure should end in a queryable session state, not an unbounded retry loop.

The published numbers are architectural clues

Google says Agent Substrate can snapshot idle sessions, restore them in under 500 milliseconds, and reach 10x container density compared with traditional approaches. Those numbers are from Google’s release material; I have not reproduced them on my own cluster, so I would not turn them into a generic promise. They do show the intended model: persistent identity and state, with compute activated when the agent actually has work to do.

The security boundary needs the same discipline. Google includes gVisor in the open stack to isolate code generated and executed by agents. That does not mean production credentials belong in the sandbox. I would keep credentials behind an external broker or short-lived authorization service, restrict egress, and record each tool request with a result summary and an approval decision. A sandbox reduces the blast radius; it does not remove the need for authorization.

A small acceptance test is enough to start

I would begin with a research agent that has no production write access. Give it a tiny explicit state model: task ID, last completed step, input hash, and retry count. Let it run, suspend, and resume. Then cut the model network and the tool network separately and inspect whether the resulting state explains what happened. Only after that would I test gateway policy, rate limits, and telemetry.

The Agent Sandbox documentation is the useful reference for installation and resource behavior. The Agent Gateway documentation covers the traffic-control side. My view is that the first infrastructure decision for production agents is not which model to pick. It is whether sessions, execution, network access, and cost can be operated as separate systems with separate failure modes.

I would not accept a green Pod status as proof. The test needs one real suspend and resume, one rejected egress request, one rate-limit response, and a trace that connects a session to its tool calls. If resume loses state, or the gateway sees only aggregate traffic without an agent identity, the platform is not ready to operate even if the demo looks healthy.

发表评论