Chinese
Uncategorized

Agent Substrate on GKE: Agent Sandboxes Are Becoming Infrastructure

Google has donated Agent Substrate to the CNCF and made version 1.0 generally available on GKE. The operational question is how agent sessions pause, resume, and stay isolated.

Jacky Wang 4 min read 103 views

Last week I was looking at a very ordinary Kubernetes question: what happens when a coding agent works for ten seconds, then waits for a model turn, an external API, or a human approval for several minutes? Why should a Pod keep its memory reservation during the entire pause? The execution pattern is different from a web request, and the old “one workload, one live container” mental model starts to look wasteful.

On November 5, Google Cloud announced that it was donating Agent Substrate, gVisor, and OpenRL to the Cloud Native Computing Foundation. The same announcement says Agent Substrate 1.0 is generally available on GKE, while the open-source project can run on other Kubernetes clusters. I read that as a fairly concrete infrastructure change: an agent session’s identity and state no longer have to be tied to continuously allocated compute.

The useful unit is a resumable session

Google says an idle agent can be snapshotted to storage and restored in under 500 milliseconds. It also describes a target of 10x the density of standard container runtimes. I would not copy either number into a capacity plan without testing. Image size, hardware, network setup, and the amount of state being restored will matter. Still, the direction is clear. During the model’s thinking phase, an agent does not need an active machine.

For a SaaS platform, I would model a session with explicit state: a task ID, workspace snapshot, agent identity, current tool call, and reason for waiting. On resume, the system should continue from a verifiable checkpoint rather than asking the model to reconstruct what happened from a chat transcript. That distinction becomes important when a tool call is retried or a customer sees a partial result.

The platform abstraction I would want looks roughly like this:

apiVersion: sandbox.agents.example/v1
kind: AgentSession
spec:
  suspendAfter: 30s
  workspace: snapshot
  network:
    egress: [api.example.com]
  credentials: injected-outside-sandbox

That is a boundary sketch, not an official Agent Substrate API. The important parts are the policy decisions. The platform should own suspension, egress allowlists, and credential delivery. Individual agents should not be trusted to interpret their own security perimeter.

Isolation has to survive generated code

Agents can write and execute code, install packages, and call tools that were not known when the session was created. Google places gVisor next to Agent Substrate because ordinary containers share the host Linux kernel. gVisor adds a userspace application kernel to create a stronger boundary while keeping a container-like operating model.

That does not mean I would hand a sandbox a production secret. Egress should go through a policy enforcement point. Credentials should be injected outside the sandbox, and logs should connect every tool call and denial to an agent identity. A malicious session that can be resumed is more interesting to an auditor than a malicious process that simply crashed.

CNCF is useful, but it is not a warranty

Moving the projects into the CNCF Sandbox gives the work a more neutral community home. It may also reduce the fear that an agent runtime is tied to one cloud. Google says Agent Substrate can run on any Kubernetes cluster and across CPUs, GPUs, and TPUs. I would treat that as an architectural promise to test, not as a universal performance guarantee.

My first experiment would use 20 sessions: half actively calling tools, half waiting, with deliberate restore failures and blocked network requests. I would record snapshot size, resume latency, identity continuity, and cleanup behavior. The first agent-infrastructure bill is often not the model bill. It is the fleet of processes that stayed awake while doing nothing.

The primary references are Google Cloud’s CNCF announcement and the Agent Substrate repository. I would use them to shape an experiment, not to turn a vendor target into an acceptance test.

There is also a product-design question here. A suspended session still owns context, customer data, and a network policy. Restoring it should not silently refresh permissions or inherit a new service account. I would make identity expiry and snapshot encryption part of the same test, because a fast resume that crosses an authorization boundary is not an operational win.

发表评论