A blue cloud-computing diagram shows Kubernetes orchestrating containerized AI agents across servers.
Microsoft has put an Agent Sandbox cluster extension for Azure Kubernetes Service (AKS) into public preview. It is meant to take the plumbing out of running AI agents that execute untrusted, model-generated code. The Azure Updates entry is dated November 17, 2026, and carries the "In preview" label.

Related Azure CLI pull request commands reference an extension type of Microsoft.AgentSandbox. Microsoft's public material is high-level, though, and it leaves several practical questions open. They are listed below.

A blue cloud-computing diagram shows Kubernetes orchestrating containerized AI agents across servers. What Microsoft announced​

The Azure Updates entry starts with the problem. An agent that runs code produced by a model needs an isolated environment for each session. Teams on Kubernetes currently build their own sandbox lifecycle, storage and warm-up logic. That adds operational overhead, and users wait while fresh environments initialize.

Microsoft's answer is a managed extension built on the open-source Kubernetes Agent Sandbox project. It is described as providing:

  • Declarative sandbox creation.
  • Sandbox lifecycle management.
  • Pre-warmed pools.

Microsoft also pairs it with the separate Pod Snapshot cluster extension. Together, the two let isolated agent environments be snapshotted and restored while keeping runtime state across the sandbox lifecycle.

Added detail from the Azure platform blog​

A November 17 post by Jeremy Winter, Azure Platform's chief product officer, puts the extension in the context of Microsoft's Ignite infrastructure announcements. The post says agent-generated code runs in managed sandboxes with:

  • VM-level isolation for every session.
  • Pre-warmed pools for fast startup.
  • A serverless option with no infrastructure to manage.

It adds that Azure confidential computing can extend protection to sensitive data and workloads while in use.

These are Microsoft's own claims. The post gives no benchmarks, and no threat model for the isolation.

What the upstream project does​

Independent coverage describes the upstream project this way. The Pixel Robots write-up of the Azure Day at KubeCon NA 2026 characterizes Agent Sandbox as a Kubernetes SIG Apps project that gives you an isolated, stateful pod per sandbox with a warm pool you claim from. The same author notes that if Agent Sandbox becomes a managed extension, that removes a lot of the work of running it yourself. That is the operational case Microsoft is making.

The upstream project documentation also shows how people have been doing this by hand on AKS. One example combines the agent-sandbox extension CRDs with the AKS Pod Sandboxing feature (Kata Containers on Microsoft Hyper-V) to run a fleet of AI agents. That example required registering a preview feature and adding a dedicated Kata-capable node pool. The managed extension looks aimed at removing this sort of hand assembly. Microsoft has not said whether it uses the same underlying isolation mechanism, so treat the Kata example as background, not as a description of the managed service.

How it fits with other announcements​

Other Windows Forum coverage links this extension to its neighbors:

  • Our Pod Snapshot piece notes that the troubleshooting page also mentions an Agent Sandbox cluster extension restore workflow as an alternative path.
  • Agent Grid, a separate preview, lists AKS Agent Sandbox, with isolated execution using Kata Containers and Kubernetes orchestration, as one of its supported environments.

Agent Grid is its own announcement and not part of Agent Sandbox. The overlap does suggest Microsoft sees the sandbox as the execution layer in a larger agent stack.

Management-plane evidence​

Microsoft Learn's AKS managed-cluster reference for API version 2026-08-02-preview includes a cluster-level Agent Sandbox profile with an enabled flag. The flag is described as controlling whether the Agent Sandbox core extension is enabled, and it defaults to false on cluster creation. The same reference lists a matching Pod Snapshot Restore flag.

This shows that the feature exists in the preview API schema. It is not a setup guide, and it does not by itself establish the supported onboarding path for every customer. I haven't written deployment steps because the supplied material doesn't contain verified ones.

What isn't stated​

The announcement leaves out several things admins will want:

  • Supported AKS and Kubernetes versions.
  • Regions.
  • Pricing, including for the serverless option.
  • Prerequisites such as identity requirements.
  • Network defaults and resource limits.
  • Exactly what runtime state a snapshot captures.
  • How pool sizing and warm-up behave.

Practical takeaways​

  1. Treat it as a lab feature. Azure Updates defines "In preview" as available for non-production use and testing. "Launched" is the production-ready label. This item isn't there yet.
  2. Don't treat isolation as a security sign-off. VM-level isolation per session is a strong architectural claim. Agent code still needs egress controls, identity scoping and review, and none of those are covered in the available material.
  3. Evaluate against what you run today. If you already maintain your own sandbox controller, warm pools and snapshot logic, compare the managed extension against that on startup time, cost and operational burden.
  4. Check current documentation before piloting. Preview terms, regions and API versions can change quickly.

Bottom line​

This is a meaningful step for AKS teams running agentic workloads. Microsoft is taking an open-source Kubernetes project and offering it as a managed extension with pre-warmed pools and snapshot integration. The detail that matters for production decisions is still missing. For now it is a preview to test, not a platform to build on.

 

References

  1. (In preview) Public Preview: Agent Sandbox cluster extension Azure Updates 2026-11-17T20:15:03Z
  2. (aks-preview) Add direct Agent Sandbox extension commands by alimaazamat · Pull Request #10517 · Azure/azure-cli-extensions · GitHub github.com
  3. Your Privacy Choices Opt-Out Icon azure.microsoft.com