<p>Kubernetes SIG Apps is building <a href="https://github.com/kubernetes-sigs/agent-sandbox" target="_blank" rel="noopener nofollow">Agent Sandbox</a>, a new project that adds a <code>Sandbox</code> Custom Resource Definition specifically for running AI agent workloads. Version 0.21 is out now.</p>
<h2>Why not just use StatefulSets?</h2>
<p>The existing Kubernetes primitives — StatefulSets, headless Services, PersistentVolumeClaims — weren’t designed with agentic workloads in mind. Agents are stateful, need a persistent identity, and often need to start near-instantly. Stitching together existing resources to meet those requirements is awkward at best.</p>
<p>The <code>Sandbox</code> CRD treats a single agent instance as a first-class object: a stateful singleton with its own lifecycle, network identity, and isolation guarantees.</p>
<h2>What the Sandbox resource provides</h2>
<ul>
<li><strong>Strong isolation</strong> via gVisor or Kata Containers, useful when agents execute untrusted or generated code</li>
<li><strong>Stable network identity</strong> — each sandbox gets a dedicated hostname and address, which makes multi-agent coordination straightforward</li>
<li><strong>Scale-to-zero with state preservation</strong> — sandboxes can be paused and resumed without losing state</li>
<li><strong>Declarative lifecycle management</strong> through a clean API rather than a pile of coordinated resources</li>
</ul>
<h2>Eliminating cold starts</h2>
<p>The Extensions API adds a <code>SandboxWarmPool</code> resource that keeps a pool of pre-provisioned pods ready to go. When a new sandbox is requested, it picks up a warm pod instead of waiting for one to spin up from scratch.</p>
<h2>Getting started</h2>
<p>The project is installable via <code>kubectl</code>. A Python SDK is available along with example templates. The <a href="https://github.com/kubernetes-sigs/agent-sandbox" target="_blank" rel="noopener nofollow">GitHub repo</a> has the full documentation.</p>


Collections

Kubernetes SIG Apps has released Agent Sandbox (v0.21), a new project introducing a `Sandbox` Custom Resource Definition (CRD) designed specifically for AI agent workloads. Unlike existing primitives like StatefulSets, the Sandbox CRD treats each agent instance as a first-class object with its own lifecycle, stable network identity, and strong isolation via gVisor or Kata Containers. Key features include scale-to-zero with state preservation, declarative lifecycle management, and a `SandboxWarmPool` resource that pre-provisions pods to eliminate cold starts. A Python SDK and example templates are available, installable via kubectl.

Kubernetes SIG Apps introduces Agent Sandbox, a new CRD for running AI agents