Agent Sandbox: Kubernetes Project Brings Isolated, Stateful Workload Management to AI Agents
Key Takeaways
- ▸Kubernetes SIG Apps introduces Agent Sandbox CRD for managing stateful, isolated AI agent workloads with stable identities and persistent storage—filling a gap for workloads that don't fit Deployments or StatefulSets
- ▸Agent Sandbox leverages existing sandbox runtimes (gVisor, Kata Containers) for security isolation while providing a declarative, Kubernetes-native API that abstracts orchestration complexity
- ▸The project includes extensions like SandboxWarmPool and SandboxTemplate to enable efficient resource pooling and template reuse for managing large fleets of sandboxed workloads
Summary
Agent Sandbox is a new Kubernetes Custom Resource Definition (CRD) and controller being developed under the Kubernetes SIG Apps umbrella to manage isolated, stateful workloads ideal for AI agent runtimes. The project provides a declarative API for managing long-running, stateful, singleton containers with stable identities and persistent storage—essentially offering a lightweight, single-container VM experience built on Kubernetes primitives. Rather than reinventing low-level container isolation, Agent Sandbox delegates to secure sandbox runtimes like gVisor or Kata Containers via Kubernetes' RuntimeClass feature.
The core Sandbox CRD enables stable hostname and network identity, persistent storage, and comprehensive lifecycle management (creation, deletion, pausing, resuming) for individual workloads. An extensions module builds on this foundation with additional constructs including SandboxTemplate for reusable templates, SandboxClaim for allocating Sandboxes from a pool, and SandboxWarmPool for pre-warming instances to reduce startup latency. The project includes both a standard Kubernetes controller implementation and a Go SDK for programmatic interaction, with installation available via kubectl apply or kustomize.
