Agent Sandbox by Kubernetes SIGs

Agent Sandbox is an open-source Kubernetes project under SIG Apps that provides a declarative, standardized API for managing long-running, stateful, singleton container workloads. It introduces a Sandbox CRD and controller that fills the gap between Kubernetes Deployments (stateless, replicated) and StatefulSets, offering a lightweight, single-container VM-like experience built on Kubernetes primitives. The project is particularly well-suited for AI agent runtimes, development environments, Jupyter notebooks, and other stateful single-pod services that require stable identity and persistent storage.

• Stable Identity: Each Sandbox gets a stable hostname and network identity, making it easy to address and connect to running workloads consistently.
• Persistent Storage: Sandboxes support persistent storage that survives pod restarts, enabling stateful workflows without manual PVC management.
• Lifecycle Management: The controller handles creation, scheduled deletion, pausing, and resuming of Sandbox pods automatically.
• SandboxTemplate: Define reusable templates for creating Sandboxes, simplifying management of large numbers of similar workloads.
• SandboxClaim: Allows users to request Sandboxes from a template, abstracting away underlying configuration details.
• SandboxWarmPool: Manages a pool of pre-warmed Sandboxes for rapid allocation, reducing cold-start latency for new workloads.
• Strong Isolation: Supports runtimes like gVisor and Kata Containers for enhanced security, kernel isolation, and network isolation — critical for running untrusted or LLM-generated code.
• Python SDK: A high-level Python client library enables programmatic creation and management of Sandboxes for agent and automation workflows.
• Getting Started: Install via kubectl apply using a versioned release manifest from GitHub, then define a Sandbox resource with a pod template to launch your first isolated workload.
• Community-Driven: Developed under the Kubernetes SIG Apps umbrella with an active Slack channel, mailing list, and open contribution model.