k8s-agent-sandbox

Kubernetes Agent Sandbox provides CRD-based isolation for AI agent execution using network policies and runtime sandboxing (gVisor or Kata). Deploy a SandboxTemplate and use SandboxClient to provision and manage isolated Pod environments with automatic lifecycle handling.

Sandbox CRDs (Sandbox, SandboxTemplate, SandboxClaim, SandboxWarmPool) define templates, claims, and warm pools for on-demand sandbox provisioning. SandboxTemplate defines the blueprint; Sandbox represents a runtime instance; SandboxClaim requests access; SandboxWarmPool pre-provisions instances for faster deployment.

Yes. The Python SDK includes SandboxClient for direct sandbox interaction and SandboxRouter for on-demand provisioning and routing. Use SandboxClient(template_name="python-sandbox-template") to programmatically create and manage sandbox instances.

Agent Sandbox supports three connection modes: Tunnel (proxied connection), Gateway (ingress-based access), and Direct (point-to-point connection). Choose based on network topology, security requirements, and latency constraints for your deployment context.

Network policies restrict traffic between sandbox instances and external services. Agent Sandbox applies policy-based isolation automatically through CRD configuration, preventing unauthorized inter-pod communication and enforcing egress/ingress rules at the Kubernetes network layer.

Yes. Agent Sandbox supports both gVisor and Kata Containers as runtime isolation options. Select your preferred runtime when defining the SandboxTemplate to balance between container efficiency and security isolation level.