# Agent Sandbox

> Apache-2.0 Kubernetes Sandbox CRD + controller — warm pools for agent runtimes

**Category:** AI Agents  
**Pricing:** open-source  
**URL:** https://github.com/kubernetes-sigs/agent-sandbox  
**GitHub:** https://github.com/kubernetes-sigs/agent-sandbox  
**Added:** 2026-10-08

---

## The Hook

Agent Sandbox (Kubernetes SIG Apps) adds a `Sandbox` custom resource for long-running, stateful, single-pod workloads with a stable identity — the shape AI agent runtimes and RL environments need — plus warm pools so a new sandbox is ready in seconds.

## What's Good

Apache-2.0 (kubernetes-sigs/agent-sandbox). Release v1.0.5. Install with one `kubectl apply` of the release manifest. Core `Sandbox` CRD handles lifecycle (create, scheduled delete, pause, resume) with persistent storage; extensions add `SandboxTemplate`, `SandboxClaim`, and `SandboxWarmPool` for pre-warmed allocation. Upstream Kubernetes project, not a vendor SaaS. ★~4.2k.

## The Catch

It schedules and manages sandboxes; it does not isolate them. Isolation comes from the RuntimeClass you pair it with — gVisor (see the gvisor gem) or Kata Containers — and without one you just have ordinary pods. Assumes you already run and operate a Kubernetes cluster. Vs e2b (hosted sandbox API with an SDK): Agent Sandbox is the self-hosted orchestration layer on your own cluster, with no hosted control plane.

## Verdict

Apache-2.0 sandbox orchestrator for Kubernetes when agent runtimes need stable, warm, stateful pods — pair it with gVisor or Kata for the actual isolation.