Agent Sandbox is a Kubernetes project that makes it easy to run isolated, stateful, singleton workloads on a cluster. It was designed with AI agent runtimes and reinforcement-learning environments in mind, where each workload needs its own stable identity and persistent storage.
Developed under Kubernetes SIG Apps, Agent Sandbox provides a Sandbox Custom Resource Definition (CRD) and controller that give you a declarative, standardized API for a single long-running pod with a stable identity and persistent storage, much like a lightweight single-container VM built from Kubernetes primitives. Rather than implementing isolation itself, it delegates low-level sandboxing to secure runtimes such as gVisor or Kata Containers by configuring pods with the appropriate RuntimeClass. This fills a gap between stateless Deployments and numbered StatefulSets for workloads that are fundamentally one-of-one. The project is open source under the Apache-2.0 license, has over 4,300 stars on GitHub, and is under active development, with updates continuing through 2026.
Agent Sandbox is free and open source under the Apache-2.0 license. You run it on your own Kubernetes infrastructure; there is no commercial edition or hosted offering.
Platform engineers and infrastructure teams who run AI agents, code-execution sandboxes, or RL environments on Kubernetes and need safe, stateful isolation per workload.