Agent Sandbox is a Kubernetes operator that manages isolated, stateful, singleton workloads ideal for AI agent runtimes and reinforcement learning applications.
The tool addresses the mismatch between workloads that require long-running, stateful containers with stable identities and Kubernetes' native abstractions like Deployments and StatefulSets, which are designed for stateless or numbered-replica patterns. It introduces a Sandbox Custom Resource Definition and controller that provides a declarative API for managing single pods with stable hostnames, network identities, and persistent storage that survives restarts. The controller handles lifecycle management including creation, scheduled deletion, pausing, and resuming. The project delegates low-level container isolation to secure sandbox runtimes like gVisor or Kata Containers by managing pods configured with appropriate RuntimeClass settings.
Adoption makes sense for teams running AI agents or reinforcement learning workloads on Kubernetes that need persistent state and stable identity without the complexity of StatefulSets. The tool includes extensions beyond the core Sandbox CRD: SandboxTemplate for defining reusable configurations, SandboxClaim for allocating sandboxes from a pool, and SandboxWarmPool for maintaining pre-warmed instances to reduce startup time. A Go SDK is available for programmatic interaction. Installation can be selective, allowing users to deploy only the core components or include extensions as needed.
The project shows substantial real-world adoption, with most open issues coming from external users rather than the core team. Maintainers respond to new issues and pull requests within a day. Work in the issue tracker centers on feature requests, bug reports, and flaky test fixes.
