This is an early release preview. You may encounter bugs.
Agent Sandbox logo
Unclaimed

Platform deployment

Agent Sandbox

Easy management of isolated, stateful, singleton workloads, ideal for use cases like AI agent runtimes and reinforcement learning (RL).

A+ 93/100 GitHub score ? This grade is derived from GitHub signals, not user votes. Open for the full breakdown.
Discussion

01 / Summary

About Agent Sandbox.

Agent Sandbox adds a Sandbox Custom Resource Definition (CRD) and controller to Kubernetes, developed under SIG Apps. A Sandbox is a single, long-running, stateful pod with a stable identity and persistent storage, a shape that fits neither the stateless replicas of a Deployment nor the numbered pods of a StatefulSet. Use cases include runtimes that execute untrusted, LLM-generated code for AI agents, high-throughput sandboxes for reinforcement learning and evaluation loops, developer environments, notebook sessions, and single-instance services.

Agent Sandbox is a sandbox orchestrator rather than an isolation layer. It delegates container isolation to secure sandbox runtimes such as gVisor or Kata Containers by creating pods configured for them through a RuntimeClass, and follows the standard Kubernetes controller pattern: you declare a Sandbox resource and the controller manages the pod, its lifecycle, scheduled deletion, and pausing and resuming.

An extensions module builds on the core API. SandboxTemplate defines reusable Sandbox specs, SandboxWarmPool keeps pre-warmed Sandboxes ready to hand out, and SandboxClaim takes a Sandbox from a warm pool without the caller knowing its configuration. Go and Python SDKs consume the API from code, and an optional Sandbox Router reverse-proxies HTTP traffic from SDKs and external clients to sandbox pods where direct port-forwarding is unavailable. Planned characteristics include deep hibernation, automatic resume on network connection, and memory sharing across sandboxes.

  • Sandbox CRD: a declarative API for one stateful pod with a stable hostname and network identity
  • Persistent storage: storage that survives Sandbox restarts
  • Lifecycle management: creation, scheduled deletion, pausing, and resuming handled by the controller
  • Runtime isolation: delegates kernel and network isolation to runtimes such as gVisor or Kata Containers
  • SandboxTemplate: reusable templates for creating many similar Sandboxes
  • SandboxWarmPool: a pool of pre-warmed Sandboxes for low-latency allocation
  • SandboxClaim: claims a Sandbox from a warm pool without exposing its configuration
  • SDKs: Go and Python clients for creating and using Sandboxes programmatically
  • Sandbox Router: an optional HTTP reverse proxy that routes SDK and external traffic to sandbox pods
  • Performance tuning: controller flags, warm-pool refill shaping, multi-pool sharding, and API Priority and Fairness insulation for high-throughput use

02 / Discussion CREDIBILITY-GATED

Discussion

Reading is open to everyone. Posting and voting need a verified identity or a GitHub grade of B or higher.

  • No discussions yet.

04 / Build

Build with Agent Sandbox.

Browse the catalog for frameworks, tools, and harnesses, each scored on real GitHub credibility.

Get Agent Sandbox →

Browse the catalog