Agent Sandbox

Kubernetes-native Sandbox Engine for AI Agents

---

Overview

Agent Sandbox is a Kubernetes Operator that manages AI agent sandbox Pod lifecycles using a pre-warmed Pod pool with in-place image upgrades. Instead of scheduling a new Pod for every sandbox request — which incurs 15–60 seconds of cold-start latency — Agent Sandbox pre-warms a pool of idle Pods and reassigns one to an incoming request in under 100ms.

It is purpose-built for workloads where sandbox allocation speed is critical:

• Reinforcement learning training pipelines (SWE-bench, Terminal-bench, and custom RL environments)
• AI coding agents that need on-demand isolated execution environments
• Multi-agent systems requiring dozens or hundreds of sandboxes simultaneously

---

Key Features

Feature	Description
< 100ms Allocation	Pre-warmed Pod pool eliminates scheduling overhead; sandboxes are ready in milliseconds
In-Place Image Upgrade	Running Pods are updated with a new image without recreation, preserving pool warmth
Cross-Cluster & Multi-Region	ExtProc-based routing dispatches requests transparently across multiple clusters
E2B SDK Compatible	Drop-in replacement for the E2B API — existing E2B clients work without code changes
Optimized for RL Training	Purpose-built for SWE-bench, Terminal-bench, and large-scale RL environment rollouts
Kubernetes Native	Managed via CRDs (SandboxPool, SandboxTemplate); integrates with RBAC, namespaces, and autoscaling
Any Image, No Rebuild	Bring any container image; no custom base image or agent installation required
Prometheus Metrics	First-class observability with a Prometheus endpoint and pre-built Grafana dashboards

Architecture

Components

Binary	Purpose	Ports
cmd/sandbox	Operator + REST API Server	:8080 (API), :8090 (E2B-compat), :8082 (metrics)
cmd/envoyextproc	Data-plane ExtProc for cross-cluster routing	:9002 (gRPC), :9003 (control-plane)
cmd/wsproxy	WebSocket reverse-proxy sidecar for terminal access	:9003 (WS), :9004 (sync)

CRDs

• SandboxPool (sbp, namespace-scoped) — defines a pre-warmed Pod pool with Replicas, optional autoscaling, and an inline or referenced template
• SandboxTemplate (sbt, cluster-scoped) — reusable Pod template with idleImage and runtimes

Performance

Metric	Traditional Kubernetes	Agent Sandbox
Sandbox allocation latency	15–60 s	< 100 ms
Pod churn per request	1 create + 1 delete	0 (pool reuse)
Image pull on every request	Yes (cold start)	No (pre-warmed)
Autoscaling to zero	Supported	Supported
Cross-cluster routing	Manual / external LB	Built-in ExtProc

Quick Start

Prerequisites

• Kubernetes 1.26+
• kubectl configured against your cluster
• helm (optional, for chart-based install)

Use Cases

Reinforcement Learning (SWE-bench / Terminal-bench)

Agent Sandbox is designed to serve as the environment backend for large-scale RL training runs. Thousands of rollout workers can each request a fresh isolated sandbox in milliseconds, dramatically reducing the environment-reset bottleneck:

Cross-Cluster Scheduling

Deploy sandbox pools across multiple clusters or regions. The ExtProc component routes API requests to the appropriate cluster transparently — no changes needed in client code:

Development

Prerequisites

• Go 1.25+
• make
• Docker (for image builds)
• controller-gen, oapi-codegen (installed automatically by make)

Build

# Build all binaries
make build

# Build individual binaries
make build-controller   # sandbox operator + API server (linux/amd64)
make build-extproc      # envoy extproc (linux/amd64)
make build-wsproxy      # websocket proxy

Code Generation

make manifests          # Regenerate CRD YAML + RBAC
make generate           # Regenerate DeepCopy methods
make gen-all-api        # openapi.yaml → Go + TypeScript + Python SDK
make sync-crds-to-helm  # Sync CRDs + manager ClusterRole into Helm charts

Test

make test               # Unit tests (no cluster required)
make test-e2e           # E2E tests (requires a real cluster)

Lint

make lint-fix

---

Contributing

We welcome contributions of all kinds — bug reports, feature requests, documentation improvements, and code. Please read CONTRIBUTING.md before submitting a pull request.

All commits must include a Signed-off-by line (see DCO). Use git commit -s to add it automatically.

---

License

Apache License 2.0 — see LICENSE for details.

Copyright © 2026 ScitiX.