Fast Startup
With <60ms sandbox allocation.
Pre-warmed pools keep isolated environments ready for agent loops, eval batches, and RL rollouts instead of waiting on a new Pod every request.
Pre-warmed PoolsHigh-volume Rollouts
Open-Source Kubernetes Sandbox Engine
Fast, Multi-Cloud
Sandboxes for
AI Agents.
Agent Sandbox enables AI agents to interact with real-world APIs and tools in a safe, isolated environment.
Kubernetes native
Scale across any clouds.
Use the Kubernetes model your infrastructure team already trusts: CRDs, namespaces, RBAC, autoscaling, in-place updates, and multi-cluster routing.
CRDs + RBACInplace UpdatesMulti-cluster
Agentic RL
Easily run Agentic RL with zero rebuilds.
Stay compatible with E2B / SWE-ReX workflows while adding deterministic resets, any-image runtimes, and the scale needed for asynchronous agent training.
E2B compatibleSWE-ReXAny Docker image
A complete console for sandbox operations.
Manage preheated sandboxes, runtime images, datasets, logs, and monitoring from a single control plane built for agent infrastructure.
Sandbox lifecycle and pool management
Manage sandbox runtimes, dataset mounts, image versions, and reset policies without forcing researchers to operate Kubernetes directly.
RL Training
RL Sandbox
RL Sandbox
Connect to any cloud.
AI infrastructure is rarely balanced. Agent Sandbox routes execution environments to the best Kubernetes cluster without forcing teams to rebuild their stack.
Compatible with E2B-style agent workflows.
Keep the sandbox programming model your team already understands while running the backend on Kubernetes. Agent Sandbox is designed for coding agents, evaluation harnesses, and RL pipelines that need fast, repeatable environments.
SWE-Bench VerifiedHugging FaceSWE-GymHugging FaceTerminal-BenchGitHub
Custom RL environmentsBring your own tasks
The sandbox backend AI teams can grow into.
Start with fast allocation, then scale into multi-cloud routing, direct Docker image runtimes, training-framework SDKs, and console-grade observability without replacing your Kubernetes platform.
Speed
<60ms allocation
Pre-warmed Pod pools hand requests to idle sandboxes fast enough for high-volume agent loops and RL rollouts.
Platform
Kubernetes-native adoption
Run on the Kubernetes estate your team already operates, with CRDs, namespaces, RBAC, autoscaling, and in-place updates that preserve warm capacity.
Routing
Cross-region and cross-cloud routing
Dispatch requests across clouds, clusters, and regions without forcing application teams to manage bespoke ingress or routing logic.
Runtime
Zero rebuild runtime changes
Use Docker images directly for SWE tasks, RL environments, terminals, and internal tools without rebuilding VM images.
SDKs
SDKs for agent training
Support E2B-compatible clients, SWE-ReX workflows, and reinforcement learning frameworks with a familiar sandbox API.
Console
Console-grade observability
Give operators a complete view of pools, sessions, logs, metrics, routes, and failures from the product console.
Start with Helm and Kubernetes.
Start with the Helm chart, then follow the installation guide for CRDs, routing, runtime pools, and production configuration.
helm install agent-sandbox oci://ghcr.io/scitix/agent-sandbox-worker