Skip to main content

​
What is OpenReward’s Architecture?

OpenReward provides infrastructure for hosting and running AI agent environments. The platform has two main components: 
Environment → Sandbox
Environments are persistent servers that host your evaluation tasks, and Sandboxes are temporary containers for running code.

​
The Two Components

​
Environments: Persistent Evaluation Servers

An environment is a long-running server that hosts your evaluation: What it provides:
  • Evaluation server that agents connect to
  • Tasks for agents to solve
  • Tools agents can call
  • Session management for multiple agents
  • Automatic scaling based concurrent sessions
  • Isolated storage for datasets and artifacts
Think of it as: A persistent web service (like SWE-Bench) that agents connect to for evaluation.

​
Sandboxes: Temporary Execution Containers

A sandbox is an isolated container for running code: What it provides:
  • Isolated execution environment
  • Configurable resources (CPU, memory)
  • Network isolation options
  • Automatic cleanup after use
Think of it as: A secure virtual machine that runs agent code during a task session.

​
How They Work Together

​
Common Patterns

Pattern 1: Environment-Only For environments that don’t need code execution: 
Agent → Environment Server → Evaluation Results
Example: Math problems where the environment checks the agent answer against a ground truth. Pattern 2: Environment + Sandboxes For environments requiring code execution: 
Agent → Environment Server → Creates Sandbox → Executes Code → Returns Results
Example: SWE-Bench where agents write code to accomplish the task. Pattern 3: Direct Sandbox Usage For running code without going through an environment server: 
Your Code → Creates Sandbox → Executes Code → Returns Results
Example: Quick code execution or testing.

​
Data Storage

​
Environment Storage

Each environment includes isolated cloud storage that is shared between the environment and sandboxes. Access:
  • Environment server mounts storage at: /orwd_data/
  • Sandboxes can choose the path to mount the storage at
  • Data persists across restarts and sandbox runs

​
Deployment Flow

​
Deploying an Environment

1. Push code to GitHub
   ↓
2. OpenReward builds container image
   ↓
3. Environment server deployed
   ↓
4. Server automatically scales based on connections
   ↓
5. Environment ready for agents

​
Creating a Sandbox

1. Request sandbox (specify image, resources)
   ↓
2. Sandbox provisioned
   ↓
3. Run commands or code in sandbox
   ↓
4. Cleanup sandbox on disconnect or timeout

​
Getting Started

Environments

Learn about environment servers

Sandboxes

Explore ephemeral execution containers

Storage

Configure cloud storage access

Build Your First Environment

Deploy your first environment