Docs

Volcano Engine Function as a Service

Cloud sandbox

Overview of cloud sandbox

veFaaS cloud sandboxes are the next-generation infrastructure for application scenarios such as reinforcement learning (RL) training and AI agents. A cloud sandbox provides a secure, isolated, and efficient cloud-based environment. This topic introduces the concept and advantages of cloud sandbox.

Basic concepts

Cloud sandbox: a serverless service built for AI task execution, offering high elasticity, low cost, stateful operation, secure isolation, easy maintenance, and rich features. You can create cloud sandbox environments within seconds based on your business requirements without the need to be concerned about the underlying infrastructure. In inference scenarios, a cloud sandbox provides an AI-driven task execution environment. In training scenarios, such as model Reinforcement Learning (RL) training or Reinforcement Fine-Tuning (RFT) training, a cloud sandbox execution environment supports code execution, enhanced tool calling, and complex task handling.

Sandbox applications: they are the basic units of cloud sandbox, responsible for scheduling and operation of resources. A sandbox application is built based on sandbox images and related configurations.

Sandbox instances: they are secure and isolated container runtime environments realized by loading and running sandbox images, with configurable lifetime for each instance.

Sandbox images: they are standardized packages that include all of the files, binaries, libraries, and configurations to run sandbox instances. Cloud Sandbox supports custom images and ready-to-use public images.

Advantages

Rapid provisioning of resources at scale: A cloud sandbox offers ultimate elasticity and massive resource support. During the RL training phase, it can concurrently launch a large number of sandbox resources, thereby effectively improving the overall training efficiency.

Deep integration with RL frameworks: You can use the out-of-the-box Sandbox Fusion that is deeply integrated with the veRL (Volcano Engine Reinforcement Learning for LLMs) framework. You can also use custom sandbox images to flexibly adapt to various RL frameworks.

Rich templates available: Cloud sandboxes provide ready-to-use runtime environment templates for Code, Browser, and Computer. These runtime environments help users focus on business logic development. Custom sandbox images are also supported to meet personalized business requirements.

Stateful support: Users can specify sandbox instances to run for extended periods, with requests sharing the same sandbox runtime environment.

‌Elasticity without maintenance: Based on a Serverless architecture, a cloud sandbox can start up in less than 100 milliseconds and supports large-scale resource scaling.

Image distribution: Warm-up of thousands of images is supported to help meet the need for rapid sandbox startup in multi-image scenarios.

Millisecond-level billing: Sandbox resources are billed by request volume and execution duration down to milliseconds, with no charges incurred when they are not in use.

‌Flexible access‌: A cloud sandbox supports convenient access methods such as MCP, API, and SDK, allowing flexible and quick use.

Application scenarios

RL training

Under the Agentic AI trend, model RL/RFT training focuses on enhancing tool invocation, handling complex tasks, and generating code. Such reinforcement learning tasks require starting a large number of sandboxes as execution environments for model task running and experience collection. Cloud sandboxes, with their extreme elasticity, large-scale resource support, and batch image distribution capabilities, can effectively improve overall RL efficiency.

AI application inference

General-purpose agent-based AI applications need to launch cloud sandboxes as isolated environments to execute AI-driven tasks when completing complex tasks, such as in-depth research. Cloud sandboxes provide public images for runtime environments, such as Code, Browser, and All-in-One environments. They allow developers to implement custom development or use them out of the box. This effectively improves the development efficiency of AI applications. Additionally, the extreme elasticity of a cloud sandbox enables the system to instantly launch the corresponding number of sandbox instances based on user requests. This effectively supports the high-concurrency demands of online AI applications.

Usage flow

Follow the steps below to use cloud sandbox:

Prepare the sandbox images to be used:

If you intend to use the public images provided by veFaaS, there is no additional preparation required.

If you intend to use custom images, you need to first upload the image to the Container Registry (CR) and perform image preheating.

Create a sandbox application: create a basic unit of the cloud sandbox configured with gateway routing, which facilitates scheduling and operation of resources in the cloud sandbox.

Create a sandbox instance: create a sandbox instance based on a sandbox image you selected.

Test the sandbox application: when sandbox instances are running, use the curl command or the Playground tool to perform tests.

Billing description

The sandbox will be charged during its lifetime. For billing details, see Pay-as-you-go - Cloud sandbox.

Last updated: 2026.08.19 14:06:07