Agent Sandbox is a next-generation sandbox compute service designed for AI agents, enabling large-scale, production-grade applications.
Introduction
Agent Sandbox is a sandbox compute service for production-grade AI agents. It provides a MicroVM-level isolated execution environment, featuring memory-level hibernation and resume, checkpoint and cloning capabilities, and massive elastic scaling of up to 15,000 sandboxes per minute. The service is fully compatible with the native Kubernetes ecosystem and seamlessly integrates with mainstream AI agent frameworks and tools, such as the E2B SDK and AgentScope.
Agent Sandbox is currently in public beta and is available only in the China (Guangzhou), China (Heyuan), China (Qingdao), China (Zhangjiakou), China (Hohhot), and China (Chengdu) regions.
Core features
Strong security isolation
Each sandbox has its own independent, secure execution environment with MicroVM-level isolation.
It offers end-to-end isolation across compute, network, and storage. It supports sandbox-level logging and monitoring for auditing and troubleshooting.
Large-scale, low-latency elasticity
It supports image cache acceleration, which makes an image ready in seconds and reduces pull times by over 90% in typical scenarios. This significantly shortens the image preparation phase of sandbox creation.
It supports pre-scheduling optimizations based on workload characteristics to improve creation efficiency in high-concurrency scenarios. It can create up to 15,000 sandboxes per minute, accelerating AgentRL iteration efficiency.
It leverages warm pool optimizations to support rapid sandbox creation in hundreds of milliseconds.
State persistence
It supports on-demand hibernation for running sandboxes, preserving their memory state. This allows sandboxes to resume rapidly to respond to interactive AI agent requests. In typical scenarios, resuming from a memory state takes 1 to 10 seconds.
It supports checkpoint and restore for a sandbox's memory state, allowing you to save and migrate AI agent states. This accelerates scenarios that involve parallel branch exploration.
Flexible and rich ecosystem
Rich scenarios: It supports diverse AI agent sandbox use cases, such as Code Interpreter and Browser Use.
Integration methods:
E2B-compatible SDK (Recommended): It provides an integration method compatible with the E2B ecosystem. You can continue using the E2B SDK calling pattern to create, connect to, run, and release sandboxes, which lowers migration effort.
Sandbox CR (Recommended): It provides a declarative integration method. You can use a custom resource to manage sandbox templates, runtime parameters, and lifecycles.
Kubernetes ecosystem: It offers deep and comprehensive integration with the native Kubernetes ecosystem, ensuring compatibility with existing storage, networking, and monitoring systems.
Use cases
AgentRL scenarios: Ideal for reinforcement learning training, trajectory sampling, environment interaction, and multi-path exploration. Agent Sandbox supports the concurrent creation of large-scale sandboxes, fast teardown, and state reuse to improve training throughput and resource utilization.
AgentServing scenarios: Designed for online AI agent services, such as in-depth research, tool calls, and multi-turn conversations. It provides a strongly isolated execution environment, elastic startup capabilities, and hibernation and resume features to balance responsiveness with operational costs.
OpenClaw personal assistant scenarios: For the rapid development and deployment of applications like personal assistants or digital workers. Agent Sandbox supports multiple tool environments, including code, browser, and desktop, facilitating a smooth transition from prototype to production deployment.
Billing
Billing is based on the vCPU and memory specified at creation time. For unsupported vCPU and memory configurations, the cluster automatically adjusts them to the nearest supported specification, and billing is based on the adjusted configuration. Agent Sandbox supports hibernation, and billing differs based on the sandbox state.
Running state: The first 30 GiB of ephemeral storage is free. Any usage beyond this is charged based on cloud disk prices.
Hibernation state: No charges for vCPU or memory. There is no free tier for ephemeral storage; all usage, including the first 30 GiB, is billable.
GPU computing is not currently supported. Only CPU and memory configurations are available.
Billing formula
Single sandbox cost = (Number of vCPUs × vCPU unit price + Memory size × Memory unit price) × Billing duration.
Billing method
Billing is based on each sandbox instance on a pay-as-you-go basis, with per-second billing and hourly settlement.
Region | Compute type | Billable item | |
vCPU | Memory (GiB) | ||
Chinese mainland | sandbox | CNY 0.00003006/second (CNY 0.108/hour) | CNY 0.00001499/second (CNY 0.054/hour) |