Agent Sandbox

Introduction

Agent Sandbox provides MicroVM-level isolated runtime environments with memory-level hibernation/wake-up and checkpoint/restore. It scales up to 15,000 sandboxes per minute and is fully compatible with the native Kubernetes ecosystem, integrating with the E2B-compatible SDK and AgentScope.

Note

Agent Sandbox is in public preview. To report issues or suggestions, submit a ticket or contact us through other support channels.

Key features

Strong security isolation

Each sandbox runs in a MicroVM-level isolated environment with an independent, secure execution space.
End-to-end compute, network, and storage isolation with sandbox-level logging and monitoring for auditing and troubleshooting.

Elastic scaling

Image caching reduces pull times by over 90%, accelerating sandbox readiness.
Predictive scheduling based on workload characteristics enables up to 15,000 sandbox creations per minute, accelerating AgentRL iteration.
Warm pools enable sandbox creation in hundreds of milliseconds.

State persistence

On-demand hibernation preserves sandbox memory state for rapid wake-up (1–10 seconds typical), enabling responsive interactive agent requests.
Checkpoint and restore saves and migrates agent state, accelerating parallel branch exploration.

Ecosystem

Use cases: Supports AI agent sandbox scenarios including Code Interpreter and Browser Use.
Access methods:
- E2B-compatible SDK (Recommended): Create, connect, execute, and release sandboxes with the E2B SDK, minimizing migration cost.
- Sandbox CR (Recommended): Declaratively manage sandbox templates, runtime parameters, and lifecycles through custom resources.
Kubernetes ecosystem: Fully compatible with existing Kubernetes storage, networking, and observability systems.

Use cases

AgentRL: Concurrent sandbox creation at scale with rapid destruction and state reuse for reinforcement learning, trajectory sampling, environment interaction, and multi-path exploration, improving training throughput and resource utilization.
AgentServing: Strongly isolated execution with elastic scaling and hibernation/wake-up for online agent services—deep research, tool invocation, and multi-turn conversations—balancing responsiveness with cost.
OpenClaw personal assistant: Build and deploy personal assistant or digital employee apps with code, browser, and desktop tool environments, from prototype to production.

Billing

Billing is based on the vCPU and memory specified at creation. Unsupported configurations are automatically normalized, and you are billed for the normalized specification. Billing differs by sandbox state:

Running state: The first 30 GiB of ephemeral storage is free. Any usage beyond this amount is billed according to cloud disk prices.
Hibernating state: CPU and memory are not billed. There is no free tier for ephemeral storage; all usage is billed.

Currently, GPU compute is not supported. Only CPU and memory configurations are available.

Billing formula

Single sandbox cost = (vCPU Count × vCPU Unit Price + Memory Size × Memory Unit Price) × Billing Duration.

Billing method

Each sandbox is billed pay-as-you-go, calculated per second and settled hourly.

Region	Compute type	Billable items
		vCPU	Memory (GiB)
Chinese mainland	Sandbox	CNY 0.00003006/second (CNY 0.108/hour)	CNY 0.00001499/second (CNY 0.054/hour)