AgentFlow Sandbox is a unified execution environment designed specifically for Agent Data Synthesis. It provides standardized tool invocation and environment interaction capabilities for large-scale Agent trajectory data synthesis, supporting VM desktop automation, RAG retrieval, command line interaction, web automation, and more.

Why AgentFlow?

In Agent data synthesis scenarios, we need:

• Standardized tool invocation interface: Unified calling method for different types of tools
• Complete trajectory recording: Automatically record every interaction between Agent and environment
• Large-scale parallel execution: Support multi-worker concurrency to improve synthesis efficiency
• Resource isolation and management: Independent execution environment for each worker
• High reliability: Complete error handling and resource cleanup mechanisms

AgentFlow is designed to solve these problems.

Core Features

🤖 Agent Data Synthesis

AgentFlow is optimized for Agent trajectory data synthesis:

• Standardized response format: Unified JSON response for easy data collection
• Complete trajectory recording: Automatically records tool calls, parameters, return values, execution time
• Batch execution support: Supports batch tool calls to improve synthesis efficiency
• Worker isolation: Independent execution environment for each synthesis task

# Typical Agent data synthesis workflow
async with Sandbox() as sandbox:
    await sandbox.create_session("vm")
    
    # Agent decision -> Tool call -> Record trajectory
    result = await sandbox.execute("vm:screenshot", {})
    trajectory.append({
        "action": "vm:screenshot",
        "observation": result["data"],
        "metadata": result["meta"]
    })

🖥️ Multiple Resource Backends

AgentFlow supports various types of backend resources:

Backend	Type	Description
VM	Session Resource	Virtual machine desktop automation
RAG	Shared Resource	Document retrieval service
Bash	Session Resource	Command line interaction
Browser	Hybrid Resource	Web automation
Code Executor	Session Resource	Code sandbox execution

📦 Session Management

Flexible session lifecycle management:

• Explicit Session: Created via create_session(), can be reused multiple times
• Temporary Session: Automatically created during execution, destroyed immediately after use

🌐 HTTP API

Standardized RESTful API interface:

# Execute tool
POST /execute
{
    "worker_id": "sandbox_xxx",
    "action": "vm:screenshot",
    "params": {}
}

# Create Session
POST /session/create
{
    "worker_id": "sandbox_xxx",
    "resource_type": "vm",
    "config": {"screen_size": [1920, 1080]}
}

🔌 Extensible Architecture

Two extension methods:

• Lightweight API Tools: Using @register_api_tool decorator
• Heavyweight Backend: Inheriting Backend base class

Quick Start

from sandbox import Sandbox

async def main():
    async with Sandbox() as sandbox:
        # Create VM Session
        await sandbox.create_session("vm")
        
        # Take screenshot
        result = await sandbox.execute("vm:screenshot", {})
        print(result)
        
        # Click
        await sandbox.execute("vm:click", {"x": 100, "y": 200})

import asyncio
asyncio.run(main())

Next Steps

• Architecture - Learn about system architecture
• Installation Guide - Start installation
• First Sandbox - Quick start tutorial