Skip to Content
Coding Agents

Last Updated: 3/7/2026

Coding Agents

E2B sandboxes are perfect for building AI coding agents that can write, test, and execute code safely.

Why Use E2B for Coding Agents?

Isolated Execution

  • Each agent runs in its own secure sandbox
  • No risk to your host system
  • Perfect for untrusted or AI-generated code

Fast Startup

  • Sandboxes start in seconds
  • On-demand creation and destruction
  • Optimized for agent workflows

Full Linux Environment

  • Complete access to filesystem
  • Install any packages or tools
  • Run any programming language

Basic Example

import { Sandbox } from 'e2b'
 
// Create a sandbox for code execution
const sandbox = await Sandbox.create()
 
// AI-generated code
const code = `
import pandas as pd
import matplotlib.pyplot as plt
 
# Create sample data
data = {'x': [1, 2, 3, 4, 5], 'y': [2, 4, 6, 8, 10]}
df = pd.DataFrame(data)
 
# Create plot
plt.plot(df['x'], df['y'])
plt.savefig('output.png')
print('Plot saved!')
`
 
// Write code to file
await sandbox.files.write('/home/user/script.py', code)
 
// Execute the code
const result = await sandbox.commands.run('python /home/user/script.py')
console.log(result.stdout)
 
// Download the generated plot
const plotData = await sandbox.files.read('/home/user/output.png')
 
await sandbox.close()

Advanced Patterns

Persistent Code Context

Keep a sandbox alive across multiple agent interactions:

// Create once
const sandbox = await Sandbox.create()
 
// Multiple operations
await sandbox.commands.run('pip install numpy pandas')
await sandbox.files.write('/home/user/data.csv', csvData)
await sandbox.commands.run('python analyze.py')
 
// Close when done
await sandbox.close()

Template with Pre-installed Tools

Create a custom template with all your tools pre-installed:

FROM ubuntu:22.04
 
RUN apt-get update && apt-get install -y \
    python3 python3-pip \
    nodejs npm \
    git
 
RUN pip3 install pandas numpy matplotlib scikit-learn

Then use it:

const sandbox = await Sandbox.create({
  template: 'your-template-id'
})

Integration with AI Frameworks

LangChain

import { Sandbox } from 'e2b'
import { Tool } from 'langchain/tools'
 
class E2BCodeExecutor extends Tool {
  name = 'code_executor'
  description = 'Executes Python code in a secure sandbox'
  
  async _call(code) {
    const sandbox = await Sandbox.create()
    try {
      const result = await sandbox.commands.run(`python -c "${code}"`)
      return result.stdout || result.stderr
    } finally {
      await sandbox.close()
    }
  }
}

OpenAI Function Calling

const functions = [{
  name: 'execute_code',
  description: 'Execute Python code in a secure sandbox',
  parameters: {
    type: 'object',
    properties: {
      code: {
        type: 'string',
        description: 'Python code to execute'
      }
    },
    required: ['code']
  }
}]
 
// Handle function call
async function executeCode(code) {
  const sandbox = await Sandbox.create()
  const result = await sandbox.commands.run(`python -c "${code}"`)
  await sandbox.close()
  return result.stdout
}

Best Practices

1. Resource Management

  • Always close sandboxes when done
  • Use try/finally blocks to ensure cleanup
  • Set timeouts for long-running operations

2. Error Handling

  • Check exit codes: result.exitCode
  • Capture stderr: result.stderr
  • Implement retry logic for transient failures

3. Security

  • Never pass sensitive data directly in code strings
  • Use environment variables for secrets
  • Validate AI-generated code before execution

4. Performance

  • Reuse sandboxes for multiple operations
  • Use templates for faster startup
  • Enable caching for dependencies

Common Patterns

File-based Communication

// Write input
await sandbox.files.write('/home/user/input.json', JSON.stringify(data))
 
// Process
await sandbox.commands.run('python process.py')
 
// Read output
const output = await sandbox.files.read('/home/user/output.json')
const result = JSON.parse(output)

Streaming Output

const process = await sandbox.commands.run('python long_script.py', {
  onStdout: (data) => console.log('stdout:', data),
  onStderr: (data) => console.error('stderr:', data)
})

Multi-step Workflows

// Step 1: Install dependencies
await sandbox.commands.run('pip install -r requirements.txt')
 
// Step 2: Run tests
const testResult = await sandbox.commands.run('pytest tests/')
 
// Step 3: Execute main script
if (testResult.exitCode === 0) {
  await sandbox.commands.run('python main.py')
}
Ci CdLifecycle