Using a Docker Sandbox for a Coding Agent

This guide shows a concrete, end-to-end example of how Docker Sandboxes can be created and used to safely run autonomous coding agents that can install packages, modify files, and even run Docker — without touching your host machine.

Scenario

You want to let a coding agent:

• Modify a real codebase
  


• Install system dependencies
  


• Build and run containers
  


• Run unattended with permissive flags
  

…but without risking your laptop or credentials.

Docker Sandboxes solve this by running the agent inside a disposable microVM with only your project workspace mounted.

Step 1: Create a Sandbox

Create a new sandbox using your local project directory as the workspace:

docker sandbox create   --name agent-sandbox   --workspace ./my-project

What this does:

• Creates a dedicated microVM
  


• Mounts only ./my-project into the sandbox
  


• Keeps your OS, home directory, and secrets isolated
  

Step 2: Enter the Sandbox

Start an interactive shell inside the sandbox:

docker sandbox exec agent-sandbox bash

You are now inside the sandbox, not your host machine.

From this point on:

• Any package installs
  


• Any config changes
  


• Any Docker commands
  

…are fully isolated.

Step 3: Run a Coding Agent (Unattended)

Inside the sandbox, run your coding agent in permissive mode:

claude-code run   --dangerously-skip-permissions   --project /workspace

Why this is safe:

• The agent runs inside a microVM
  


• Only the project directory is writable
  


• No access to your host OS, SSH keys, or credentials
  

This is the intended workflow for Docker Sandboxes.

Step 4: Let the Agent Install Dependencies

The agent can freely modify the environment:

apt-get update

apt-get install -y nodejs npm

No permission prompts.


No approval loops.


No risk to your machine.

Step 5: Let the Agent Use Docker

Inside the sandbox, the agent can build and run containers:

docker build -t my-app .

docker run -p 8080:8080 my-app

Important notes:

• This does not use your host Docker daemon
  


• Containers run entirely inside the sandbox microVM
  

This capability is what makes Docker Sandboxes fundamentally different from regular containers.

Step 6: Review the Results on the Host

Exit the sandbox and inspect the changes:

git diff

You’ll see only the intentional code changes made by the agent.

No stray system packages.


No modified OS files.


No lingering background processes.

Step 7: Delete the Sandbox

When you’re done, delete the sandbox:

docker sandbox delete agent-sandbox

The microVM is destroyed immediately.

• The environment is wiped
  


• Nothing persists except your code changes
  


• You start clean the next time
  

Why This Pattern Works

Docker Sandboxes give agents:

• A real operating system
  


• Package managers and system tools
  


• Docker access
  


• Full autonomy
  

While giving you:

• Strong isolation via microVMs
  


• Disposable environments
  


• Zero host contamination
  


• Confidence to use permissive agent modes
  

When to Use Docker Sandboxes

This pattern is ideal when:

• Running coding agents unattended
  


• Using flags like --dangerously-skip-permissions
  


• Allowing agents to install tools dynamically
  


• Letting agents build and run containers
  


• Experimenting aggressively without fear
  

Agents need freedom.


Your machine doesn’t.