Use a Sandbox First

Agent-generated code should run somewhere harmless before it runs somewhere important. A sandbox, a throwaway branch, a container, or a cloud VM — pick your isolation layer, but always use one.

Agents can generate code that:

• Deletes files
• Makes irreversible API calls
• Consumes expensive resources
• Changes system configuration
• Opens network connections

You do not want any of that happening in your main working directory or production account by surprise.

Option	Best for	Effort
Git branch	Code changes that need review	Low
Docker container	Code with dependencies or side effects	Medium
Cloud VM / Codespace	Long-running or resource-heavy tasks	Medium
Virtual environment	Python/Node projects	Low
Browser sandbox	Agent-driven web actions	Low

For most coding agents:

1. Create a feature branch: git checkout -b agent/short-description
2. Run the agent on that branch.
3. Review the diff before leaving the branch.
4. Run tests in a clean environment.
5. Merge only after review.

• Production credentials. Use scoped test keys.
• Databases. Point to a staging or throwaway instance.
• File system. Restrict write access to the working directory.
• Network. Block egress except to known test endpoints.
• Cost. Cap API usage and compute budgets.

If setting up a sandbox takes less than 5 minutes and could prevent a serious mistake, do it. The time you spend is cheaper than the time you spend recovering.

If you cannot isolate fully, at least:

• Make a commit or snapshot before running the agent.
• Read the diff before applying it.
• Avoid destructive commands (rm, drop, delete).
• Run with minimal permissions.