I've tried every method for sandboxing AI coding agents. Here's what worked and what didn't.
Method 1: Prompt engineering
Write detailed rules in the system prompt:
- "Don't touch .env"
- "Don't delete files"
- "Ask before making changes"
Result: Works until it doesn't. The model eventually breaks a rule. Context gets truncated. Instructions get lost.
Verdict: ❌ Unreliable
Method 2: Permission systems
Use the tool's built-in approval gates:
- "Approve this action?"
- "Allow file edits?"
Result: Slow. Friction-filled. Still doesn't prevent mistakes — you're just approving them before they happen.
Verdict: ❌ Slows you down without actually preventing damage
Method 3: VM
Spin up a virtual machine for each task.
Result: Works, but 30 seconds to boot, 2-4 GB of RAM, networking setup, file sharing setup. Overkill for most use cases.
Verdict: ⚠️ Works but heavy
Method 4: Docker container
Run the agent in a Docker container.
Result: Works, but Dockerfile setup, volume mounting, permission issues, learning curve. Overkill.
Verdict: ⚠️ Works but complex
Method 5: Mapped drive letter
Map a folder to a drive letter on Windows, point the agent at it.
Result: Works perfectly. 10 seconds to set up. No overhead. No learning curve. OS-level boundary.
Verdict: ✅ Best option for most people
The tool I use
ScsSubst (https://subst.scsoi.com/) — $9.99 lifetime, one-time purchase. Makes persistent drive mappings that survive reboots.
The bottom line
You don't need a VM or Docker to sandbox an AI agent. You just need a drive letter.
It's the simplest, fastest, and most reliable option.
Top comments (0)