The AI safety debate is stuck on one question: should agents have permission to take actions?
The answer is neither yes nor no. The answer is: give them a sandbox where it doesn't matter.
Permission systems are broken
Every AI tool has some version of:
- "Approve this action?"
- "Allow file edits?"
- "Confirm before running commands?"
These are all friction. They slow down the agent. They interrupt your flow. And they don't actually prevent damage — the agent will still do something unexpected eventually, and you'll still have to clean up.
Sandboxes are underrated
What if you didn't have to approve anything? What if the agent could do whatever it wanted, and you knew it couldn't hurt anything important?
That's what a mapped drive letter gives you.
Create a folder, map it to a drive on Windows, point the agent at that drive. The agent can delete files, rewrite everything, run wild — it's fine. It's all inside the sandbox.
The tool
I use ScsSubst (https://subst.scsoi.com/) for this. $9.99 lifetime, one-time purchase. It wraps the Windows subst command with a GUI and persistent mappings.
No permission prompts. No approval gates. Just a drive letter.
The workflow
- Spin up a new drive for the task
- Point the agent at it
- Walk away
- Review the results when done
The agent works faster. You work less. And nothing important gets broken.
The point
Permission systems ask "should the agent be allowed to do this?"
Sandboxes ask "what happens if the agent does anything at all?"
The second question is easier to answer. And the answer is: nothing bad, as long as it's on its own drive.
Top comments (0)