Every AI agent demo looks great. The agent does exactly what you ask, doesn't break anything, and finishes the task perfectly.
Then you try it on your real project. And it breaks something.
Why demos lie
In demos, the agent is working on a clean, empty folder. There's nothing important to break.
On your real project, there's your .env, your database, your production config, your 48,000 files. And the agent can touch all of them.
The fix
Give the agent its own drive letter.
Create a folder for the task, map it to a drive on Windows, point the agent at that drive. Now it's just like the demo — there's nothing important to break.
The agent can run wild. It can delete everything, rewrite everything, do whatever it wants. 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.
What this means
Suddenly the AI agent becomes trustworthy. Not because the model got better, but because the blast radius is contained.
You can let it run for hours. You can let it experiment. You can let it "go off script" — worst case, it deletes the drive and you start over.
The point
AI agents aren't trustworthy on their own. But they don't need to be.
Give them a sandbox, and suddenly they're as safe as they need to be.
The best AI safety isn't a smarter model. It's a smaller filesystem.
Top comments (0)