← All terms

define sandbox --plain-english

Illustration for "Sandbox" from the Non-Technical Technical Dictionary

Sandbox

TLDR:The scariest part of letting an AI agent loose on your computer is the small voice asking what happens if it does something dumb.

The scariest part of letting an AI agent loose on your computer is the small voice asking what happens if it does something dumb. A sandbox is the answer that lets you say yes anyway.

A sandbox is a walled-off space where code or an AI can run and make a mess without touching anything outside the box you put around it. It can create files, delete files, run commands, even break things outright, and your real computer, your real files, and your real accounts sit on the other side of the wall.

The name is literal. A toddler's sandbox has high sides for a reason: the kid can build, dig, and knock the whole thing down, and the living room stays clean. A sandbox for software is the same deal. You give the agent its own throwaway version of a computer, let it do whatever the task needs, and a well-built box keeps the mess from reaching the real thing.

This is the safety idea that matters most the moment you hand an AI real power. An agent that can run commands on your machine is enormously useful and a little terrifying, because one confident wrong move could delete the wrong folder. Run it in a sandbox and the worst case shrinks dramatically: it wrecks a box you can throw away and recreate in seconds, not your laptop.

It sits right next to two other ideas:

  • Permission scope decides what an agent is allowed to reach, the doors it holds keys to. A sandbox is the bigger wall around the whole thing, so even if it tries a door it should not, there is nothing real waiting behind it.
  • A container is one common way to build a sandbox: a sealed box with its own little world inside, the kind of thing you are happy to delete the second the work is done.

One honest catch: a sandbox is only as strong as its walls. If the box can still reach your real files, your logged-in accounts, or the open internet, the mess can still leak out. A good one is sealed tight by default.

So the question to ask first, before you let an AI do something with real consequences, is simple: where is it running, and how sealed is that box? On your actual machine, with your actual files in reach, you watch it like a hawk. In a proper sandbox, you let it work and just check the result. Powerful and caged beats powerful and loose.

A sandbox is a box where software can run and fail safely, the high walls that keep the mess off your real machine. It turns letting an agent loose from a gamble into something you do on purpose. Let it run. Just build the walls first.