Sandbox

What is a Sandbox

A sandbox is a restricted execution environment, commonly a container or a virtual machine, where an agent can run commands and edit files without those actions reaching the real system outside it. It exists to contain the blast radius of a mistake: a destructive command run inside a sandbox affects only the sandbox, not production data or a developer's actual machine.

This matters most for agent work with real uncertainty attached: running code from an untrusted source, processing content that might carry a prompt injection attempt, testing a risky change before it touches anything real. A sandbox lets that uncertainty play out somewhere contained instead of somewhere consequential.

A sandbox is related to a worktree but broader, a worktree isolates by directory within the same real system, while a sandbox isolates the whole execution environment itself. Work done inside a sandbox usually still needs a deliberate step to promote it out, a review, a merge, a copy, rather than affecting the real system automatically.