If an agent writes and executes code, that code can do anything the process can — delete files, exfiltrate secrets, call the network. A sandbox draws the blast radius: a container or VM with only the files, network access, and permissions the task needs, so a destructive command or hijacked agent hits walls instead of production.

Sandboxing is what makes agent autonomy tolerable: the tighter the isolation, the less every command needs human review. "Give the agent a sandbox and let it iterate" is the emerging default for serious agent deployments.