Sandbox

An execution environment with explicit isolation and resource boundaries intended to contain code, data, network, or process effects.

A sandbox may isolate filesystem paths, processes, credentials, devices, network destinations, system calls, time, or resource usage. Boundaries differ across containers, virtual machines, browser sandboxes, and hosted agent runtimes. Mounted host folders, inherited secrets, privileged modes, kernel vulnerabilities, or open networks can pierce the intended containment.

Builder example

Sandboxing reduces impact but does not establish correctness or make promotion to production safe. Use least privilege inside the environment, scan outputs, separate production credentials, and test which resources the sandbox blocks. Some tasks can run safely in a restricted real workspace without a copy-and-manual-promotion ceremony.

Common confusion: A sandbox isolates where an agent runs; a guardrail blocks a specific forbidden action before it runs. What separates them is scope: the sandbox contains the blast area for any mistake, while a guardrail stops one named action no matter where the agent is. Strong setups use both, since a sandbox limits damage and a guardrail prevents a known bad move outright.