AWS has released Strands Box, an open-source sandbox designed to rein in autonomous AI agents that operate in what the company calls “YOLO mode” — approving actions without meaningful human review. The tool pairs OS-level isolation with AWS's existing open-source policy engine, Dogwood, to enforce rules deterministically rather than relying on an agent following instructions.
According to AWS, Strands Box addresses a gap in typical container or microVM isolation: those approaches keep agents contained but do not provide contextual rule enforcement. The new tool adds temporal awareness, so policies can consider both an agent's current request and its prior activity. For example, an agent might be allowed to post Slack updates, but no more than three times every ten minutes. It also includes shell and Python interpreters that expose operations like file deletions to the same policy engine, making agent actions more auditable.
AWS emphasized that Strands Box enforces policies without trusting the agent, and that the agent cannot talk its way around the rules. Still, the company cautioned that developers remain responsible for deciding what access to grant and where human review is needed. Strands Box supports any agent or harness and is available on GitHub now, though only for macOS; Linux support is in development and a Windows client is planned but not yet dated.