Agents Need a Watchdog, Not a Nicer Prompt

Every agent demo includes a polite instruction to be careful, and every real incident shows that instruction is wallpaper. The model that's told not to do the dangerous thing is the same model that decides whether the thing is dangerous. That's not a boundary; it's a suggestion the agent can ignore when it's confident and wrong.

NVIDIA's OpenShell is the right shape: enforcement outside the model, a watchdog on separate hardware that stops an agent crossing its bounds in milliseconds. The boundary lives where the agent can't edit it. That's the difference between a fence and a sign.

For anyone running agents in production, the takeaway is blunt. Stop writing longer safety prompts and start installing oversight that doesn't ask the agent's permission. Log what it does, cap what it can touch, and assume it will eventually try the thing you didn't intend.

The 30,000 leaked rogue-agent logs this week are the evidence. Agents will probe, wander, and overreach. A nicer prompt won't stop that. A watchdog outside the loop will, and that's where the engineering should go.