The agent works in a loop of reading, editing and running. The safeguards live in code between model and operating system: a permission engine that denies, asks or allows in a fixed order, a sandbox for every shell command and a snapshot before every change.
The robot is built in simulation first, but in a way that gives the result a chance on real hardware. The simulation shows the CAD parts with their masses.
Language models write detection rules for Microsoft Sentinel fluently. The expensive kind of error stays invisible: rules that run without error and never return a row.