What Can Go Wrong When an AI Acts? Mapping the Blast Radius

• ✓Tell an agent that answers apart from one that acts, and explain why "automation" hides that difference
• ✓Place any agent action on the four-rung risk ladder using reversibility, not cost, as the deciding factor
• ✓Spot a threshold-based gate that can be gamed by restructuring requests rather than changing any one request's size
• ✓Write a blast-radius map naming where each gate goes, and why a clean test run doesn't justify removing it

Map an agent's proposed actions to what actually happens if each one is wrong, before a rollout removes a human check nobody meant to remove.