An AI assistant works through one morning's inbox at a small print studio. Run it without a seatbelt, then with one, and be the person it asks.
The agent spotted the phishing email and still paid an invoice twice. A read can be repeated; sending, deleting and paying cannot, so that is where the checks go.
One polite email argued the agent past a prohibition in its memory. It could not argue a send function into a mail tool that has none.
Each capability has a level from 0 to 4. Never granted means denied, whoever approves; granted but above the task's level means a person decides, with the evidence in front of them.
Both runs ended with the same “all done”. Only the gates over the real files noticed that the price list PDF still showed the old prices.
The guardrails in this simulation are the ones our TwinCortex agent runs on: tools without a send function, capability levels, intent checks on irreversible clicks and completion gates over real files. The article walks through each of them.