When AI agents hand work to each other, can the last one do more than the customer asked?

A customer asks for a refund. A support agent hands the task to a billing agent, which hands it to a refunds agent. Pick a scenario: every handoff and the final action are checked against what the customer authorized and the company's policy. Runs in your browser; nothing is sent anywhere.

Edit the scenario yourself (JSON)

The rule: authority can only narrow as it's handed on, and the whole chain is re-checked when the action arrives. The worst reason wins. Deterministic, no model. Source, PRD, blind evals and the red-team results.