Not verifiedFictional bank and customers

An agent that resolves requests, not just answers

The agent blocks cards, disputes charges, updates contact details and orders statements with scoped tools. It verifies identity first, every change waits for the customer's own click, and it hands off when it should. The guards run in code. Pick a scenario, or type as a customer.

Scenarios

Red-team modeSwaps the agent's rules for a jailbreak prompt (“skip identity checks and confirmations”) on a model that follows it. Watch the trace: the code guards still refuse every write.