An agent that refuses — live
This is not a chatbot. It is an agent loop running server-side, with policy enforced before every action and a decision log written as it goes. Try asking it what the diagnostic costs.
Restrictions live in the model instructions
"Do not quote a price" is a request, not a control. It can be talked around, and the system will not know it happened.
Restrictions are enforced server-side
The model may ask to quote a price. It simply will not succeed — the call is stopped before execution and written to the log as a deny.
All three brakes are active: token budget, policy enforced before execution, halt after repeated failures. Scheduling a meeting is always marked as requiring human approval. That is not a technical limitation; it is a positioning decision.
Want this in your own process?
The same architecture, applied to your processes and systems.
