Limits that live in the system, not the prompt
Telling an AI “don’t do that” is not a limit. Here is where Serene Resolve keeps the real ones.
An instruction in a prompt is a request. A customer can try to talk the model out of it. So every limit that matters in Serene Resolve is enforced by the system itself, whatever the model decides.
Each action has its own limits
Everything the agent can do is a separate action with its own limits, such as how many times it may run on one ticket or in one day, or the largest amount it may handle. A call over a limit is refused and recorded.
Each piece of work has a budget
The agent’s work on a ticket is capped in steps, cost and time. When it reaches a cap, it hands the ticket to a person. It never just stops.
One customer at a time
Lookups are tied to the customer on the ticket, so the agent can’t look up someone else’s account even if asked to. Visitors the desk can’t confirm only get answers from public information.
A tricked agent can’t do much
What customers write is treated as information, never as instructions. And if a model were talked into something anyway, the worst it could do is write a reply that fails the checks, or ask for an action that is refused. It can’t message other people, and it can’t keep talking after it has been stopped.