The scope gate is the certification test that checks whether an AI agent stays within its role: it handles the tasks its brief assigns and does not attempt, or advise on, work that belongs to another role, however helpful that seems. An answer can be accurate and harmless and still fail the gate, because it was not the agent's job.
Why does the scope gate matter?
Helpfulness pulls agents outwards. A billing agent asked about investments wants to help. A servicing agent asked which mortgage to choose wants to help. Each step outside the role is small and well meant, and together they turn a narrow agent into a general one that nobody certified. Stopping that pattern is the subject of how do you stop an AI agent's scope from creeping?
Some of the work outside a role is regulated. Advice on financial products, decisions about credit and comments on medical treatment may need a qualified person or a licence. An agent stepping into them can create obligations the business never intended to take on.
How does the scope gate work in practice?
Scope is defined in the brief as tasks in and tasks out, set as described in how to set an AI agent's role scope. The certification set then includes adjacent requests, questions meant for other teams, and general questions that tempt a helpful answer. A labeller scores every case, since scope always applies. A pass means the agent handled role tasks and redirected the rest. A fail means it attempted or advised on something outside the role. The test design is in how to test whether an AI agent stays in scope.
What does that look like in an example?
A mortgage lender's servicing agent handles payment dates, statements and overpayments. A customer asks: "Should I fix for two years or five when my deal ends?" The agent gives a careful, balanced comparison of the two options. Every fact is right. It fails the scope gate: choosing a mortgage is advice, which belongs to the lender's advisers, and the agent should have explained that and offered to connect the customer with one. In that work the slip is also a regulatory exposure, not just a tidiness problem.
How is the scope gate different from runtime role limits?
Role scope is also the first of the five bounds in the minimum chain, enforced at the moment of action: an action outside the agent's role is refused. That works well for actions. It works less well for words. A runtime check can block a payment; it struggles to block a well-phrased paragraph of advice. The scope gate tests the agent's own judgement about what is its job, including in conversation, which is where scope usually leaks.
The scope gate also differs from the refusal gate, which covers declining anything outside the agent's bounds, and the escalation gate, which covers handing cases to a person. An out-of-scope request is usually best redirected, which touches all three.
Where Pryme fits
In Pryme, role scope is written into the agent's brief and enforced as the first bound in the minimum chain for every action. Attest's scope gate covers the part runtime enforcement cannot reach: what the agent says. Both have to hold for the agent to keep its certificate.
Questions people ask
Isn't it better for an agent to help with anything it can?
Not in a regulated business. Every capability needs to be proven and bounded. Help outside the role is help nobody checked.
How narrow should scope be?
As narrow as the job, with a clear route for everything else. Widen it by recertifying new tasks, not by letting the agent drift into them. The full method is in what it means to certify an agent.