What does an AI agent certificate cover?

One configuration of the agent, one organisation's operation, one depth and one period. Within those, it covers the behaviour shown on the test cases, gate by gate, and any conditions attached. Everything outside those lines, including other agents, changed configurations, other customers and the time after it lapses, is not covered.

What should be written on the certificate?

A certificate is only as useful as it is precise. Each field narrows what it covers.

A ready format is in the AI agent certificate template, and the concept itself is explained in what an agent certificate means for AI agents.

Why so narrow?

Because each field is something that, if changed, can change the agent's behaviour. A certificate that named only the agent would claim to cover every configuration that agent might ever have. That claim would be false the first time a source was updated.

Narrow certificates are also easier to rely on. An auditor or supervisor asking "was this action covered?" can check the fields against the action and get a yes or no.

What does the certificate not cover?

• Actions that need a higher depth. Each action needs its own depth, and the certificate names one.
• Other organisations. Evidence from one operation does not transfer; see whether one agent certificate can cover every customer.
• A changed configuration. A new model, source, brief or tool limit takes the agent outside the certificate.
• The period after lapse. Evidence ages; the certificate stops counting on its date.
• Correctness of the sources. A certificate shows the agent follows its sources faithfully, not that the sources are right.
• Cases unlike the test set. If a new kind of case appears, the certificate has no evidence about it.

What controls follow from the certificate?

A certificate is only meaningful if it is enforced. Several controls follow directly from it:

• Depth enforced at the moment of action. The agent cannot act beyond the certified depth, whatever its instructions say.
• Conditions enforced too. If the certificate says hardship cases go to a person, the platform holds them.
• Change detection. A change to any listed field voids the certificate until a rerun passes.
• Lapse enforcement. On the lapse date, depth drops unless a renewal has passed.
• A record linking every action to its certificate, so coverage can be shown later.

Why certificates end, and what happens then, is covered in why an agent certificate lapses.

An illustrative example

Take a mortgage servicer's agent that answers borrowers' questions about their current balance and next payment. Its certificate covers it at depth 3, on the servicer's own cases, using a named model version, brief version 4, and the servicing system and payment schedule as sources. A condition holds any conversation mentioning financial difficulty for a person. It was accepted by the head of servicing operations and lapses in 90 days.

Six weeks later, the product team adds a new rate-switch product. Borrowers start asking the agent about it. The certificate does not cover those questions: the product guide is not a listed source, and rate-switch questions were not in the case set. The agent hands them off until it is re-certified on cases that include them.

How should certificates be kept?

In a register, with each certificate linked to its evidence, its reviewer and the actions taken under it. At renewal, a certificate renewal checklist for AI agents helps confirm nothing listed has drifted. The reasoning behind the whole approach is in what it means to certify an agent.

Where Pryme fits

An Attest certificate records the agent, the customer, the depth, the configuration tested and the results for each of seven gates. It lapses after 90 days. Every action filed in Pryme references the certificate in force, so coverage can be checked action by action.