Agent certificate: what it means for AI agents

An agent certificate is a dated record that an AI agent has been proven to a specific depth, for a specific customer, by passing certification on that customer's own cases. It states what the agent may do, on what evidence, under what conditions and until when. Without a current certificate, the authority it granted does not exist.

Why does an agent certificate matter?

It answers the three questions a supervisor, an auditor or a board member will ask about any agent that acts: what is it allowed to do, on what evidence, and until when? A vendor's claim that its agent is "enterprise-ready" answers none of them. A certificate answers all three, for one agent at a time.

It also enforces something. In our model, every action meets five bounds at once: role scope, certificate, sources, the customer's own limits and the platform guardrails. The certificate is one of them. An agent with a generous role and a generous customer limit is still stopped at the depth its certificate supports.

How does a certificate work in practice?

A certificate is issued after a certification run clears every gate for at least one depth. It records the depth requested, the ceiling the agent's role allows, the depth the evidence supports, and the lower of the two as the effective depth. It lists the cases used, the score at each of the seven gates, the conditions that applied during the run and the sources it depends on. The AI agent certificate template sets out every field.

It ends in one of three ways. It reaches its expiry 90 days after issue, which is certificate lapse. A later run falls short, which leaves the agent at the highest depth it cleared, or at depth 1 the same day if it cleared none. Or something it depends on changes, such as its sources. The reasons for a fixed term are explained in why does an agent certificate lapse?

What does that look like in an example?

A motor insurer certifies its claims agent. The certificate names one depth: depth 4 requested, but the evidence supported only depth 3. Acknowledging new claims and requesting missing documents need depth 3, so the agent does them alone. Settling windscreen claims moves money and needs depth 4, so each settlement is held for a handler. Conditions: settlements under the insurer's limit only. Expiry: 90 days from issue. A claims manager reading it knows exactly what the agent may do alone, and what it may not.

How is it different from other certificates?

A website's security certificate proves a server's identity; it says nothing about behaviour. An ISO/IEC 42001 certificate shows an organisation runs a certified AI management system; it does not test any one agent's answers. A vendor certification says a product met a standard in general. An agent certificate is narrower and more useful for operations: this agent, this depth, this customer, on these cases, until this date. What falls inside and outside that scope is covered in what does an AI agent certificate cover?

Where Pryme fits

Attest issues certificates after replaying the agent against the customer's own cases and scoring every answer against a floor at each gate. The certificate is checked at the moment of action as one of the five bounds, and the verdict names it when it is the bound that decided. The full method is in what it means to certify an agent.