How do you write a brief for an AI agent?

Write it the way you would brief a capable new hire for one job: what the role is, which sources to trust, which systems to use, when to hand a case to a person and when to refuse. Keep it to one job. Limits do not belong in it: the platform enforces them as bounds.

What is a brief, and what is it not?

A brief is the written description of an agent's job: its role, its sources, its connections and its escalation contacts. Our glossary entry on what an agent brief means gives the full definition. It tells the agent how to do the work well.

It is not a control. A sentence such as "never refund more than 500" in a brief is advice to a model, and models can misread advice or be talked out of it. The refund limit must sit in the agent's bounds, outside the model, where it is enforced at the moment of action. The brief can mention it so the agent knows to hold larger refunds, but the limit holds even if the brief is ignored.

What goes in a brief?

1. Role: one sentence naming the job and who it serves. "Drafts replies to business customers asking about failed payments."
2. Capabilities and target depth: each thing it does, and how far it may go: read, draft, or act within bounds.
3. Sources: which systems and documents it may use, and which wins when they disagree. A system of record beats an emailed attachment.
4. Connections: which tools it may call, and for what.
5. Escalation: which cases go to a person, which person, and what to include.
6. Refusal: what it should decline and how to say so.
7. Style: tone and format, with two or three approved examples.
8. Owner: the named person who answers for it.

The agent brief template lays these out as fields to fill.

Who writes it, and in what order?

1. Process owner: names the job and the owner. Evidence: a one-paragraph scope.
2. Experienced practitioner: writes how the work is done in plain language, including the cases they find hard. Evidence: a first draft.
3. Platform owner: lists the sources and connections with their provenance. Evidence: a source list.
4. Risk or compliance: reviews escalation and refusal rules against policy, for example the firm's AI agent policy. Evidence: review comments, resolved.
5. Agent, depth 1: runs on a sample of past cases with the draft brief. Evidence: outputs where the brief was unclear.
6. Practitioner: revises where the agent misread the brief. Evidence: a versioned brief.

The practitioner matters more than the technologist here. The best briefs read like the notes a senior handler would leave for a stand-in, not like a specification.

What does a good brief look like next to a weak one?

Say a commercial bank wants an agent to help its trade finance team check documents presented under letters of credit. A weak brief says: "You are an expert in trade finance. Check documents carefully and flag discrepancies."

A good brief says the role is to compare presented documents against the credit's terms and draft a discrepancy list for an examiner. It names the credit terms in the trade finance system as the governing source, and says presented documents are external claims to be checked, not trusted. It lists which discrepancies are common and how examiners describe them. It says any case involving a sanctioned country or a goods description that changes between documents goes straight to a senior examiner. And it says the agent never advises a customer on whether to proceed. The second version gives the agent something to be tested against.

Which controls keep briefs safe?

• Versioned briefs. Every edit creates a new version and needs recertification.
• Limits in bounds, not prose. Anything the agent must not exceed is enforced outside the brief.
• One job per brief. Two jobs make two agents.
• Escalation names a role, not "a human".
• Briefs reviewed with runbooks, since the people side is in runbooks for AI agent operations.

A brief usually follows a business case; see how to write a business case for AI agents. The wider sequence is in our guide to deploying an AI agent.

Where Pryme fits

In Agent Studio, you describe the job in plain language and it assembles the brief: role, sources and connections. What it produces is a candidate, not a deployed agent. The candidate then goes through Attest on your own cases, and depth is set by that evidence.