An agent brief is the written definition of an AI agent: the job it does, the sources it may rely on and the systems it may connect to. In a governed agentic operating system the brief is the record that gets reviewed, versioned and certified, so it is where an agent's behaviour is set and later changed.
What does an agent brief contain?
Three parts. The role says what the agent is for and what it is not for, in plain language: "prepares supplier payment runs for the finance team; does not onboard new suppliers." The sources list what the agent may treat as fact, each with its provenance class. The connections list the systems it may read from or write to, and through which operations.
A brief is not a grant of authority. Depth, how far the agent may act, is earned through certification and capped by its role's ceiling.
Why does the brief matter once an agent acts?
Once an agent drafts, executes or moves money, every gap in the role becomes a decision the model makes on its own. A brief that says "handles refunds" with no limit and no exclusions invites the agent to fill the gap.
The brief is also what an auditor reads first. If an action was refused as outside role scope, the reviewer needs the version of the brief that applied that day. That is why we hold that the brief is where an agent's behaviour lives, not in scattered prompts and settings.
How does a brief work in practice?
A team describes the job in plain language. A tool assembles a candidate brief from that description, proposing a role statement, the sources the job needs and the connections it implies. A person reviews and edits it. The candidate is then replayed against real cases from the customer's own operation before it earns any depth above read-only.
Most teams start from the agent brief template; how to write a brief for an AI agent covers the drafting.
What does a brief look like for a real job?
Take a collections agent at a mid-sized lender. Role: contact customers whose payment is overdue, offer the standard payment plans, and pass any mention of hardship to a named person. Sources: the collections policy (signed by its owner), loan records from the core system (signed), income statements issued with a verifiable seal (sealed), notes typed by staff (declared). Connections: read loan records, draft messages in the messaging system, create a task in the case queue.
Nothing in that brief lets the agent change a balance or waive a fee. If the business later wants it to waive small late fees, that is an edit to the brief and a new capability, and the agent has to be re-certified with it.
How is an agent brief different from a system prompt?
A system prompt is text sent to the model at the start of each call. It shapes how the model responds, but nothing outside the model enforces it. An agent brief is a governed record: the platform reads its role scope, sources and connections and applies them at the moment of action, whatever the model proposes. Part of the brief may be rendered into the prompt, but the prompt follows from the brief, not the other way round.
Which terms sit next to the agent brief?
- A pre-trained agent ships with a starting brief for a common job, which the customer then edits and certifies.
- Role scope is the first of the five bounds in the minimum chain, and it is read straight from the brief.
- The governed agentic operating system is the layer that enforces what the brief declares.
Questions people ask
Who should own an agent brief?
The business owner of the process, not the engineering team. They know what the job includes and, more importantly, what it excludes. Engineering maintains the connections and compliance reviews the sources and exclusions.
Does changing the brief change the model?
No. A brief changes what the agent is asked to do and what it may use. The model's weights stay as they are; the agent is requalified on the new brief rather than retrained.