The brief is where an agent's behaviour lives. Its role decides what work it takes on, its sources decide what it believes, and its connections decide what it can touch. Swap the model underneath and a well-written brief still produces broadly the same conduct. Swap the brief and the same model becomes a different agent.
Why does the model get the credit for behaviour it does not decide?
Model choice is visible, so it attracts the attention. Teams compare providers, argue over reasoning ability and run side-by-side tests on generic prompts. Those tests measure fluency and general skill. They say little about whether an agent will refund a fee it should not, or quote a policy that has since been withdrawn.
That conduct is set elsewhere. A model asked to "help customers with their accounts" with a connection to the payments system will try to help by moving money. The same model asked to "answer balance and statement questions for retail customers, using the statement service only" will not, because it has no reason to and no means to. The difference is in the brief, not the weights.
This matters because the model is the part you control least. You rent it, it changes when the provider changes it, and you cannot inspect why it chose one word over another. The brief is the part you write, version and own. If behaviour lives there, behaviour is something you can govern.
What does each part of the brief actually decide?
A brief has three load-bearing parts. Each one constrains a different kind of mistake, and a gap in any one shows up as a different failure in production. Our explainer on what an agent brief means defines the term; the table below shows what each part settles in practice.
| Part of the brief | What it decides | Failure when it is vague |
|---|---|---|
| Role | Which requests the agent accepts, which it declines, and who it hands off to | The agent answers questions outside its job because nothing told it not to |
| Sources | Which facts the agent may rely on, and how much weight each carries | The agent cites a stale document or treats a customer's claim as fact |
| Connections | Which systems the agent can read from or write to, and at what depth | The agent can take an action nobody intended it to take |
Notice that none of these rows mention tone, phrasing or style. A prompt can shape wording. The brief shapes scope, belief and reach, which are the things a regulator, an auditor or an injured customer will ask about.
How does the role set the agent's boundaries?
A role is a job description with edges. "Collections assistant for consumer loans in early arrears" tells the agent what it is for and, by implication, what it is not for. It should not discuss mortgage switching, give debt advice or change a payment date beyond the plan rules. Each of those exclusions is a behaviour.
The role also names the escalation path. An agent that knows it hands vulnerable customers to a named team behaves differently from one that only knows it should "be careful". You can see the difference when you test an agent's escalation behaviour: the role-bound agent escalates for a stated reason, while the vaguely briefed one escalates at random or not at all.
Roles also make accountability possible. If the role is "drafts dunning letters for an operations manager to approve", there is a named owner for every letter. A prompt cannot carry that, because a prompt has no owner of its own.
Why do sources shape behaviour more than reasoning ability?
An agent reasons over what it is given. Give it the current fee schedule and a strong or weak model will both quote the current fee. Give it the superseded schedule alongside the current one, with no indication of which is authoritative, and a stronger model may simply argue more persuasively for the wrong one.
So the brief should say which sources are in scope and how each is classed. A policy issued and signed by the compliance team is not the same as a note pasted into a ticket by a customer. When the brief records that difference, the agent can weigh facts correctly, and the platform can refuse actions that rest on weak evidence. When the brief is silent, every fact looks equal, and behaviour becomes a matter of which document the model happened to read first.
Why do connections matter most of all?
Connections turn words into effects. An agent with read access to a ledger can describe a balance; an agent with write access can change one. Most of the harm an agent can cause is set by its connections, not by anything it says.
A good brief lists each connection with its purpose and its depth: read, draft for a person to send, or execute within bounds. That list is also the smallest useful description of the agent's risk. If a connection is not in the brief, the agent should not have it, and the platform should refuse any attempt to use one. Writing the list forces the question every review eventually asks: why does this agent need to touch that system at all?
Where Pryme fits
In Pryme, Agent Studio assembles a brief from a plain-language description of the job: the role, the sources with their provenance classes, and the connections with their intended depth. The result is a candidate, not a deployed agent. Attest then replays that candidate against cases from your own operation, so the brief is proven on your work before it gains any depth.
What should you do differently?
Spend less time comparing models and more time writing briefs. Start from a structured agent brief template rather than a blank page, and treat each section as a control. Version the brief, review changes the way you review a policy change, and recertify after any change to role, sources or connections.
When an agent misbehaves, look at the brief before you look at the model. Ask whether the role allowed the request, whether a source was missing or mis-classed, and whether a connection gave the agent reach it did not need. The guide on how to write a brief for an AI agent walks through each part. For how the brief fits alongside enforcement, certification and the record, see the governed agentic operating system, explained.
Questions people ask
Does a better model make the brief less important?
No. A better model follows a brief more faithfully, which makes the brief more important, not less. It will also act more confidently on a poor brief, so gaps in role, sources or connections show up faster.
Can two agents share one brief?
They can share a template, but each deployed agent should have its own brief with its own owner. Two agents with different connections are different risks, even if their roles read the same.
How often should a brief be reviewed?
Review it whenever its role, sources or connections change, and at least as often as its certificate is renewed. A brief that has not been read since launch is likely to describe an agent that no longer exists.