- Home
- Insights
- AI Agent Manager
- How do you write a brief for an AI agent?
How do you write a brief for an AI agent?
Most of the mistakes an agent makes come from an incomplete instruction, not from the model. This article gives a seven-heading brief, a template to copy and a filled-in example.
Why does the brief matter so much?
When an employee is told to “look at the reports”, the employee asks questions: which reports, how often, for whom? An agent usually does not ask; it fills the gap with its own guess and carries on. A vague instruction becomes a misunderstanding in a person and a wrong action in an agent.
The brief is the document that closes those gaps before the agent runs. It is also a management record: when an agent does something unexpected, this is the first place to look.
Which seven headings belong in a brief?
| Heading | Question it answers | Common mistake |
|---|---|---|
| Purpose | What result does this agent exist to produce? | Writing an activity instead of a result: “prepares a report” rather than “brings a summary to Monday’s meeting that a decision can be made on”. |
| Scope | Which tasks does it do? | Fitting three unrelated tasks into one brief. |
| Out of scope | What does it never do? | Leaving this heading empty. A limit that is not written down does not exist. |
| Access | Which data does it read, where may it make changes? | Granting admin rights because it is easier. |
| Stop points | In which situations does it stop and ask a person? | Saying “ask if you are unsure”. The situations have to be listed one by one. |
| Output | To whom and in what form does it deliver the result? | Not describing the format; a different layout arrives every time. |
| Success measure | How is it known that the task was done well? | Setting the measure after the agent has started running. |
A template to copy
The template below can be copied and filled in as it is.
AGENT BRIEF Agent name: Owner (one person): Last updated: 1. Purpose This agent exists to produce the following result: 2. Scope Tasks it does: - - 3. Out of scope Tasks it never does: - - 4. Access Sources it may read: Systems it may change: Authority to spend or send: 5. Stop points It stops and asks its owner when: - - 6. Output Format: Delivered to (person or channel): Frequency: 7. Success measure Baseline (today): Target: How often it is measured:
What does a filled-in example look like?
Take an agent that prepares the weekly advertising performance report.
- Purpose. To bring a one-page performance summary to Monday’s meeting that decisions can be made on.
- Scope. Pulls last week’s data from the ad accounts, compares it with the week before, flags campaigns that moved outside their target, writes no more than three suggestions.
- Out of scope. Makes no change to campaigns, budgets or bids. Sends nothing to a client or outside the company.
- Access. Read-only in the ad accounts and the analytics tool. No write access anywhere.
- Stop points. If data is missing or contradictory, if spend changed unexpectedly against the previous week, or if conversion tracking appears to be broken, it does not send the report; it notifies its owner.
- Output. One page with fixed headings: summary, the three most notable changes, suggestions, notes on the data. To the team channel on Monday morning.
- Success measure. The time it takes to prepare the report and the number of lines that need correcting in the meeting; both compared with the manually prepared report at the start.
How can you tell a brief is well written?
A simple test is enough: give the text to someone who has just joined and does not know the company. If that person can start the task without asking anything, the brief is ready. If they ask questions, each question is a line to add.
The second test is the out-of-scope heading. If the three riskiest things the agent could do but should not are not written there, the brief is incomplete.
When is the brief updated?
In three situations: when the agent does something unexpected, when the task itself changes, and when the agent is given a new permission. The date and reason for the update go at the top of the document. Over time the brief becomes the record of how the agent is managed.
How to choose the first task is covered in the starting guide, and which steps to tie to human approval in the article on human approval.
Our recommendation
The brief is written before the agent is set up. A brief written afterwards does little more than document what the agent has done so far.
- Fill in the template together with the person who does the task today; bring the technical team in on the second pass.
- Do not run the agent until the out-of-scope heading has at least three items.
- Have someone who does not know the task read the text, and add every question they ask to it.
- Give the document an owner and a review date. A document without a date does not get updated.
Frequently asked questions
How long should an agent brief be?
No more than a page. A brief that keeps growing usually means the task was chosen too broadly; in that case the task is split in two and each part gets its own brief.
Who should write the brief?
The person who does or manages the task today writes it, and the agent’s owner approves it. The technical team adapts the text to the system, but what counts as correct is decided by whoever knows the work.
Is a brief the same thing as a prompt?
No. The brief is a management document that people read and agree on. The prompt is the technical counterpart of that document given to the model. The prompt is not changed unless the brief changes.
Does every agent need its own brief?
Yes. Two agents using the same tool have different tasks, permissions and stop points. Shared rules live in a separate company policy; the brief is specific to the agent.
Read next
All insightsWho is accountable for AI agents? The AI Agent Manager role
Agents no longer stop at answering; they act on a company’s behalf. This piece describes the role that owns those actions: what it takes on, what it is not, and where a company starts.
Read the article AI Agent ManagerWhen does the role arrive? Analysis · 7 min readWhen do you need an AI Agent Manager? Authority decides, not the number of agents
Companies usually think about this role when the number of agents grows. What decides it is not the number but the authority the agent is given. This piece covers the signs that the role is needed, and who takes it on at what scale.
Read the article