Answers
How do you write a job description for an AI agent?
The same way you would for a new hire, on one page, before the agent starts. Harvard Business Review's March 2026 advice on onboarding AI agents says every agent should have a job description that spells out what it is responsible for, where its authority stops and when it must ask for human input. A usable version has seven lines: what the agent exists to do, the person who answers for it, what it does alone, what it hands over for approval, what it never touches, whom it tells when it is stuck, and where its record goes. The sheet only works if the agent reads it before it acts and if it changes when the team decides something else. Arroway is where the sheet can live as a set of rules every AI on the team reads at the start of a session, each with who decided it and when it ends. The install page is the first step.
Last updated October 5, 2026
The seven lines of the sheet
Purpose: one sentence on the job, so the agent and the team can tell when a request falls outside it. Owner: one named person. Alone: the actions it may take without asking, stated as actions and not as intentions. Approval: what it prepares but a person releases — anything that leaves the company, spends money or changes a client's files. Never: what it does not touch even if asked politely. Stuck: whom it tells, and how, when it cannot proceed. Record: where it writes what it did and where it stopped. A new hire's description is read once; an agent's is read at the start of every session, so each line has to be short enough to be read that often.
Where the sheet goes, so the agent reads it
A job description in a shared document is read by the people who open it and by no agent that was not given it. The same holds if you paste it into one tool's instructions: the colleague using another assistant never sees it. The sheet has to live where every AI on the project looks first. In Arroway each line is a rule in the project, with the name of whoever decided it and the condition that ends it, and an assistant that wants to widen its own scope can only propose the change; a person approves it.
When to use Arroway
Use Arroway when you have more than one assistant, or more than one person, and the agent's scope has to be the same for all of them and change in one place. When the owner decides the agent may now draft the client email, that is one recorded replacement that names the line it ends, and the next session of every connected tool reads the new version. Arroway does not enforce the sheet: the tool's own permissions do that. It makes sure the sheet is what the agent reads. The first step is the install page.
What it looks like in practice
A consultancy brings a reporting agent into its monthly close. Before the first run, the finance lead writes the sheet: it exists to assemble the month's figures from the exports; she answers for it; alone, it reads the exports and builds the draft table; for approval, any number sent outside the firm; never, the bank portal; when stuck, it tells her in the project log; its record is the log. In week three the lead decides the agent may also reconcile the card statements, still reading only. She records that as replacing the 'alone' line. The next month a new analyst starts and asks her own assistant what the agent may do. It reads the current sheet and answers, with the lead's name and the date of the change. The old version stays in the history.
Questions people ask about this
- How long should an agent's job description be?
- One page, and in practice the seven lines above. It is read at the start of every session, so anything that does not change what the agent does should come out.
- Can the agent update its own job description?
- It can propose a change. In Arroway, what an assistant writes on its own stays a proposal until a person approves it, so the agent's scope widens only when someone decided it should.
- How do I try Arroway?
- Install it in the tool you already use — the install page has one route for Claude, ChatGPT, Codex, Cursor or any MCP tool — and sign in when the tool asks. From then on, sessions open by reading the project and close by recording what was decided.
Where this is verifiable
Harvard Business Review, 'Create an Onboarding Plan for AI Agents' (Joseph B. Fuller, 25 March 2026), for the advice that every agent have a job description stating responsibilities, where authority stops and when it must ask for human input; the seven-line sheet is this page's own working template, not the article's. Product documentation on this site (How it works, Install), the answer on who is accountable for an AI agent on a team, and the sanction, supersession and expiry-condition rules in the sanctioned product spec. Everything described here about Arroway is behaviour the tools apply today, not roadmap.
https://www.arroway.app/en/answers/ai-agent-job-description-before-it-joins-the-team