How do you hire an AI agent?
You hire an AI agent by giving it the same six things a new member of staff needs: a role, access, context, a first task, a routine and feedback. None of it needs technical skill. On Monmo you write in plain words, by typing or by voice note, in English or Japanese, and the agent starts work on its own computer.
- Step 1: Write the role. Say what the job is, what good work looks like and what to hand back to you.
- Step 2: Give access. Connect the apps the job needs through your own accounts and choose which ones this agent may use.
- Step 3: Share context. Give it your policies, templates, past examples and the names it will come across.
- Step 4: Set the first task. Pick one small, real piece of work whose result you can check.
- Step 5: Agree a routine. Decide when it works without being asked, such as every weekday at 7:30.
- Step 6: Review and correct. Read the result, say what to change, and allow or decline what it asks to do.
Step 1: write the role
Write the role as you would brief a person on their first morning: the job in one sentence, the tasks in the order they happen, how you want them done and what must come to you. Short and specific works better than long and general. A role that fits on half a page is usually enough to start.
For example: you run the support inbox for Fieldwork, an online store. Read each email, check the order and draft a brief, friendly reply. Refunds over $100, complaints and anything about safety come to me. That is a complete role. It names the job, the steps, the tone and the limits, and the agent can start from it.
Steps 2 and 3: give access and share context
Give the agent the least access that lets it do the job. Apps connect through your own accounts, and you choose which ones each agent may use: a support agent needs the inbox and the store, not the bank. Reading needs no approval. Sending, paying, posting or changing a record shows you a card first, and you can take an app away again later.
Then tell it what a colleague would tell a newcomer in the first week. Share your policies, your templates, a few examples of work you were happy with and the names of customers, suppliers and products. The agent keeps this as background and remembers your preferences, so you do not have to repeat them with each task.
Steps 4 and 5: set the first task and agree a routine
Choose a first task that is real, small and easy to check: yesterday's twenty support emails, last month's bank lines or a list of ten prospects. You learn more from one real task than from a test, and a small one keeps any mistake small. Ask it to show its work, such as the sources or the order it checked.
Once the first task comes back right, agree when the agent should work without being asked. A routine can be a time, such as every weekday at 7:30 or every Monday at 8:00, or an event in a connected app, such as a new email at your support address. From then on the agent starts by itself, whether or not you have the app open.
Step 6: review and correct
Review the agent's work as you would a new hire's: closely at first, then by sample. Open the drafts, check a few facts against the source and read every approval card before you allow it. When something is wrong, say so in plain words, such as shorter, never offer a discount, or put the order number in the subject line.
The agent keeps a correction as a preference and applies it to later work. If the same mistake returns, the role is probably unclear: add the rule to the role or give an example of what good looks like. You remain responsible for what you approve, so do not allow a batch you have not read.
What should you check in the first week?
Check seven things in the first week, and write down what you find. The aim is not to catch the agent out. It is to learn where its work can be trusted by sample and where it still needs you to read every line, and to turn what you learn into clearer instructions.
- Does the work match the role you wrote, or has it drifted into tasks you did not ask for?
- Are the facts right? Open the source, the order or the ledger line for a sample.
- Does it sound like you? Compare three drafts with something you wrote yourself.
- Do the approval cards ask about the right things, and can you decide from what they say?
- Did the routine run at the agreed time, and did the result arrive where you expected it?
- Does it have more access than the job needs? Remove any app it did not use.
- What did it hand to you, and should any of that become a rule in the role?
When should you add a second agent?
Add a second agent when the first is doing useful work and a different job is waiting. Give each agent one role: a support agent and a bookkeeper are easier to brief, check and correct than one agent asked to do both. A Chief of Staff coordinates the team, so agents can hand tasks to each other and bring one result back to you.