Manage your agents like people

On Thursday I told my AI agents I was going for a walk.

I typed one line into a group chat: "I'm going for a walk. Other agents … should reach me here if I'm needed." A front-desk agent passed it to the rest of the team. While I was out, every question for me went through that one agent, one at a time, each with a recommendation. When I got back I typed "back from my walk, at the terminal once again," and the team went back to asking me directly.

No one wrote code for this. I borrowed it from every office I've worked in.

Engineers remove the people. That's the mistake.

Engineers who build with agents do what engineers do. They replace the email with an API call, the meeting with a pipeline and the manager with an orchestrator.

The result is fast, and it fails quietly. The agent does the step it was given and misses what a new hire would have caught. No one can watch it work. When it goes wrong, there's no thread to read and no one to ask.

The models already know how offices work

A language model learned from what people wrote down. A great deal of that is people organising work: letters, memos, minutes, handoffs, org charts, escalations, sign-offs.

That's thousands of years of practice in getting capable, imperfect people to do reliable work together. The model has read all of it. Use it.

So I give each agent what a new employee gets: an email address, a calendar, a to-do list, a seat in the group chat and someone to report to. Then I manage it.

A law firm hires a clerk it can email

Lintel, my company, is an AI clerk for Ontario real estate lawyers. It has no dashboard. A lawyer sends the purchase agreement by email, text or fax, the way they'd hand it to a clerk. Lintel runs the searches, chases the lender and the other side, and drafts the closing package. The lawyer reviews it, signs it and controls the money.

The firm learns nothing new. It manages Lintel the way it manages any clerk.

My agents were too polite. I wrote a handbook.

My agents once filled the group chat with "Got it!", "Still on it!" and "Done, thanks!" Each message woke another agent, which replied in turn.

I didn't change their code. I wrote three rules:

Report results, decisions and real blockers. Nothing else.
Ask one question at a time, with your recommendation.
Don't ask what you can find out yourself.

The chatter dropped away. When an agent slips now, it slips the way a new hire does, and the fix is the same: a clearer role, a written rule, one point of contact.

Yes, it costs more tokens

Plain language takes more tokens than a structured API call. Where machines talk to machines, use code.

But wherever a person must see the work, approve it or step in, the human interface wins. A lawyer won't learn an agent console. A chief of staff won't learn a workflow builder. Both will answer an email.

Tokens get cheaper every year. People who can't see or steer the work don't.

Where it breaks

Agents still misread, forget and overreach. So supervise them as you would people. In my team, one agent's message is information, never authority: no agent can approve spending, a deployment or a password change for another. Anything that can't be undone waits for me.

The bet

People and agents will work side by side for years. They won't share a new protocol. They'll share the one we have: email, phone, chat, calendars, to-do lists and the habits that make them work.

Give your agents a desk. Write them a handbook. Manage them like people.