My chief of staff is a team of AI agents

At ten to four every morning, a Mac mini starts working. It pulls the day’s news and research into a briefing. A little later another job takes the next book on my reading list, runs it through a fixed set of questions, and turns the result into an audio review I can listen to on a run. Once a week a different job studies a few leaders I admire and writes up the rules they lived by.

I call the system BEAST. It’s a set of AI agents organised like a small company. An orchestrator plays chief operating officer. A handful of department heads look after knowledge, money, brand and tools. An expert council of three reviews everything that matters: one checks facts against sources, one rechecks logic and numbers, and one looks for blind spots and weak writing. More than fifty scheduled skills hang off that structure.

I didn’t build it because AI is fashionable. I built it because I wanted capacity, not activity. Here’s what it has taught me so far.

Give agents job descriptions, not prompts

Most people use AI like a very clever search box: ask, read, forget. The step change came when I treated agents the way I’d treat new hires. Each one got a role, an owner, a schedule, defined inputs and outputs, and a clear definition of done.

That sounds bureaucratic. It’s the opposite. Clear roles are what let a team, human or not, work without you hovering over it.

The builder never grades its own work

This is the most useful rule in the whole system. Anything an agent produces goes to a different agent before it reaches me. Facts are checked against sources. Numbers are recalculated from scratch. Claims are argued against.

A model reviewing its own output is like an employee approving their own expense report. It’s not dishonest; it just can’t see what it can’t see.

Memory is a liability until you prune it

Agents want to remember everything, and for a while I let them. Then one morning the system wouldn’t start. Old session files, 4,354 of them, had piled up until they crashed it.

The fix was boring and important: a janitor that deletes anything older than two weeks, hard ceilings on how large core memory can grow, and a written post-mortem so the lesson lives in the system rather than in my head. Most organisations have the same problem with their documents, their meetings and their processes. They just don’t crash as visibly.

Every automation needs an off switch

Each new capability ships with a kill switch: one setting that turns it off without touching anything else. That’s what makes it safe to move quickly. Speed is only reckless when you can’t undo it.

The last twenty percent is mine

Agents are very good at gathering, structuring, drafting and checking. They are not good at deciding what matters, what to refuse, or how something will land with the person across the table.

So the rule for everything they produce, including the first drafts of pages on this site, is the same. It doesn’t go anywhere until I’ve rewritten it and added a judgment only I could make.

What I’m testing next

Different models have different temperaments. I’m weighing a loop where one model gathers evidence and argues against my view, another does the analysis, and a third writes the narrative, all working from one shared folder. If it works, using AI will feel less like operating a tool and more like chairing a good meeting.

The real lesson

The skills that make someone good at running AI agents turn out to be the skills that make someone good at running people: clear roles, honest review, pruning what no longer serves, and knowing when to step in yourself.

That’s why I think “Can you run a team of agents?” will soon be an ordinary question in management interviews. The managers who answer well won’t be the ones who know the most about models. They’ll be the ones who already know how to build a team.