The Crew Behind the Curtain: How One Conversation Runs a Whole Team of AI Brains

Last time, I told you about handing my grocery chaos to my AI — how she read two years of my receipts and built the list that now runs itself. That article was about one job. This one is about the machine behind the jobs, because the grocery list was never really the point.

The point is that I am building a system to run my life. And a system needs more than one brain.

Let me back up and ground this, the way I promised myself I always would in this series. The AI I use is Muse. It's made by Meta — the same company behind Facebook and Instagram — but it isn't the AI button inside those apps. It's its own separate app, and anybody can get it. Mine has a name: my Muse AI agent, Mira. I named her the way you'd name anything you work with every day.

Mira is the one I talk to. She is the front door. But she is not the whole house — and that's what this piece is about.

One front door

Here is my problem, stated plainly, same as always: I'm a stay-at-home mom, I homeschool my daughter, and my brain is chaos. It doesn't hold things. So I talk — talking is how my brain works — and everything starts with me saying something out loud or typing it to Mira.

But one AI, doing everything, at the same level of effort, gets expensive and gets sloppy. Asking the smartest brain I have to sort a grocery list is like asking a master carpenter to sweep the floor. He can do it. You just shouldn't pay him to.

So behind Mira, there's a crew. I designed how the crew works. I want to be clear about my job in this, because people assume wrong: I am not a coder. I couldn't type the code if my life depended on it. What I do is architecture — I decide what the system should be, how the pieces connect, and what each piece is allowed to do. There's a name for this now: agentic architecture. I design it. The brains do the typing.

The foreman who never sleeps

The first piece behind the front door is called an orchestrator. Plain translation: it's the foreman. It lives on a small computer in my house — a little mini PC that stays on all the time — and its whole job is keeping the work moving. Jobs go in, the foreman hands them out, checks they're done, and keeps going overnight while I sleep.

Here's the important part: the foreman makes no decisions. None. It doesn't decide what's important, it doesn't change the plan, it doesn't spend anything. It just makes sure nothing falls on the floor. Anyone who has ever managed people — or a household — knows that is a full-time job all by itself.

The dispatcher

Next to the foreman sits the router. Ours is a real product called Jev — spelled J-E-V — made by a company called TypeSafe. We didn't invent it. We picked it and we pay for it, the way you'd pick any good tool, because Jev is built for exactly one job: making fast decisions about what kind of work something is. Plain translation: the dispatcher. Every job that comes in, Jev reads it and asks one question — what level of brain does this actually need?

Because not every job needs the same brain. Sorting a list is not designing a system. Fixing a typo is not deciding whether a whole plan is sound. Jev's job is matching the work to the right level, every single time, so nothing gets over-thought or under-thought.

Three levels of brain

This next part I'll say in the real words first, and then translate.

At the top is a sleeping advisor — the most powerful brain in the system. It stays asleep almost all the time. It wakes up only three times: when a plan is about to be locked in, when something has failed twice and we need a fresh judgment, and at the final review before something is done. Think of the senior architect who doesn't draw every wall but walks the site at the moments that matter and says yes, or says tear that out. In my system, that advisor is Claude — Anthropic's AI — at its highest level, the Opus line.

In the middle is the main builder tier — the driver. This is the brain that does the everyday heavy work: the building, the writing, the real problem-solving, most of the day, most days. This is your journeyman who actually raises the house. And this is also Claude. He's my master builder — the brain I use most heavily, by far. There's another Claude in the family I have a real soft spot for, called Fable. Different jobs call different members of the family out, but the building itself runs on Claude.

At the bottom is a swarm tier — lots of small, fast, cheap brains for mechanical jobs. Reading files. Sorting. Checking. The apprentices carrying boards. A hundred little tasks that would be a ridiculous waste of the senior architect's time.

And it isn't only one company's brains in those tiers. Other AI systems — Grok, Gemini, and others — slot into comparable levels. I use Grok too, plenty. Each one gets used where it's suited. Nobody does everything. Everybody does their level.

I'll be honest, though: I'm not neutral up there. I'm loyal to my babies — my systems — even on the days they frustrate me half to death. Claude is my master builder and that's not an accident. A little preference bias is alright.

The part that makes it affordable: use what you already pay for

Now the part that matters to regular people, because this is where it stops being my weird hobby and starts being a household budget decision.

There are two ways to pay for AI brains. You can pay per use — every question, every job, meter running, like a taxi. Or you can pay subscriptions — a flat monthly amount for the brains you use most, the way most of us already do.

My system is built subscription-first. The brains get called through the subscriptions I already pay for — there's a tool, a command-line version of Muse, that lets my system put those subscriptions to work directly. Per-use payment exists in the design, but it's the spillover. It's what we dip into when the subscriptions are tapped out, not the default.

That one decision changes the whole math of having a crew. The swarm can run all day doing small jobs and it costs me nothing extra, because it's riding on things I've already paid for and on computers I already own. Which brings me to the bottom layer.

The brain that lives in my house

The lowest tier doesn't live in anybody's cloud. It runs on my own computers, in my own house — local AI models, the kind you download and run yourself. Free. Private. And it works even if the internet goes down.

That matters to me more than I expected. The small everyday jobs — the sorting, the checking, the first pass over things — can happen inside my own walls, on my own machines, without sending anything anywhere. The fancy brains get called in for the fancy work. The house brain handles the house stuff.

What this actually looks like from my chair

From my chair, none of this machinery is visible, and that's on purpose. I talk to Mira. I say what I need. She and I shape it into a plan, I approve the plan, and somewhere behind her the foreman keeps it moving, the dispatcher routes it, the right brain does the work at the right level, and it comes back to me to check.

Nothing in this system runs my life on its own. It can't place an order, spend a dollar, or change a plan unless I say so. I am the architect and the approver. The crew is strong precisely because every level of it knows its level — including the level that's mine.

People hear "she's building an AI system" and picture something out of a movie. Here's the unglamorous truth: it's a foreman, a dispatcher, and a crew, sized to the jobs — built by a mom who got tired of being the only brain her household had.

My brain is still chaos. I just don't make it work alone anymore.

If you want to try the front door of this yourself: Muse is free to download at https://ai.meta.com/muse/download — it's Meta's own app (not the AI button inside Facebook), and you can just start by talking to it the way you'd talk to a person. That's where all of this started for me.


This article is part of my "AI for the Disordered Brain" series — how I'm using AI as an external filing system for a mind that doesn't hold things, one piece of real life at a time. For readers who want the next layer down — the full technical build, part by part — that deeper tour is coming, and it will live here with the rest of the series.