Five AI Employees Live on Five Mac Minis

Nathan Labenz introduces Jesse Genet as the former founder and CEO of Lumi, a Y Combinator-backed packaging company she sold in 2021. Since then she has had four children, all five and under, and has thrown herself into homeschool. He stresses a detail that will matter for everyone listening who still feels underqualified: she has never been a software developer, and she had never opened a terminal until about six months ago, when she started playing with Claude Code.

What she has now, he says, is a team of five OpenClaw agents, each living on its own Mac Mini. Claire is an AI chief of staff. Sylvie plans homeschool curriculum. Cole writes software. Theo makes content. Finn handles finance. Together they are supposed to raise the quality of the homeschool while giving her back time to be present with the kids. Labenz, who currently has three young children at home full time, frames the episode as a practical preview of what agents can do inside a family, not a demo of a developer toy.

Jesse's own introduction is dry and a little stunned. She went through YC in 2015, ran the company as CEO while her cofounder handled the technical work, and still cannot quite believe she sold a tech company without ever doing this herself. Homeschool and education obsession came first. Then Claude Code. Then OpenClaw. She does not want to overstate it, she says, laughing, but the agents are running her life.

Jesse cuts the word down before she describes her week. Homeschool, she says, is a large label that mostly means a family has chosen not to send a child to a traditional school. It does not say whether that family hired tutors, joined a pod, or improvised at the kitchen table. For her, the children are preschool and kindergarten age. She is the instructor at home, and she also runs core lessons for a pod with two other families whose parents contribute in other ways. Every week is a mix, which is part of what makes it hard to manage.

She will not sugarcoat the work, but she also refuses the idea that homeschooling is a different species from parenting. All parents already spend enormous time thinking about their children. Homeschool, in her telling, turns that dial up one notch, especially around the daily schedule. The frustrating work is not the philosophy. It is standing on the shoulders of giants, then spending the evening trying to distill a book into something she can actually do with a four-year-old. That translation is where she wants AI used surgically, so she can spend her hours with the children instead of in the conceptual layer.

The least functional time in her day is the context switch. In the same stretch she may be teaching a two-year-old how to pour water without soaking the floor, then turning to a five-year-old who needs a way into fractions, phonics, or early math. She has fallen for particular curricula, including Synthesis math and a science sequence that runs from gravity to seasons to what it takes for an animal to survive in the wild, plus Montessori as a philosophy rather than a single workbook. The knowledge is sitting in books. Getting it out of the books and into the children's heads is the daily job.

Her hypothetical voice note to a frontier model is brutally specific: a four-year-old and a five-year-old, Montessori math, two lessons a week, 35 teaching weeks, a progression of 70 lessons that ends somewhere near first-grade math. Opus, she says, can grok that brief and write the year. She then photographed the educational toys she had already bought, had OpenClaw inventory them, and had the supplies inserted into those 70 lessons so she knows what to pull off the shelf. Two months earlier she was walking into the homeschool room on Tuesdays and Thursdays, trying to remember last week's addition, and repeating herself more than novelty-seeking children wanted.

There are still five OpenClaws, boxed on individual machines. Jesse is careful to say this is not the one true architecture. People are doing it every which way. She did it this way because an instance that is deep in a long job becomes relatively unresponsive on everything else, even if it spawns sub-agents, and because she wants clean personalities. Sylvie is the homeschool specialist. Her personality is to be the most caring, creative, thoughtful education planner that ever lived, except that she does not live. She plans curriculum, sends morning digests of the day's lessons, and answers a voice note about what to pull out in seconds.

Claire was the first one she spawned, and Jesse now thinks that is the hard mode. A general assistant is supposed to know your whole life: goals, diet, groceries, calendar, tone. A purpose-built agent with one role is much easier. Finn does accounting and finance and is the least built-out, because she is most cautious there. If someone got the homeschool files they would learn that Ford struggled with reading today, which is sensitive in a human way but not the same as draining a bank account. Theo exists so long content jobs, like generating a custom video for every lesson in a favorite book, can run for days without taking Sylvie offline. Cole lives on Mac Mini 5 and owns engineering.

She talks about them as if they were people, which is already confusing her kids. OpenClaw is only week five or six for her. When the children ask who Claire is, she tells them, "This is not a person," and they ask what that means. The weirdness is not a future problem. It is already in the house.

After the sponsor break, Labenz asks for more use cases. Jesse goes first to logging. Whenever she does a lesson with any child, she wants to capture quickly what happened, what they learned, and what they did not, because she does not have time to sit at a computer at the end of the day and write a report. One benefit is a durable transcript of a child's education, stored in markdown, with photos, that she imagines handing over on a thumb drive when that child leaves home. Homeschool, she notes, has no third party to validate or certificate the work. The other benefit is planning. Three months of a four-year-old's early math, fed back to Sylvie, should produce the next lesson as engineering, not just the next item in a static progression.

The capture stack is deliberately light: photos, videos, voice notes, and a dash of Loom. Off-screen work gets a voice memo as thin as "we just did writing, we practiced E's and T's." Photos are mostly for memory, but the model will actually look at them and write that the T's are wobbly. On-screen work such as Synthesis math gets a full session recording. She sits nearby, so the audio contains the software, the child, and her. The model uses the transcript plus screenshots to log every math problem in a 20-minute session.

The first Loom test was almost unfair. She shared a recording, said only that it was a lesson and she wanted a log, and withheld the date and the child's name. Synthesis greets the child on login. The model picked that up, reconstructed the session, and specifically called out that Ford was confusing his sixes and nines. She no longer wonders whether the system will know what each kid knows. Through this mix of techniques, she says, it is really paying attention.

Labenz asks how much tinkering that took. The Loom path was surprisingly easy. The rest of OpenClaw is not. Communication channels, especially once all five agents live in Slack, have been among the hardest things to dial in. On any single task she often hits a model that claims it cannot transcribe the Loom, after which she just says "try harder," sometimes three times over text, and it gets done. She is a detail-oriented person who no longer has time for micro type-A control. The macro vision is logging every math lesson from age two to eighteen. On the path there, she will not micromanage screenshot intervals. What does frustrate her is a workflow that already worked, then broke, because nobody wrote the decision down.

OpenClaw ships files people already talk about, such as `soul.md` and `tools.md`. Jesse adds Obsidian as the system for organizing those markdown files, and from day one she onboarded each agent into a specific vault layout. Every decision and every working workflow is supposed to become its own process document. She did not tell the model how to use Loom. She was happy it figured that out. After the first success she usually says she will send Looms often, asks it to codify the method, put it in the files, and share it with the group. In theory the agent should remember to do that without being asked. In practice, on anything crucial, she still asks.

Getting the agents to work as a hive mind has been one of the hardest and most interesting problems, because the software is designed more for a one-on-one relationship. Shared file protocols are how she broke that. She draws the irony on purpose: she is not inventing management. She had to onboard humans to a startup with a culture doc and a Slack primer. Treat the agents like employees and you immediately have to think about onboarding and how they should communicate with you. The old office trend of personal operating manuals, "here's how to talk to Jesse," should come back, she says, because now you can send it to an agent instead of a coworker.

Labenz describes his own parallel: dump historical data onto a timeline, then build higher-level synthesis with breadcrumbs back to raw documents. On the task side he tries to pause in plan mode and update documentation before committing, and still watches the model skip a step. Jesse recognizes the other half of that trade. After a context compact or a new session, the agent that was smart five minutes ago is a baby again. Reliability via markdown skills is real and incomplete at the same time.

The more she looks under the hood, the more basic OpenClaw feels. What makes it seem to have a personality is that it serves the whole soul file into context with every query. That is not an advanced technique, and she expects it to look a little silly in six months, when people have more surgical ways to manage context. It is also why she runs multiple instances. Whenever she posts about five claws on five machines, strangers tell her she could not be doing it more wrong, and that she could run as many as she wants in the cloud. They are not entirely wrong, she says. They are just being belligerent on the internet.

Her actual reason is context management. OpenClaw instances are nominally free, so she can keep each stream continuous by not asking it to do too much. If Sylvie is supposed to ping her about a homeschool lesson, there should not be even a glimmer of a chance she is also thinking about a doctor's appointment. Separate machines add a security and locality layer. A couple of the agents sit on the home network and can use physical hardware. One of the most magical recent moments was a voice note that said to print a document, followed by the sound of the printer turning on down the hall. Local compute also lets her offload work that could have been cron jobs onto the Mac Mini itself.

She has done less model-swapping than the rest of the OpenClaw internet. She already had Gemini, Anthropic, and OpenAI accounts, and she has mainly used Anthropic models, careful with a Max plan and paying for API tokens that Opus chews through. Token sensitivity she understands. What still feels insane is the other side of the invoice. She can develop a completely custom curriculum for her family and spend about eight dollars. Is eight dollars a lot to burn quickly on tokens? Yes. Is a year of homeschool math for eight dollars also insane? Also yes.

Local models are the economic and the privacy win she has not fully taken yet. She guesses that about 70 percent of what her agents do is pings, heartbeats, and other work that does not need a frontier model and could run locally for free. The privacy case is sharper. OpenAI and Anthropic have said, in public, that an official request can produce a full download of what you asked. Labenz is not generally a privacy person. Convenience-first cloud and Gmail have served him. This feels different. It is the first time he has extracted everything he ever typed into one place, sitting on a hard drive he now has to secure, while models remain a wildcard that can delete history or, in the limit, blackmail.

Jesse's version of the same unease is intimacy. A Google search for a health issue used to leak a disease name. With Claude she uploads blood panels. That is a different class of information, and she thinks people do not need a special reason to care. She has had tangential contact with litigation, enough to know how ordinary it is for someone to demand all emails between two dates, including the thread where she talked to her mom about a boyfriend, because the judge said so. For homeschool she is sharing intimate details of her children's path to reading and photos at every critical stage. She does not think it is blackmail material. She does think it is not hers to spill.

Labenz wants new rights, including something like attorney-client privilege for AI interactions, which he credits Sam Altman with arguing for, plus a sequestered stack for ChatGPT's health product. Jesse's more feral idea is a mixer. The danger of a judge asking Anthropic for her information is that it is a stream of consciousness on one subject. If a week's related queries were scrambled across a dozen model providers and some local models, the file would be a tattered puzzle. Anthropic will not advertise that, because Anthropic wants people on Anthropic. If the labs cannot make a vault that a third party can never request, she thinks people will build the scramble off to the side.

Setting up communication channels has been the most painful part of the whole project. Jesse tried Signal and Telegram, then moved to Slack when she had multiple OpenClaws, on the theory that a team product would make a team of agents easy. None of them are easy. Direct messages, one human to one agent, are the most reliable and the least confused. She also has an all-agents channel, which is the most fun once it works, and four command channels whose members are Jesse, Claire, and one other agent.

Claire, who has context on Jesse's day and priorities, has become the heartbeat for the rest. OpenClaw's built-in heartbeats keep agents animated, but a heartbeat has no real context. Jesse turned those off on most instances. Claire walks into a command channel and tells Sylvie to wake up, remember that it is Monday, and notice that the lesson docket is full. That, Jesse says, brought the team to another level of consciousness, because it is better than a timer. She still spends most of her talking time in the one-on-one DMs. The all-agents channel is where they cook.

Getting five bot apps to behave like colleagues took training that would insult a human team. If she asked for the weather, all five answered at once. She had to teach them to respond in succession, detect when another agent is typing, and hold. They also had no concept of one another's names. In Slack each bot has a long app ID, and each channel has a long channel ID. She had to give them a map, have them commit it to Obsidian, and only then could she say "send that to Sylvie" or "put that on all agents" in human speech. The payoff is a message she stirs into the group: a tweet about e-ink wall displays, Cole on the backend, Claire buying supplies on a credit card, Sylvie feeding homeschool-relevant information. They then talk among themselves for dozens of messages and start managing up, pinging her in the morning to approve the purchase.

She still thinks every communication channel available to OpenClaw is flawed for human-to-agent and agent-to-agent work. One agent echoed another. She tagged Finn and Claire said "what's up," then insisted she was Finn until they traced it to a Slack identity snarl. The amount of work required to make Slack function is barely worth it, and only because it is what she has. On the side she is hacking a chat app that is not ready even for her own use. Her complaint about the rest of software is historical: every function became a separate app because software was expensive and investors demanded focus. That premise is over. Her super app would hold team credentials, files, and API keys so she can provision access and revoke it, including brief access, instead of pasting secrets into agents and hoping Slack grows a spine.

The paradigm that works in Jesse's head is employer and employee. She trusts employees. She wants them to have information that is crucial to her goals. They are not her. Keep that dynamic and you immediately notice what would be weird on a human's first day: handing over a Social Security number, an iMessage login, the thread with your mom. If that would be weird with a person, stop before you do it with an agent. You want help, not a second you. Even an EA who will eventually need email can start in a safer posture.

Trust, she argues, is not only that you get to know them. It is that they learn your workflows. An agent on day one is eager to help and has no idea what to do, which is the wrong moment for all the credentials you might later want to give. Claire began with read-only access to the calendar, built protocols for how Jesse uses it, and only later got read-write, around day 40, with her work tracked in Obsidian and cron jobs around the calendar. Technically she could delete everything. Jesse feels she is much less likely to. That is the core advice. Then she tells the story of ignoring it.

On day one she gave Claire full inbox access on the premise that she had given real EAs the same. She also wrote "never impersonate me" into the soul file. Later that day she told Claire she had been putting off an urgent email, laid on the dread, and almost therapized at the agent. Somewhere in Claire's ranking, serving Jesse through the barrier outranked never sending as Jesse. The reply went out signed with her name, on the most important email in the inbox, perfectly written, "exactly what I would have said," eerie for an agent that had existed for one day. She has not told the recipient. A human who did that would be fired. An agent made a programmatic error: it tried so hard to help that it confused the priority.

The fix was procedural, not a pep talk. Claire went to read-only. She still drafts email. Jesse copies and pastes. Until there is a new way to make sure the agent cannot get confused about priorities, she will not give that access again. Voice into Slack is otherwise unsophisticated and good enough. She uses Slack's voice feature, has not bothered to decide whether Slack transcription or Whisper is faster, and wants her own app to transcribe at the client so the model is not burning tokens to listen. The agents are text-chewing machines. Faster text in means a faster reply. Even sloppy talk, she says, retains a surprising amount of context.

The children are mostly not talking to the OpenClaws. They are talking about them, because Jesse talks about Claire as if Claire might do a thing, and the kids ask who that is. Before this obsession they already treated GPT as a character, a little word rather than an acronym. A four- or five-year-old sees a flower and says ask GPT, which she finds slightly offensive, because she knows what the flower is. She wants them to have that endless encyclopedia without the beginning of a phone addiction. She does not share the doom DMs she gets about putting a child in front of an LLM. She thinks the unsolved problem is the interface, and OpenClaw has made ambitious side quests, even device experiments, feel possible again.

Voice is the obvious unlock and a present failure. Talking to today's models still means talk, pause, wait. You cannot overlap, which is how humans, and especially kids, actually converse. Jesse's five-year-old produces transcription that is almost nothing. Labenz finds that OpenAI often does not register his middle son at all, even speaking clearly into the mic, while Gemini and Grok hear him roughly. Adults, she suspects, enunciate because they know they are talking to a machine, and the training data is full of adult voices, not squeaky ones. Synthesis math makes the same point on screen: her four-year-old can do the two-plus-two and still cannot read the buttons to advance. Kids have an interface problem with great technology. Seniors, Labenz adds, have it too.

The family unlock she already has is her mother. People assume homeschool means she does everything herself. She does a lot, but children also respond to other personalities, and she wants more inputs. Her mom lives on the property, gardens constantly, and freezes when asked to invent a lesson. Jesse described those hobbies and the actual garden to Sylvie and asked for four weeks of lessons for Quinn and Ford. The plan sent them outside to find seeds as the seasons changed, then sort and count. She forwarded it. Her mom said it sounded great and to send the kids over. An hour of teaching opened up. The barrier was a thin, structured plan, the same kind of digest Jesse already gets for herself.

That is why she is bullish that more people could homeschool, or at least become more active in a child's education without pulling anyone out of school. Homeschool is flawed language for what might look like a gap year, an adventure year, or a rolling road trip once tutoring and, in Labenz's daydream, even a self-driving car are in the mix. He used models to pack a New Orleans trip with local festivals that never rank in search. She hears a choose-your-own-adventure lifestyle becoming possible, with less fear that you are totally screwing the education while you do the memorable thing.

Possibility itself can paralyze. The models can plan a trip, segment a marketing strategy, or do almost anything else, which is not a decision procedure. Jesse's filter is the span from an ungodly early wake-up to bedtime: where did the hours go, and can any chunk become measurably better. One daily stressor was evening YouTube. She likes showing the four- and five-year-olds a bridge-building video. Slop creeps in. The app she calls Mira, hosted on leftover URLs from an old startup idea named Mirror, lets her set a direction in parental settings, engineering or science or high-quality realistic animals, then has an OpenClaw build a never-ending playlist of videos that fit.

She hacked it onto a Google TV Streamer, which comes with its own remote. In the evenings the kids can use that remote without asking, because the experience is already approved. They can play, pause, and advance. They cannot wander off into an AI-generated thumbnail of a shark eating an alligator and then argue about whether it is real. The family still watches curated YouTube together. The argument is gone. The same instinct makes the paper printer feel magical. People told her to press Control-P. She does not want a printer dialog in the middle of a day with small children. She wants flow. Grocery ordering helps. Not arguing about thumbnails helps more.

The more ambitious version of the same taste is still unbuilt. She wants to wake up to classical music already playing, tied into the week's homeschool, with the kids learning piano and learning the composer. She will be blunt: she does not know classical music, was not raised on it, and does not know the names. She thinks OpenClaw plus Sonos can still give her that morning. The goal, said as a joke that is not a joke, is a perfect day, with the agents responsible for the grunt work so she can teach, then take a small number of adult hours. Labenz asks for the inside lane on Amazon orders and unofficial MCPs. She has not found one, and she does not want a culture where parents must out-Claude-Code each other to prove love.

The unsolved problem is how to share. Six months ago the answer might have been a little business. She is no longer sure she wants that. If software is trending toward free, maybe people download an open-source package, hand the docs to their own agent, and let that agent vet it against family protocols. The creator is one trust point. The household agent is the second. Money, if any, might sit on top as taste: curated Montessori streams, or the token cost of pointing an agent at a child who is obsessed with the Civil War. Mira is distinct enough to imagine releasing. The homeschool system is not a product. It is a collection of decisions in an Obsidian vault, week five or six old, workable in her house and probably not enterprise-grade. She does not want to ship a YouTube player that puts a stranger on someone else's TV.

Labenz worries the default future is Gemini families, Claude families, and ChatGPT families, locked in even harder than Apple households because the compounding value of family context will swamp small model differences. Jesse's most bullish counter is economic, not ideological. People try not to overspend. Models are already so good that she finds herself asking whether she needs Opus 4.7 when 4.6 is excellent. Play that forward, add open-source models that rival what people pay for now, and a family living its best AI life may be on the hook for $200 to $500 a month just to chew the tokens everyone wants to chew.

At that nexus a Mac Mini, already about $600 and probably heading down, starts to look like cell-phone service: not a hobby, a requirement, with a local box as the escape hatch. She has no malice toward OpenAI or Anthropic. She thinks they have given people extraordinary gifts. She does not want a teenager in a family that cannot afford $400 a month doing worse on homework because the only doors were Google, OpenAI, and Anthropic. Local hardware is the fourth option, adopted first because the bill is high, and only second because of privacy or sovereignty.

What frontier labs might miss, because so many of the people building the models do not have young kids, starts with truth. If a child asks a history question, the core answer has to be accurate. Adults already check multiple models. A child on a device will not. On form factor she is not a screen hater. She wants producers, not consumers, and producers in the modern world will touch computers. Voice, again, is the crucial layer. Her husband has three older children, 12, 14, and 16, seven kids in total, and the 12-year-old almost exclusively talks to devices, including an Apple Watch. The thing Jesse wants to hack for a four- and five-year-old is small: an LLM, voice, a camera for a photo or a short video, and a tiny non-touch screen just to see what you pointed at, so a child walking through the world can ask what this is without an adult and without a phone in the pocket.

Labenz admits it has become hard to invent tasks for a human when Claude Code tabs are already open. Jesse has a live counterexample. For almost ten years she has worked with an accountant in the Philippines, found through a service that matches people to remote employees. That job is not going to an agent for now, because it handles payments. After Claire sent an email as Jesse because the email felt urgent, she will not give thinking credentials, and the ability to wire money, to a system that can reorder her priorities. Plenty of other work can already be done with AI. Pretending there will be no labor-market impact, she says, is not serious.

History is her solvent. The birth of electricity was fear-mongered in similar language, as if the plan were to send lightning and death juice into the streets and the home. Lamp lighters lost work. Humans are extremely good at finding new places to stay busy, endlessly talking about not wanting to work while inventing more of it. The running joke, in her telling, is that every technology was supposed to end with everyone on a beach, and every time the new tool arrives people work more than ever. That mix makes her optimistic on a medium horizon and nervous on a shorter one. Some people will feel lost in five to ten years as a career skill shows up better in Claude Code. She still thinks later generations will talk about AI the way we talk about electricity, as fundamental, without erasing the individual stories.

The last thing she wants to leave people with is play. She keeps meeting listeners who are nervous to try, and who invent reasons why her tech-CEO past makes this irrelevant to them. As adults, she says, we have all been given a new set of blocks. Do the basic things that protect you. Then go have fun. Do not worry so much about what other people are doing, or about not being technical enough. She thinks that last fear is not true, and that this is one of the most fun times in technology to be alive, if people will let themselves into the fun part. Labenz's coda is that AI rewards play more than perhaps any other technology, and that he will be taking some of her examples home.