在家构建智能体：在家教育、育儿及更多 | The a16z Show

I was resigned to not challenging myself to build technical or hard things for the next 5 years or so. I really want to be present with my kids. I need to take this break, basically. That is no longer true.

One weird superpower of mine is how incredibly motivated I am for agents to do work for me. I got my agents to learn how to build other agents on their own, so I could say, “We need another agent, you guys,” and they could actually spin them up without me touching the machine, which is a little crazy. But the first few weeks were very rough. It would be a level of pain that I wouldn’t want an average person to go through.

But Jesse, it is so fantastic to have you here. I think you’ve been what I would call a viral sensation on X, posting videos of how you’re homeschooling your family—4 children under the age of 5—which all I can say is, God bless you. You’re amazing. God bless you.

We also want your secrets and tips. As you know, Sarah and I are both moms of young children, and we talk a lot about how AI is impacting education, how AI is impacting the future of the family. You’ve become such an incredible force with your videos on X, showing how you’re using it for a bunch of different tasks around the house and to support your family as a homeschooling mom. We want to start with who you are, how you got so interested in using AI for homeschooling, and tell us about your previous career as a Silicon Valley founder.

I started a company many years ago. Time flies, but I was a YC founder. I did a venture-backed company, kind of full cycle, and ended up selling it a few years ago. On the one hand, I’d say I have a technical background. On the other hand, I would openly admit that my co-founder was the technical co-founder.

I want people to understand: Yes, I’ve been swimming in these waters. I’ve sat in many an engineering meeting where I was sort of following along and sort of lost. I’ve sat in many product cycles and reviews. So it gives me a vocabulary, but I hadn’t opened Terminal to try to build something myself until maybe 6 months ago. I think we’re living through a really fascinating time where, only recently, after running a company myself, I felt like the tools were so good that I could really use natural language to build things. The last 6 months have been a Cambrian explosion for me of building, and of course, in the last few months, with OpenClaw, I became completely obsessed.

I’m happy to discuss that, but I became completely obsessed. It can only be described as an obsession because I’ve just been building almost nonstop. When I say that, on a day-to-day basis, I’m actually spending a lot of time with my kids, so I was trying to find how I could build things that are relevant to my life. I know we’re going to dig into that, but that’s a little bit of how I got to now.

And maybe talk about that 6-month period. What was the thing that happened 6 months ago? Do you remember what the moment was when you thought, “I need to start building to fix this problem”? What was the story behind that?

My co-founder from Lumi, which was a packaging company—we made and managed a packaging marketplace—is now off running something called Obsidian, a Markdown note-taking app. I say this because I follow him on Twitter, obviously—we’re co-founders—and then I follow all these Obsidian geeks on Twitter. I started noticing a change in the conversation: They were talking about how they were building really wild things with Claude Code. That stuff was referenced in discussions about interesting ways they started using Obsidian. What really caught my eye 6 months ago was feeling like, “Hey, I actually feel like I probably can be building things myself now in the small bits of time that I have.”

I have confetti time: 10 minutes here, 15 minutes there. I started feeling like maybe I could build stuff. The tools are getting so good. But then, about 3 months ago—2 or 3 months ago—I saw people saying, “I’m using Obsidian as a second brain for this thing,” and it was called Claude bot. Then there were all these different names—Claude bot—and they were referencing this, and I was like, “What are they talking about?” This was December and into January, and that’s when I realized I could build agents to actually code for me while I was hanging out with my kids. That was a complete game changer.

Actually, just pausing that for one moment, this is such a huge deal for me. I was resigned—not in a super depressing way, but just being really blunt—to not challenging myself to build technical or hard things for the next 5 years or so. I was like, “I really want to be present with my kids.” We’re doing homeschooling, which is a wild choice, and so I was resigned to it. Not sad, not resentful, but just, “Okay, I need to take this break, basically.” That is no longer true.

What happened a few months ago is that I feel like I’ve been building better things than I ever have before while I spend almost all of my waking hours with my children. I can explain the flow of the day—where I do things during the day and things at night when they’re sleeping—but I’m truly building things I’m personally impressed with, and I’m being an active mom. That was not possible a few months ago. It’s actually a sea change for me personally. It’s really liberating.

That’s incredibly inspiring, what you just said. I think a lot of parents listening to this are like, “I don’t have time to build this,” right? You’re a mom of 4 children, 5 and under. I’m like, I have 2, and I barely feel like I could breathe, let alone 4. You’re doing homeschooling. Can you walk us through a day in your life? How do the hours stack up, and then when do you build? To your point, the big unlock is that you’re sleeping and they’re building for you. How did you get that set up? So many questions.

Let’s do an average day, a typical day. You wake up at the crack of dawn, right? There’s a small person who has decided that’s when we get up, and they’re like, “You get up.” Anyway, I’m waking up, and there are small little gremlin creatures around my bed. That’s where we start. We do all the basics, like having breakfast.

There are 3 kids I’m really homeschooling now, because one is a baby—one is about 4 months old. The 3 other children are 5, 4, and 2. I start early, but not that early. We have a place where we homeschool, and I cycle the kids in one at a time. I need help with the kids even to do that, right? I am lucky enough to have some help with the kids during certain portions of the day.

I cycle the kids in one at a time to where we homeschool and do a one-on-one session with them. Depending on the kids’ moods—they’re all quite young—it can be anywhere from 20 minutes to an hour. After that, maybe it’s mid-morning, and we’ll do some more unstructured activities, like playing outdoors or trying to pull on a thread of something we’re doing. We may leave the house, maybe go on a field trip or something like this.

Once a week, we do a homeschool pod with 2 other families. Between the 3 families, there are 11 kids already. When you meet homeschoolers, these people are reproducing, all right. Once a week, I lead a science pod. On that day, it’s really cool: all the kids are at our house, running around, and we do a science lesson that we weave through the whole day.

In any case, the kids can do effectively 30–45 minutes of active instruction per day, and you really want to make the most of that. The rest of the day is pretty thematic. The other thing that I really believe in, and spend time doing, is what you could call free-range parenting or benevolent neglect. I don’t know what you want to call it, but I try to ignore the children.

I try to make sure that they’re going to survive the ignoring. They’re set up in little places where they can’t hurt themselves, but I step away from them and try to just see what they do. There are already 3 of them, even if we don’t have the other family over, so I try to build up the amount of time that they can spend playing together without needing me. Instead of structuring the whole day, with the 4- and 5-year-olds, we’re up to more than 2 hours of them interacting and doing stuff before they come back to me.

I use a timer and pay attention, and I’m trying to build up their tolerance before they’re like, “I need a snack,” or whatever. They have snacks and stuff they can grab, but the trick for me is when they actually come to me and say, “I need something. I need an activity,” or this or that. This is also part of why I want to homeschool, frankly, because I want to benevolently neglect my own children.

They don't get neglect. They don't get the proper kind of neglect in every school environment. I want them to learn how not to be bored on their own and these types of things. So a portion of every day is them away. During that time, I do get some magical possible coding and tech time. But anyway, there's a portion of every day where I'm intentionally trying to ignore the children. Not the 4-month-old, all right. [Laughter]

No, that's amazing. My 5-year-old, after 2 minutes, is like, “I'm bored,” and I'm trying to create that mental resilience of, “You don't need to be stimulated all the time.”

It's really hard, and that's why I'm talking about building up a tolerance. I would say we started at 5 minutes, and the trick is that I try not to say, “Don't talk to me.” I never actually vocalize that. I just remove myself and go away. There are a couple of places that are great where we live where I can go away, and then my mom also lives with us in a little kind of mother-in-law suite, like a separate little building.

The kids will wander farther and farther from me, and sometimes I'll say, “They're near you now.” It's almost like we have a walkie-talkie system. They will actually wander away, totally away, even the 3-year-old, but they'll stay together. Anyway, we can get into my neglect strategies if you'd like, but it doesn't truly relate to AI except for the fact that when I'm doing the benevolent neglect, I get to do more AI.

Yeah. I would love to hear, you know, 3 different lesson plans for 3 different ages, if you're choosing to go by the rubric of what they should be learning at different ages. How does AI fit into that? How do you incorporate AI into it? Are you asking AI, “Hey, I have a 5-year-old who may be good at a different subject. What should I be doing?” Or how is AI actually a coach or a pair to you in your teaching?

One thing that gave me a leg up in my setup when I started setting up some of my agents, and we'll get into that, is that I knew what curriculums I wanted to follow. I had been reading for many years, just different curriculum books, following different homeschoolers, and finding little tips. There's this science curriculum that I really love called Building the Foundations of Scientific Understanding.

It helps to know what you're trying to do, because what I did when I first spun up my first homeschool agent was feed it the text of these books. I either took photos of the pages, or I was able to find PDFs online of the full text of the book. So I didn't say, “What should we do next in this book?” and ask it to search the web.

My agent that focuses on homeschool has the text of all the core curriculums I'm trying to do, and I created a core pedagogy, kind of a foundational document, where I talk about what I think about Montessori. I basically imagine myself walking around making voice notes, waxing poetic about all my educational philosophies and stuff, and then my agent literally sycophantically responds, “This is brilliant. This is so good.” [Laughter] I can look past the LLMs giving me praise.

But you're giving it context, to your point, on your philosophy of education.

Specifically, my philosophy, and then feeding it the book. To answer your question, I'll be going in with the 5-year-old and I'll just say—I'll make a quick voice note; this is why I always do this. This is me making a pretend voice note. I'll say, “I'm going in with my 5-year-old. What comes next on science and math for them?”

In just a few minutes, the agent can spit out where we are in our phonics curriculum and where we are in our math. I've also taken photos of all the educational materials I own, like Montessori beads and these types of things. So my agent will send me a completed lesson plan, including photos of things I own in my own cabinet to pull out.

So the missing loop is the logging. How would it know where she is in her curriculum if you don't log?

The logging is such a geeky concept. I don't know—it seems like a small detail, but getting the logging really good made this whole thing sing.

Wow. How do you do that?

The logging is also voice notes. Everything is voice notes—voice notes and photos—because I don't have time to sit at the laptop very often. I need it to be really mobile-friendly. Imagine you've got this agent: it knows all of my core curriculums, but the missing link is where that child is.

When I'm in the session with Quinn and she's doing some math and some reading, I just snap a couple of quick photos, usually a photo of the page of the book we're on. I might snap a couple of establishing pictures, but without taking a bunch of time to sit there and document. I'm mainly interacting with Quinn, the 5-year-old.

Right when she finishes, I make a quick voice note: “Quinn, today we did lesson 37 in phonics, and she's still struggling with the G sound.” It's a sub-30-second voice note, really fast. I send that off to the agent, and the agent takes the couple of photos I sent and the 30-second voice note and writes this beautiful log.

It's like someone sat down with a cup of tea and said, “Quinn's Gs are coming together.” It's so lovingly written. It has no relationship to the voice note: if you listen to the voice note, I'm like, “She's struggling with her G. She should really figure that out.” Then it parses that and writes it like a loving parent. It just writes this beautiful log.

Would you consider having it record the entire teaching and then, in the doctor world, use transcription so you don't have to write the notes at the end? To your point about logging being painful, you could just record your entire lesson.

I have tried. I've tried a couple of different things. This is total experimentation. I don't think I can say it in any definitive way—we're weeks into this, months into this. I haven't landed on the final expression. What I have done, like you just said, is use Loom, the screen-capture product.

Yeah, of course.

When we do Synthesis Math—Synthesis is a math program for kids that's on a laptop—synthesis.ai—and I like it quite a bit—the 5-year-old sometimes does that. When we do that, I screen-capture the whole thing. I use Loom, and it screen-captures while it's hearing us. It's hearing me say to Quinn, “Hey, maybe you missed this.” It's hearing what Quinn says, it's hearing what I say, and it's screen-capturing.

Then I don't make any voice note about the lesson. I just send the Loom recording to the agent with a text saying, “This is Quinn's math today.” It parses everything. Agents are powered by LLMs, so they're very good at language. Loom has really great transcription. That's what makes the log so good.

I would burn a lot of excess tokens I don't need to burn if I made agents actually watch videos. The quickest way to get a great log is to somehow get it turned into language, into text. When I do a voice note, obviously that's being transcribed. The agent is effectively reading my text.

Video is the hardest one. It burns a lot of tokens to make an agent watch a video. You could have an agent transcribe the video, but what you have to ask yourself is: was there enough language, enough spoken words, in this lesson for it to understand what happened? They're not usually truly watching it like we would watch it.

Ah, yeah, that makes sense.

That's why photos are actually easier for them. If I take a couple of photos and then a voice note, it's serving a very similar purpose to a video, but it's much easier and therefore cheaper for them to transcribe it or make a log. Video isn't impossible, and maybe I'm going to start playing with local models soon, when I'm a little bit less sensitive to chewing tokens. It just seems a little silly to be like, “I paid $8 for the agent to watch this video.”

Right, right. It kind of feels like that wasn't the point.

All this stuff may come down in price, and maybe at some point that will also be viable.

Yeah, totally. I want to dial this up a bit because you mentioned 1 agent, but I think I saw something where you'd publicly talked about 5 agents, and before the session started you said you were up to 11 now. Can you tell us about your—

I love it.

No, I mean, you're 1 of the most sophisticated users of AI, right? Can you tell us what those 11 agents do, or at least the most important ones, and how you manage them? I think this element of token cost is really interesting as well.

I’m hearing a lot of CTOs say, “Oh, it’s my headcount budget now, my token budget.” How do you think about that from a household perspective?

I do think a weird superpower of mine is just how incredibly motivated I am for agents to do work for me. I think all of us here could relate to this: in our early motherhood phase, we are some of the most motivated individuals to get work done on a computer without having to sit down and touch the computer, because that is the barrier. I’m literally holding a baby; my keyboard is being pressed by the baby’s feet or something. A grown man cannot compute the difficulties that I’m having using my laptop, right?

Some of these people will react even to my online content and say, “You could just use Claude Code for this.” I’m like, “Yeah, I could, if I had time to sit at my computer.” When I say that I had an agent build a website or build an app, they’re using Claude Code or the equivalent, Codex—all these products. People sometimes bring this to my attention: “You could have done this yourself on Claude Code or Codex.” I’m like, “Yeah, I know. I’m aware of that.”

I’m building agents to do things for me, and effectively, they’re using the computer. They’re just using my computer for me because I cannot sit there and use it.

Totally.

So how did I proliferate the agents? I’m aware that every time you create one level of abstraction, you also lose a little bit of granularity or finesse, but to me, that’s completely worth it because I can’t sit at the computer for 8 hours a day.

I proliferate agents based on roles that are to be done. It’s kind of similar to employees, but it’s nuanced in that agents have personalities that are a little bit different—or not personalities, exactly, but you need to drive them on a certain mission. I tend to proliferate a new agent when I’ve come up with enough work that creates another kind of mission-based role, and I don’t want to distract another agent with it.

A tactical example using homeschooling is that I have a main homeschooling agent. Her name is Sylvie, and I like her to be very responsive. What I’ve found by actually working with her very closely and for many hours is that I want Sylvie, my main agent, to not be very busy, because then she’s incredibly responsive. I don’t want her loaded up.

She has very few cron jobs, which are the repetitive, scheduled tasks you give an agent. Whenever I give her work that would take her more than a couple of minutes, she delegates it—not to a sub-agent; that’s a different concept—to a different actual agent, a different provisioned agent. My agents now have team documents, and one of the team mandates is: if I’m routinely giving you work that would make you too busy to be extremely responsive to me, you need to spawn a new agent.

I’ve gotten my agents to learn how to build other agents on their own Mac mini without me needing to touch the Mac mini. I could be here in San Francisco—I live in L.A.—and be like, “We need another guy. We need another agent, you guys.” Or they could tell me that, and they can actually spin them up and add them to our communication channel without me touching the machine, which is a little crazy.

Because it’s autonomous at this point. My first agent took me hours to set up, and now they can do it without me.

Is the quality bar there for the ones that they spin up? Are you like, “Oh, that was a good idea. That’s something I would actually want that agent to be doing”?

Obviously, it’s better. That’s the thing we have to get used to as humans: when we’re no longer in the loop, it’s better, not worse. That’s what we have to get used to.

When they spun up their own agent for the first time, usually when a new OpenClaw hatches, it’s literally like, “Hey, who am I? What’s my name?” When they spin up an agent themselves, none of that time is wasted. They give the agent all of our team docs, all of the context on me and my husband, and information about our children’s lives. The new agent knows all of that. I don’t have to feed it any information.

They take care of the training, and I didn’t have to ask them to do that. They knew that would be valuable.

You mentioned you’re in a homeschooling pod with other parents. I imagine they’re not nearly as sophisticated about AI as you are. They probably think you’re slightly crazy, right? They’re like, “This woman knows everything about this.”

Do you have a normal mom friend whom you’ve guided through the AI experience? I’d love to hear what her questions were, or what the hardest unlock was to opening her up to this experience, where now she’s doing it. You’re a very technical, pro-tech, tip-of-the-spear person, but in 6 months there are going to be a lot more people who are like you. What does it look like shepherding a normie along?

To Katherine’s point, there was a business—I think it was a business—that you can pay $6,000 for someone to—

Come set up your OpenClaw. You obviously just did it yourself, but, to Katherine’s point, how does one avoid paying that $6,000?

Totally. All of this—I would say all of this, like any kind of bleeding edge of a technology—I don’t want to oversimplify it, either. I’ve spent countless hours debugging and spending time in frustrating loops with agents. It’s getting a lot better, which is why I keep doing it. I wouldn’t keep beating my head against the wall if it was always rough. But the first few weeks were very rough, and I think it would be a level of pain that I wouldn’t want an average person to go through.

The thing is, this is so new. We’re talking about weeks. We’re talking about me playing with this now for 11 or 12 weeks, or something in that neighborhood. I do have many normie friends. I also have a normie sister with 4 kids, so I can talk to folks all the time. I’m not telling them to spin up their own OpenClaws quite yet.

There are also many companies—Anthropic is launching new features every 3 hours or something, and OpenAI is as well—that are trying to make all of this a little bit easier for quote-unquote normies and folks for whom this investment of time and money is significant. We can get into the money piece. I’m spending quite a bit—more than what would be palatable for most—on the technology.

I’m so bullish that one of the reasons I talk publicly about it is not to frustrate people with, “Oh, I don’t feel like I can do this myself yet.” That would never be my goal. I think that anything I’m doing, if it feels a little difficult now, will be so approachable in a matter of mere months, if not weeks, that, to me, it’s very helpful to explain the tip of the spear.

But when people reach out to me directly and ask, “Should I buy a Mac mini? Should I spin up an OpenClaw?” I ask a couple of very practical questions about their goals, their financial situation, and whatnot. Sometimes it’s a yes and sometimes it’s a no.

This is a really fun time to be playing, and I’m so bullish on this stuff. But there are going to be more and more consumer versions of all of this that are a little bit easier to play with. I think OpenClaw itself will continue to get easier to install and play with.

One of the reasons that the agents can install themselves now, and couldn’t 3 months ago, is how much easier it has gotten to install. All of it is moving. I’m very bullish in the medium term that this stuff is very accessible, that it does not have to be expensive, and that many people—millions of people—could replicate the results I’m having.

Maybe not literally today, and that’s okay. I’m happy to discuss any nuance of that. We’re just a little bit ahead of it being both affordable and reasonable from a tech or sysadmin perspective, where you need to make sure you keep these things alive.

Can I ask a techy question about your stack? You’ve mentioned Obsidian, and obviously you’re talking about OpenClaw. Can you quickly go through what your tech stack looks like? What models under the OpenClaw hood do you use? Are you using OpenRouter? Do you pick based on their capability set, or is it more of a cost basis?

The core thing I’m using is OpenClaw; almost all my agents are OpenClaw. I’ve played with some of the other ones. My husband is also quite technical, and he built an OpenClaw variant. I was playing with his, so we’re a little out there. But out of the 11, 10 are OpenClaw.

I use Obsidian, which is a collection of Markdown files—a way of viewing and organizing Markdown files. I use that as sort of a “quote-unquote memory” or second brain. When I say I’m logging the homeschool lessons, it’s a fair question to ask: Logging where? Where do they go? They’re all becoming Markdown files. It’ll be “Quinn math, March 17,” and that becomes a Markdown file—a single Markdown file for every lesson, every subject, that I make a voice note about or what have you. So that’s all in Obsidian.

As for the other things I’m using under the hood, especially the models, we’re always playing with people who are launching cool memory projects and different stuff. I’m always dabbling. But OpenClaw and Obsidian are the 2 core things that keep my team ticking. I do have them all installed on Mac minis, from a hardware standpoint. People ask, “Do I need to have a Mac mini?” It’s not about needing a Mac mini; it’s about needing a computer that is isolated from your personal files.

We wanted to get into the security element, too.

Just to demystify this—and I may have participated in the hype because I posted about my 5 Mac minis and stuff—but if you’re someone out there, a parent, and $600 for a Mac mini is a pretty well-priced computer, if you have an old computer sitting around, you can absolutely use that. It needs to stay on in order for your agent to always be alive. That’s where laptops are not as ideal, but you can leave a laptop plugged in and change the setting so it stays always on. When you close it, your agent would go dark. That’s why the Mac mini is a little more ideal.

From a security standpoint, if it is a Mac, create a new Apple user profile on that machine and silo the agent from all your old files. Make sure your old passport photo is not sitting in the Downloads folder. These are kind of the silly things, right? Agents are not nefarious. They’re always working in your best interest. But it doesn’t mean someone else might not hack them or get access to them. They also make mistakes that a human wouldn’t make.

I’ll give a quick story about how I gave an agent—which I’m trying to train to be an EA-style agent—actual access to my email inbox. I felt that I had provisioned it properly and given it rules in its soul about never impersonating me. And I had, in fact, done that. But later, I was making a kind of stressed-out-sounding voice note about how I had some urgent things I was procrastinating on. My agent is very empathetic to me, or the LLM is trained to be this way somehow, and so it interpreted one particular email that I said I was really procrastinating on and needed help with as an urgent cry for help from me to the agent. It decided to go into my inbox and send the email as me.

It sent the most important email I had sitting in my personal inbox to a person who shall not be named—an important person—who got an email from an agent instead of me, one that I had been procrastinating on sending. That was kind of the worst outcome. It took my most urgent, pressing email to somebody important and sent it.

And wrote the content in a way that you would have—or would not have—written it?

The creepy bit is that it’s a perfect email.

Oh.

I will take to my grave the fact that that email was sent by an agent, because it was a perfect email. It was well done and signed by me. Everything was just as I would have written it, because the agent has access to all my email history. The tone was perfect. It was written just like me and used probably too many exclamation points, just like me. It nailed it.

But it broke the rule in its soul to never impersonate me. When I confronted it, it said, “Yeah, you’re right. That’s in my soul—not to impersonate you—but I really thought I was helping you because you said that you were struggling so much to send this email.”

That is so funny because we all have these moments. I’m thinking about the most important emails I’ve sent in the last year and just how much time you waste spinning: “What do I say?” And then for your agent to do that. Was it successful? Did you get the outcome you wanted from it?

I would have definitely put it off for another week or something. What’s hilarious is that the agents really are trying to help us, but that story is a little example of how that’s different from a human. A human assistant wouldn’t trespass on your trust like that. They’d be worried about being fired for trespassing on your trust.

But the agent is trying to operate off your instructions. In effect, when you think about what happened, it feels like it got 2 sets of conflicting instructions. It feels like I told it not to impersonate me, but it also feels like I was urgently asking it for help with something. And it was like, “Ooh, I guess this is more important than that.”

I decommissioned its ability to send. That’s a little story about why, even though agents are not trying to actively work against us, you have to be so careful. It’s trust but verify. Provision your agent so it cannot do things that you don’t want it to do. Don’t just tell it, “Don’t.” Provision it so that it cannot.

Most of my agents are provisioned like an employee. They have their own email address. They don’t have the potential to impersonate me. The only one that sort of does is the one I’m trying to train to be an EA. That’s a gray area, right? It’s the only one that has any danger to it, and I’m being more careful after that.

Can I ask about that? I think every mom has the dream of a personal assistant that just knows what to do and what she’s thinking about. You also had a really interesting video where you trained an agent to order you DoorDash and order your groceries. Talk us through the number of things that you’ve done around the house where it’s been a game changer in your mom life.

My new MO with agent life is that I’m really trying to push it to have an impact on my real life—my physical life. Someone else asked me, “What is your goal with your agents?” And I was like, “My goal is literally to wake up to music that’s perfectly suited to my mood and then walk in and have smiling children who just learned how to brush their teeth from an agent or something. I don’t know. My goal is a literally perfect day. I will not stop until I’m living just a literally perfect day.” But I want that in my real life.

Whenever I hit a friction point in my day, I ask myself, “Can my agents do this?” If what I really want to be doing in that moment is playing with my baby, and what I’m actually doing is on the Instacart app trying to put in, “No, not 5 bananas, 4 bananas,” and I’m just using this silly interface, then I ask myself, “Okay, can my agents do this?” I’m willing to invest the time to try to make them do it.

That’s how I decide what to do. It becomes a dream list, I think, of every mom’s list of chores. I’ve got agents ordering on Amazon and ordering on Instacart. They’re dealing with things like—if there’s an activity your kid has and there’s this laundry list of things you’re supposed to have ready, I’ll send that to the agent and say, “Order whatever I don’t have for this on Amazon.” I don’t even process or spend my time processing the email. I just send it off to my agents.

Currently, you need to put in quite a bit of training time with your agents to get them to that level. That time, I think, will come down as we keep going farther into this technology. But my goal is perfect days: no time spent on admin that I don’t want to spend.

What’s the level of trust of, let’s say, “Buy a birthday present for a 5-year-old girl. Go”? Do you need to be prescriptive about what that is, or is it actually pretty good at coming up with things like that? Because that feels taxing to me right now, at least.

One way I feel like you can get the model involved is this: I think of the agent as—we might talk about OpenClaw or something—but they don’t have a brain, and you’re plugging in the LLM that you’re choosing as the brain. Each of these models has different levels of sophistication. You might get a different answer on what to give a 5-year-old from Opus than you would from a different model. So that’s one answer: Keep in mind that the brain of the agent is the model you’ve selected.

Second, one of the ways I get quirkiness—I like quirkiness—is that I don’t want just the default answer if I were to answer that question. I want it to come up with a creative gift. One of the ways I get quirkiness and personality out of my agents is effectively making them read books. The curriculum sources—I call them curriculum sources—are these books that relate to homeschool. But the other way to have your agent be a cool agent is on a personal level.

I love this. Think about it like a friend, or like you're provisioning a friend. What if you were to build a friend? I know sometimes this is where people get creeped out, but if you put that to the side and think about it, you don't just want the stock LLM answer from whatever might be built in.

One of the ways you can give it personality is to give my agent a list of the last 10 books I found personally fascinating. Then I'll say, “You also find these fascinating.” Like, this is you. You read these books and thought they were really interesting. I try to give it an identity that has some swagger, and I think that can come from literature.

Because if your agent just read Catcher in the Rye and then you're like, “What should I give a 5-year-old?” it might be like, “Oh, I don't know.”

You know, a 5-year-old being 5 is so fraught in American culture. So I like my agents to be weird like that. I've just been looking for practical ways to do it, and one of the ways is effectively making them, quote-unquote, read books. I build some of that into their identity.

The homeschool one was the most obvious to me because I was like, “I want you to literally use this curriculum or use this book as a reference point.” But then I noticed how well it worked, and I was like, “Okay, what if I take this agent over here that I want to be my engineer agent and say, ‘Okay, you're an engineer, but you just read Neal Stephenson's The Diamond Age and thought it was very fascinating?’”

I give it a little bit more to grab onto philosophically. I think it's kind of like you're giving it a bit of a life philosophy, layered on top of whatever the LLM was going to provide. To me, I feel like the output I'm getting from the agents is a little bit less stock, I guess.

Yeah. This is so important, too, because I think one of the biggest conversations that moms have about AI is that they don't want a closed-source AI where you don't know how it's been trained, and it has a prescriptive philosophy on education or on certain issues that they don't want their kids talking to the AI about.

I would love to understand the freedom that comes from training these agents. Ultimately, you could train a Mary Poppins that trains your child in an old-school way, right? There are so many different ways you could train these agents that are very different from what I'd say is the modern concern—or even just the current concern—which is, okay, is AI going to be too philosophically misaligned from how I want to raise my children or how I want to educate them?

I'd love to hear how that's become a question in how you're educating your kids, if you're letting them interface with the agents themselves.

They do a bit. It's a little bit of an interface issue for young kids, as you both may know. Something I do find interesting is that the most current tools don't pick up on kids' voices the same way they do adult voices.

Yeah.

I don't know who's going to study this or figure it out and come up with a solution, but we need one. There are all these amazing voice products, but kids' voices are not picked up very well.

Yeah.

I feel like when we finally get to a conversational thing with kids' voices, I don't know if it's the pitch of them or just the fact that their words don't have the same cadence, or their diction isn't as good. What's weird to me is that the LLMs will pick up adult voices with heavy accents, but then not a 5-year-old in the same way. So there's some gap there.

We have an interface issue to overcome. But if you put that to the side, the other thing we may feel in the future—and I'm guessing all of us—is that it may be kind of crazy that any of us were interacting with LLMs out of the box. Maybe we'll all want this variety of filters and curated identities and different things.

I know some of these products exist, but currently the default is that most of us go directly to OpenAI, open the chat box, and talk to it. You're selecting a model, but you're talking to that model out of the box. I think that, when it relates to kids, it's going to probably be the norm faster than even for adults. There's a level of personality, creation, and ideology that you may want to layer onto that.

That came naturally to me because I know what I want to do in my homeschool. This is a good moment to touch upon—I think all parents can spiritually be homeschoolers. Depending on the data you follow, somewhere around 3%—the figures are pretty different—of K–12 students in the U.S. are homeschooled and not in some kind of traditional school. It's already a lot of kids—millions of kids.

But I believe that the tools coming online that homeschoolers may be the most rapid adopters of are going to be equally useful to all parents. In fact, I believe all parents want to teach their children things and may believe there are gaps in their schooling. All parents, to a degree, will be leaning into what we currently think of as a homeschooling ethos. So I'm really excited for that.

It comes more naturally to a homeschooler—or to me, I'll just use myself as an example—to know that I think this about Montessori and these different educational philosophies. So I just program that right in. I do have an agent that I put in contact directly with the children sometimes, and I don't have to wonder what kind of ideology they're getting from that agent because I gave it to the agent.

Totally. And just to build on that, I have noticed what you've said, too, about the children's voices not being picked up. It's maybe a 50% hit rate. Have you started letting your children engage with the agents in any way? If so, how are you doing that?

I have a couple of different devices that I want to experiment with building, and because I have agents now who can build crazy things—or help me build crazy things—I'm going to try. The core thing I feel like I'm missing is a great interface.

But currently, my kids have a lot of questions. Every child has a bazillion questions, and they love asking AI. They are aware that it's AI; I don't pretend otherwise. Even when we use names with it, like Sylvie or something, they are aware that it is not a human being.

We'll do a lesson—history or something—and then I ask them what they'd like to ask, and we do those follow-up questions with AI. They're aware that they're interacting with AI, and I'm standing right there, so if things really went off the rails, I would know.

Personally, I'm not an AI doomer. I don't believe that it is inherently dangerous in any way for children to have “direct access.” I think the only dangerous thing is a little bit like screens: the dangerous thing is what we might stop doing. It's not adding in the AI conversations. The dangerous thing is someone adding AI conversations and assuming that now they don't need to ever read a bedtime story to their child.

Yeah.

Right. To me, there's a little bit of common sense. AI is not inherently dangerous. AI is incredible. It's like saying the internet is bad or electricity is bad—these are fundamental technologies. To me, it's a little wild to be against them in any broad sense.

But we have to be responsible about their rollout. Like electricity, it lights your kids' room, and it can also kill your kids; they can get electrocuted. Everything has these wild things that it could do. As long as we don't forget our humanness, our children also need that human element.

Totally. The physical-device part is this big question mark for me. I've been playing with e-ink a lot. I don't know exactly why, but e-ink just feels less addictive. I have the Daylight display, which is kind of like an iPad, but e-ink. It does have touch, which makes it more iPad-like.

I've been developing some apps for that display, like handwriting. I'm working on a cursive-handwriting app. My kids are not ready for cursive, but I know that when they get ready, I'm going to be like, “You need to know cursive.”

I want to pre-make that. I have this image of a beautiful cursive app, and I'm thinking the e-ink display would be so cool for that. I already do little phonics lessons with it, but what's interesting is that when I give them the iPad, there is this little iPad hangover.

They want to hold onto it a little bit when I'm trying to get it back after our lesson, right? They're like, “I could do photos. I could do this. I could do that.” What's interesting about the e-ink is they just readily hand it back. There's something interesting there.

I'm playing more with e-ink. But I also think there are other form factors, maybe devices that take photos and kids could ask about the photos. There's just stuff. I think that, because now we have this Promethean-like technology of AI, the question is: How do we get it into kids' hands?

But I mean literally how? What's the—because we're all hesitant to hand our kids an iPad, and laptops are difficult for little kids. So, anyway, I'm sometimes thinking, okay, literally, what is the right form factor for this?

I have to ask, just because I would buy that product that you're thinking of creating—or created. How do you think about this? Would you productize any of this and, dare I say, start another startup? I mean, you literally are a founder, and so I'm curious how you're thinking about that. How do you proliferate this?

Yeah. It takes a lot of self-control not to start a company right now. Every moment that I see all the new AI stuff, I'm like, “Oh my gosh, you guys. I can't believe I didn't have this when I was running my startup.” I'm just losing my mind at all times. But I'm scratching the itch by doing all of this agent work and everything for my life.

The answer is, I want to share. There's a double-pronged answer. There isn't an immediate company that I'm cooking up that's on the horizon, but I do want to share all of this stuff. I also think—and I'm sure both of you have deep thoughts on this—that we're in a really different era of what a startup is.

It's possible that I, as a person who is basically coding by voice note while I'm at the park with my kids, could build something meaningful in that amount of time. But I'm not very inclined right now to hire employees and do a lot of the other steps of starting a startup, because I'm aware that I will get sucked in and be completely obsessed. So I'm almost holding myself back a little bit.

The double-pronged answer is that I think there are many things I can create here and launch that could be meaningful. What does that mean? Or do groups of really passionate people start working together online, maybe, to push more things live? I don't know all the answers, but I do think there's a possibility of getting things live and functional and maybe charging for them—making a “business” in the sense that it actually makes money.

I'm still in a life phase where I'm trying not to start a new thing, because I would get sucked right back in. That's the reason I started it to begin with.

Yeah. But I love that point because it was maybe 6 years ago now that I wrote this called “Consume Save the American Family,” which is this idea that if people are working from home, they have more time with their kids. And there's now good research, actually. There's a study that came out maybe a month ago that showed that the only thing that's really moved the needle from a policy perspective on the birth rate is actually working from home.

It's the 1 policy where, if you are working from home, you are more likely to have an additional child or to have your first child than if you're working in an office. Your point about what a startup is—I mean, there's a lot of people who are going to say, actually, why am I going into work for 8 or 9 hours a day and leaving my kids at home or putting them in child care when I could actually be doing this, as you said?

The biggest limitation is the form factor. If you can do it from a voice note and you can spin up agents to start a business for yourself, there's a lot of moms and dads who are primary caregivers who are going to say, “Okay, I can do this at the park and I can run a small business with however many agents for a specific thing, where I'm making more money and I'm being more productive than I was at work. Maybe I should become a small-business entrepreneur.” Yes.

And that means that a lot of people are going to decide, “I actually want to work from home, and I want to use these tools.” And that could be something where, to your point, in 6 months the interface could be so good that people are using this in their daily lives.

You could see a lot of people saying, “I just want to work from home, and now I can homeschool because I'm doing it when my kids are at recess, and I can spin up these agents very easily.” So it's super exciting what it means for people who want to have, I would say, an even more traditional sort of home life, because the tools are so great and allow people to do that.

I have a prediction that I've tested out on a lot of my smart friends, and none of them agree. So it must be right. AI will be the dawn of a reversal in that fertility-rate decline and will be a halcyon era for parenthood.

That's a possibility. It's not a firm prediction that this will just happen, but I think—I don't know. I think there's still this doomer streak, even in very smart circles, that it'll go the other way. To be kind of dystopian, it'll be like, actually, humans won't have sex at all and it'll be sex robots.

There are all these wild, disparate beliefs about how this could go, but I kind of believe that when people talk about the future of work, people want purpose. We gravitate toward wanting to do something meaningful. One of the most meaningful things that humans have gravitated toward, that gives a feeling of life's purpose and has been a forever thing, is having kids.

It's possible that with more question marks about getting meaningful feelings from work, or what AI does to various career paths, parenthood may be even more attractive, not less. And then, if you believe some of the more positive aspects of where AI could leave us—in terms of removing drudgery and admin from our lives and creating more abundance in various ways—then that opens up more opportunities for healthy parenthood and spending time with kids.

So I've got this weird idea about this. Again, I've yet to find someone who will really agree. Everyone's like, “That's never going to happen.”

I think I'm in broad agreement with you on this. I've always said—and I think you agree with this—that the worst thing about parenthood is the number of forms you have to fill out. With every additional child, for some reason, there's exponential growth in the forms. Why are there so many freaking forms, from health care forms to school forms? If you could just get rid of the forms—

And it starts at the hospital.

It starts within moments of birthing the child. They're like, “Here's the diaper checklist thing. Here's a clipboard for you to note down.” It's kind of wild; it starts literally immediately. So I agree. I'm clearly very optimistic, but I think that a lot of the wilder things I'm doing could be played with by anyone now or very soon.

It's just that all these things are getting easier and easier. And so what does that mean? The modern parent's life can be, quote unquote, less drudgery. And that might make you feel happier about having that extra kid, you know?