As a teenager, I ran across Robert Heinlein’s classic sci-fi novel The Moon Is a Harsh Mistress. All great science fiction books eventually enter a historical moment where they find relevance above and beyond what they had when they were first written, and that’s inarguably happened here. The 1966 novel explores the emergence of a sentient AI out of a computing system that had reached a level of complexity comparable to the human brain.
We are, to be clear, many steps removed from Heinlein’s vision. But what we do have today is interesting on its own merits. AI researches and synthesizes without breaking a sweat, works with images, audio, and video without much fuss, and when you build an agentic framework, it can do all that autonomously in the background while you do other work.
Like many others in 2026, I took a break away from my routine to check out the possibilities of agentic AI. My work involves lots of research, data analysis, reporting, and detailed-oriented proofreading, and agentic AI can automate key tasks to support my efforts on all those fronts. I still have a lot more exploration to do, in no small part because the field is rapidly evolving.
But already, I’m seeing something important, something that’s resetting all my expectations. Start working with AI closely, and I think you’ll see it, too. It gives you a new way to learn. Let me explain.
The thing that is its own instruction manual
Recently, I had to replace a spindle on my riding lawn mower. I’m not much of a mechanic, but I’m also not afraid of trying to fix things that are already broken. What’s the worst that could happen?
Anyway. I did what good Millennials always do in such situations. I watched some YouTube videos. Found a tutorial from ten years ago featuring a dude in a trucker hat with an impossibly clean garage doing a replacement on a similar mower. Bought some parts. Worked for an hour or two. Learned that the mower’s previous owner had installed the blades upside down. And now my mower is running again.
You know what I didn’t do? I didn’t ask the mower how to fix itself. (I did swear at it a few times. Still waiting for a reply.)
How about sixteen years ago, when I learned how to overclock CPUs courtesy of an AMD Athlon II processor? I browsed a lot of forum posts. I read a lot of how-to guides.
You know what I didn’t do? I didn’t ask the CPU how to overclock itself.
I could keep going, but you see the pattern. For basically everything else that you build or repair or tweak or study, you always have to get instructions from something else. LEGO kits have instruction books. Egyptian hieroglyphics have archaeology professors. Bottles of oil paint have classic painting examples from the Italian Renaissance.
AI is different. To learn how to build an AI, you can just talk to AI. You might need some help getting to that first moment when you have a large language model loaded up in an app like LM Studio, but getting there is easy, and once you’re there, you just ask it questions. And how do you get the help you need to reach that point? Just ask Gemini. Or ChatGPT. Or drop your question in a search engine and let the AI Overview be your guide. AI is the thing that teaches you how to make itself.
There’s just nothing else quite like AI in this sense, especially not in the world of computing. For decades, PC enthusiasts have been learning about things like programming languages, BIOS features, and operating systems through college courses and forum posts and trial-and-error. All that still matters, of course. Vibe coding will only get you so far. But designing an agentic AI system largely just means talking to your agentic AI system. And I mean talking talking, not coding or even really prompting. Just everyday, natural language conversations about the things that you’re working on, the goals that you have, and your questions about what’s possible.
The importance of flexible expectations
Here’s something unexpected that comes out of this very unusual situation, this way that AI can teach you how to build itself. Basically everyone who digs into agentic AI ends up using it in ways that they didn’t anticipate going in.
The discussion threads over at the LocalLLM subreddit are full of stories like this. People will try out agentic AI with an old laptop or mini PC, expecting to use it for something like smart home integrations or keeping an eye on an email account that they’re tired of micro-managing. But then they’ll keep finding other use cases. Before long, their agentic AI system is tracking their calories and macros just from pictures snapped at meal times, managing their calendar, and organizing years of files on their NAS.
In part, these unexpected use cases come up just because the capabilities of AI systems are expanding basically every day. But they also come up just because AI works differently. Everything else you do with a PC is deterministic: input here, output there. There’s no earthly good that comes from asking your Windows “Run” dialog box an open-ended question.
But there’s an incredible number of things that can happen when you bring open-ended questions to your agentic AI system. One of the best things that I did recently was simply have a conversation with my AI about the ways that I had been using it. We talked over the sort of jobs that I’d been giving it, the questions that I’d been asking, and used that as a basis for selecting an LLM better suited for the tasks that mattered most to me. Over time, this iterative process of trying out new configurations, talking through the results, and making adjustments has led to the creation of an assistant that’s uniquely suited for my goals and needs.
Rooted in the history of the home PC
Fellow readers of Heinlein will notice a key difference between my home agentic AI setup and the sentient AI in The Moon is a Harsh Mistress. (Several, actually.) Heinlein’s AI emerged from a massive hardware array. I’m running mine on a home PC.
If you’re looking to explore agentic AI on your own, you’ll have to make a key early decision between those two hardware structures for AI. Will you subscribe to a cloud-based service that operates on massive data center installations? Or will you run your agentic AI on your own hardware?
I’ll have a lot more to say on this subject going forward, but here are some brief thoughts. Both options have their strengths and weaknesses. Going the subscription route gives you access to frontier models, the kind of artificial intelligence systems that home PCs can’t even hope to handle. That level of reasoning and responsiveness and capability might be necessary for what you’re up to, but it also ties you to paying for token usage, and that can get pricey very quickly. Cloud-based AI solutions also tend to leave you stuck in a loop of repeated instructions. It can be useful to have AI capabilities waiting around passively until you ask for something, but when AI can’t carry context forward, you end up reorganizing data, restating context, and rebuilding tasks at every turn.
A local AI system, on the other hand, creates the possibility of a system that’s always working for you. You own the storage, so you own the data, and you can keep your AI working on that data without interruption. You can try new things without wondering how many tokens your agents will burn through. You can start your workday not with a full restart, but by checking in on workflows that never stopped running.
The challenges for local AI start with the reality that your home PC can’t compete with the capabilities of a data center. And if you don’t already have the PC hardware for running agentic AI, you’ll need to make some purchases, and the upfront costs of a new PC will be much larger than the initial cost of a subscription to a cloud-based AI service.
You’ll have to balance those pros and cons for yourself, but I tend to prefer ownership over rentals. Cloud-based AI is an easy way to find out what’s possible without making a big financial investment. But one thing I like about my current agentic AI PC is that even if I stop using it for AI, I’ll still have a PC. It’s like the choice between renting a router from my ISP or buying my own. I get a much higher return on investment from buying my own, not to mention better performance.
Where learning changes, possibilities abound
Now that I’ve broken past my initial expectations for what AI can do, and I’m in a new stage of simply talking to it about what’s possible, I can’t stop thinking about what I’ll try next.
Sometimes, I think about the time-intensive but menial tasks that clog up my life. Things like managing my calendar, or basic reporting to my coworkers about recent and upcoming projects. An agentic AI system, even one running on a relatively low-powered mini PC, could handle those tasks without breaking a sweat.
But increasingly, I’m thinking about how AI might do more than simply take the reins when I run across a boring, data-intensive task. I’m finding myself feeling less apprehensive about how much work it’ll take to try out something new. When everything is just a conversation away, I feel confident. I’m still learning about home labs, music production, and lawn mower repair, but I know how to have a conversation. And in the era of agentic AI, that’s a superpower.