Imagine having a digital shadow that doesn't just follow you around, but actually anticipates your next move, handles your tedious admin, and knows your professional history better than you do. That is the promise of the always-on personal AI agent, but the real question isn't just what they can do for us—it is whether we can ever leave them once they know everything about us. Today's prompt from Daniel is about this exact landscape, specifically looking at the friction between the convenience of these "personal AI operating systems" and the looming threat of agentic walled gardens.
It is the defining tech challenge of twenty twenty-six, honestly. We are moving past the era where you just go to a website to chat with a bot. Now, we are seeing persistent, background-process agents that live across your devices. By the way, today's episode is powered by Google Gemini three Flash, which is fitting because we are talking about the very models that these agents are plugging into. Herman Poppleberry here, ready to dive into the technical weeds because, Corn, the "always-on" part of this is where the engineering gets really tricky. Think about the telemetry involved—if an agent is "always on," it’s essentially a black box recorder for your entire digital life.
It sounds great in a marketing brochure, right? "Your AI knows you." But if I spend three years training an agent on my specific workflow, my tone of voice, and my messy filing system, am I essentially married to that platform for life? Daniel’s asking about the stars of this space, like Gobii on the commercial side and Open Claw in the open-source world. Before we get into the "how," let's look at the "how much." Because the pricing models for these things are all over the map.
They really are. On one end, you have the SaaS model, like Gobii. As of April twenty twenty-six, they’ve settled into a fifteen-dollar-a-month basic tier. That gets you the persistent context and a certain amount of "agentic cycles." It is very much the "Netflix of Productivity" approach. You pay the sub, they handle the GPU orchestration on their side, and you don't have to worry about the infrastructure. But then you look at something like Open Claw. That is open-source, so the software is "free," but the actual cost of running it is usage-based. You are paying for your own tokens through an API provider, or you’re paying for the compute if you’re self-hosting a local model.
So, with Open Claw, if I have a particularly busy week where my agent is scouring the web and drafting a hundred emails, my bill spikes. Whereas with Gobii, I’m protected by that flat rate?
Usually, though most SaaS providers have "fair use" caps. The real technical difference in pricing comes down to where the "brain" lives. Gobii is centralized. They can optimize their inference because they are running thousands of agents on the same clusters. They use techniques like PagedAttention to swap user contexts in and out of GPU memory instantly. Open Claw is decentralized. You have total control, but you lose those economies of scale. There is also a middle ground emerging—projects like Aether that use a hybrid model where you pay a small subscription for the "orchestration layer" but bring your own API keys for the heavy lifting.
I suppose the "free" open-source option isn't really free once you factor in the "Herman tax"—the amount of time you spend tinkering with your server just to make sure the agent doesn't hallucinate that your calendar is empty.
Hey, I enjoy the tinkering! But you’re right. For most people, the fifteen dollars for Gobii is actually cheaper than the three hours of troubleshooting Open Claw requirements. However, what you’re paying for isn't just the "chat." It is the task capability. We are seeing these assistants move into what I call "multi-hop orchestration."
Give me a concrete example of that. What is a "multi-hop" task that an always-on agent does better than a standard LLM?
Okay, let's take a "Meeting Prep" workflow. A standard bot requires you to paste in the context. An always-on agent like Gobii is already integrated with your calendar. It sees you have a meeting with a new vendor at ten A.M. Without you asking, it goes into your personal notes, finds the last three times you mentioned this vendor's product, looks up the vendor's recent series B funding news, checks your "work" workspace for budget constraints, and presents you with a one-page brief five minutes before the call starts. That is four or five distinct steps involving different data silos, all happening in the background.
But wait, how does it actually "see" that news? Is it just doing a Google search, or is it maintaining a constant watch on specific RSS feeds or news wires?
It’s usually a mix. Most of these agents use a "perceptual loop." Every few minutes, they refresh a set of high-priority vectors. If it knows you have a meeting with "Vendor X," it adds "Vendor X" to its high-priority watch list forty-eight hours prior. It’s essentially a proactive search rather than a reactive one. It’s like having a chief of staff who reads the newspaper specifically looking for things that affect your Tuesday schedule.
And that brings up the big "church and state" problem. If this thing is always on and always listening or watching my screen, how does it know what is a "Corn the Sloth" personal thought and what is a "Corn the Professional" work task? Daniel’s prompt mentions this division between work and personal domains.
This is where the architecture really matters. Gobii uses a "Workspace" metaphor. You explicitly toggle between them, or you assign certain apps to certain workspaces. It is a logical partition. My "Personal" workspace might have access to my Spotify and my family WhatsApp, while "Work" has the Slack and the Jira. But the issue is that humans aren't that organized. I might message my wife on Slack, or I might research a work topic on my personal phone while I’m at the gym.
Well, not exactly, because I’m not allowed to say that word—but you’ve hit the nail on the head. If I’m at the gym and I have an idea for a project, I want my agent to grab it. If my agent is in "Personal" mode, does that insight get lost in the void, or does it cross the streams?
Most of these systems are moving toward a "Tagging and RAG" approach—Retrieval-Augmented Generation. Instead of hard silos, they use metadata. Open Claw, for instance, doesn't really have "modes." It just has a massive vector database of everything it has seen. When you ask it a question, it does a similarity search. The "work" versus "personal" distinction is handled by the system prompt and user-defined tags. It is more flexible, but it is also riskier. If you aren't careful with your permissions, your agent might accidentally include a private medical note in a summary for your boss because the "semantic similarity" was too high.
That sounds like a nightmare for HR. "Why does the quarterly report include my doctor's notes about my vitamin D deficiency?"
It’s a real risk! That is why the "long-tail" projects like Memex are so interesting. Memex is local-first. It lives on your machine. The memory isn't in some cloud database; it’s a local encrypted file. It segments data based on the source application. It knows that "this came from my browser" and "this came from my IDE." It provides a much higher level of "contextual hygiene" than the big SaaS players who just want to ingest everything to make the model feel "smarter."
If Memex is local-first, does that mean it can’t talk to my phone? If I’m out and about, is my "agent" effectively lobotomized until I get back to my desk?
That’s the trade-off for privacy. You can set up a personal relay—basically a private VPN for your data—but now we’re back to the "Herman tax." You’re building your own private cloud just to keep your doctor’s notes away from your boss. For most people, the convenience of the cloud wins every time, which is exactly how the walled gardens start growing their first vines.
Let's talk about the "walled garden" aspect, because this is where I get skeptical. If I use Gobii for a year, and it learns that when I say "urgent" I mean "respond within an hour" but when my boss says "urgent" he means "three days from now," that is incredibly valuable "learned behavior." If I decide Gobii’s price is going up and I want to switch to Open Claw, can I take that "wisdom" with me? Or am I starting from scratch?
That is the million-dollar question. Right now, portability is the big differentiator. Gobii offers a JSON export. You can download a giant file of your history. But here is the catch: raw data isn't the same as an "agentic state." Transitioning that JSON into a different model’s memory system is like trying to transplant a human brain into a different body and expecting it to remember how to ride a bike immediately. The "weights" of the relationships between your data points are often proprietary.
So the "JSON export" is a bit of a marketing gimmick. "Look, we give you your data!" but the data is just a pile of bricks, and you’ve lost the blueprint of the house the agent built.
Mostly, yes. It's like being given a list of every word in a book but losing the order of the chapters. But we are seeing a shift. The latest release of Open Claw in March twenty twenty-six added native support for the Model Context Protocol, or MCP. This is a big deal. It is an open standard for how agents interact with data and tools. If an agent is "MCP-native," it means its memory and tool-use patterns are formatted in a way that other MCP-compliant agents can understand. It’s like moving from a proprietary document format to a universal Markdown file.
That feels like the "USB-C" moment for AI. But does Gobii support it? Or are they sticking to their proprietary "Lightning cable" equivalent?
Gobii is "evaluating" it, which is corporate-speak for "we’ll wait until our biggest enterprise customers threaten to leave." They have a huge incentive to keep you in their ecosystem. If your agent knows exactly how you like your spreadsheets formatted and how you draft your weekly reports, the "cost" of switching isn't just fifteen dollars—it's the two weeks of frustration while you teach a new agent your quirks.
I love that "interoperable context layers" sounds so sophisticated, but it really just means "I can take my stuff and go." What about the long-tail players Daniel mentioned? You brought up Memex, but what else is out there for the people who don't want to be part of the "Big Agent" ecosystem?
There is a project called Aether that I’ve been watching. They are a SaaS, but they’ve built their entire business model around what they call a "Data Passport." Every night, Aether generates a "compressed context summary" of everything it learned about your preferences that day and sends it to a storage location you control—like your own Dropbox or a private server. If you cancel your subscription, you have a chronological, machine-readable "diary" of your agent’s evolution. It is designed specifically to be "ingestable" by other models.
That is actually a really smart middle ground. It acknowledges that you might want the convenience of SaaS but the insurance policy of open source. What about the "affordability" side of the long tail? Because fifteen dollars a month is fine for a professional, but if we want everyone to have an agent, that price needs to drop.
That is where the "Small Language Model" or SLM movement comes in. There are projects now that run entirely on your phone’s NPU—the Neural Processing Unit. They aren't as "brilliant" as a hundred-billion-parameter model in the cloud, but for ninety percent of personal tasks—scheduling, summarizing messages, filing receipts—they are more than enough. Since they run locally, the "pricing" is just the battery life on your phone. No subscription, no API fees.
It’s the "good enough" AI. I don't need a supercomputer to tell me I have a dentist appointment at two. I just need a bot that doesn't forget.
And the "always-on" nature of a local SLM is actually superior in terms of latency. If your agent has to round-trip to a server in Virginia every time you ask it to "check my last note," there is a lag. If it is living in your phone's RAM, it is instantaneous. That creates a much more "seamless" feeling of a personal OS. Plus, there's no data cap. You can have that agent summarize every single Discord notification you get without worrying about your token usage for the month.
Okay, so we’ve got the landscape. We’ve got the giants like Gobii, the open-source heavyweights like Open Claw, and the clever long-tail stuff like Aether and local SLMs. If I’m a listener and I’m ready to take the plunge into the "agentic lifestyle," what is my first move? How do I test the waters without getting trapped in the garden?
My advice is to start with a "Context Dump" test. Before you commit your life’s data to an agent, ask the provider—or check the documentation—for a sample of what their export looks like. If it is just a "History.csv" file with a list of your past prompts, that is a red flag. That is just a log. You want to see "Learned Preferences" or "Entity Graphs." You want to see that the AI is actually structuring its knowledge of you in a way that is portable.
It’s like checking the exit signs in a building before you sit down for dinner. You want to know how to get out if the kitchen catches fire. But what about the setup? Is there a way to "multi-home" my agent? Like, run Open Claw on my home server but use Gobii for work?
You can, but then you lose the "Universal OS" benefit. The whole point of an always-on agent is that it has the full picture. If your work agent doesn't know you have a flight at 6 PM because that's in your "personal" agent, it might schedule a meeting for 5:30. The real holy grail is "Federated Context"—where different agents can share specific, authorized bits of information with each other using a protocol like MCP. We're not quite there yet for consumers, but that’s the direction the wind is blowing.
That sounds like a lot of digital paperwork. I'm starting to think the "agent" is going to need an agent just to manage its permissions.
(Laughs) You’re not wrong! Also, prioritize agents that follow the "Human-in-the-Loop" principle for high-stakes tasks. We’ve talked about this before in terms of safety, but it is also a "portability" issue. If your agent is making decisions in a "black box" without explaining its reasoning, you aren't learning how to work with the AI; you’re just becoming a passenger. You want an agent that "shows its work" so you can understand the logic it is using. That logic is what you’ll need to carry over to the next system.
I think there is also a "psychological lock-in" that we don't talk about enough. It’s not just the data; it’s the habit. If I get used to saying "Hey Bot, do the thing," and "the thing" involves five different apps, I forget how to do "the thing" myself. I become "agent-dependent."
That is the "calculus of convenience." We’ve done it with everything—GPS, phone numbers, spelling. AI agents are just the next level of that cognitive offloading. The goal isn't necessarily to avoid the dependency, but to make sure the dependency is on a "standard" rather than a "vendor." If you know how to operate an "MCP-compliant agent," you are like a driver who knows how to operate any car. If you only know how to use Gobii, you’re like someone who can only drive one specific brand of tractor.
And I don't even like tractors. They are way too fast for a sloth.
Very true. But seriously, the "long-tail" projects are where the real innovation in portability is happening because they have to be better to compete with the marketing budgets of the big guys. Look at projects like "Aether" or "Memex." They are building for the "sovereign individual" model of AI.
"Sovereign individual" sounds like something you’d hear at a tech conference right before they announce a new cryptocurrency. But in this context, it actually makes sense. It is about owning your own "digital twin."
It is! Think about it: by twenty twenty-seven, your personal AI agent will likely have a more complete record of your thoughts and actions than your own memory does. It will know why you made certain decisions in twenty twenty-four that you’ve long since forgotten. That "digital autobiography" is incredibly powerful. Letting a single corporation own the "index" to your own life is a massive strategic error for any professional. Imagine a job interview in 2028 where the recruiter asks for your "Professional Agent Summary." If you can't export that because it's locked in a defunct startup's database, you're at a huge disadvantage.
So, the "takeaway" here is: go for the power, but keep the keys. Use the agents to supercharge your workflow, but treat the "context portability" as a non-negotiable feature, not a "nice-to-have."
Precisely. And don't be afraid to pilot multiple systems. You don't have to put your whole life into one agent on day one. Use Open Claw for your coding tasks, maybe try Aether for your personal research, and see which one actually lets you move data between them. The "friction" you feel when trying to move that data is the most honest review of the product you’ll ever get.
I like that. The "Friction Test." If it’s hard to leave, it’s a cage. If it’s easy to leave, it’s a tool.
And we want tools, not cages. Even if the cages come with a really nice "Meeting Prep" feature and a fifteen-dollar-a-month price tag.
Well, I’m going to go see if I can get a local SLM to run on my toaster. It seems like the right speed for me.
As long as it doesn't hallucinate that your toast is a "work task," you should be fine.
No promises. Before we wrap up, I want to reiterate how much the "always-on" aspect changes the game. This isn't just about "productivity." It is about a shift in how we interact with information. When the agent is always there, the "barrier to entry" for any task drops to near zero. That sounds amazing, but it also means we might end up doing a lot more "busy work" just because it’s easy.
That is the "efficiency paradox." When you make it easier to do tasks, people just find more tasks to do. The key is to use that saved time for actual deep work, not just to let your agent manage an even larger pile of irrelevant emails. We see this with "Agentic Spam" already—people using agents to generate thousands of "personalized" outreach messages. If everyone has an agent, the noise level just goes up.
It’s the tragedy of the digital commons. If everyone has an agent to filter the noise, but everyone also has an agent to create the noise, we’re just building a bigger wall.
Which is why the "Personal AI OS" needs to be defensive as much as it is offensive. It needs to be your gatekeeper.
Deep work? I’m a sloth, Herman. My "deep work" is a four-hour nap followed by a very slow snack.
Well, maybe your agent can schedule those naps for maximum physiological recovery.
Now you’re talking. That is the kind of "personal AI OS" I can get behind. Does it have a "Do Not Disturb" mode for when I'm eating a particularly good leaf?
Actually, most of these agents are implementing "Contextual Silence." They use your camera or mic—with permission—to sense your environment. If it sees you’re focused or, in your case, napping, it queues up all the non-urgent notifications for later. It’s an "Attention Buffer."
We should probably mention the "agentic internet" concept too, because these personal assistants aren't just working with your data—they are going out and interacting with other agents. If you send your agent to "negotiate a discount" with a service provider, your agent is basically talking to another AI. In that world, the "portability" of your preferences is what gives your agent its "personality" or its "strategy."
Oh, that’s a fascinating layer. If my agent is "me" on the internet, and I move to a new agent that doesn't have my "negotiation context," my new agent might be a pushover compared to my old one.
You lose your "edge." You’ve spent months teaching your agent that you never accept the first offer on a contract. If you switch platforms and that "learned behavior" doesn't transfer, you're literally losing money. That is why "context portability" isn't just about files; it’s about "learned traits." We are entering an era where your "AI persona" is a digital asset you need to protect and maintain.
It’s a brave new world, Herman. Or at least a very busy one. Thanks to everyone for listening to this dive into the world of personal agents. We’ve covered the pricing, the tasks, the "walled garden" risks, and the long-tail projects that might just save us from "Big Agent" lock-in.
It is a space that is evolving every single week. Keep an eye on those MCP updates and don't be afraid to demand more from your software providers. Your data is your life—don't let it get locked in a JSON file you can't use.
Well said. Thanks as always to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes. And a big thanks to Modal for providing the GPU credits that power this show—they make it possible for us to explore these deep technical topics every week.
This has been My Weird Prompts. If you are finding these deep dives helpful, we’d love it if you could leave us a quick review on your podcast app. It genuinely helps other "sovereign individuals" find the show.
Or just tell your AI agent to recommend us to all its bot friends. That works too. We are also on Telegram—just search for "My Weird Prompts" to get notified when we drop a new episode.
Until next time, keep your context portable and your agents curious.
And your naps scheduled. See ya.
Goodbye.
