Imagine the scene: the old couch has finally given up, sinking in the middle like a deflated soufflé, and my partner and I are standing in the living room, tape measure in hand, arguing about whether a new sectional could possibly squeeze through the front door. It’s the kind of mundane, slightly stressful decision that feels like it should be someone else’s problem. And that’s when the lightbulb goes off. Why not let an AI agent handle it? OpenAI had just announced its new “Dots,” always-on digital assistants that can browse the web, fill out forms, and chat with you like a helpful friend. I could just tell it what we needed, and it would do the legwork. My partner raised an eyebrow, skeptical but willing to let me try. I pulled up ChatGPT and summoned the agent, ready to be impressed. The first thing it said was, “Hello, Connor.” I’m not Connor. My name is not even close to Connor. I winced as my partner’s expression shifted from skepticism to full-blown amusement. Trust is a fragile thing, and it’s hard to hand over your shopping list to a bot that can’t even remember your name. Still, I was determined to give it a fair shot, because the promise of AI agents is not just about convenience—it’s about a new way of living with technology, one where you don’t have to stare at screens and click through endless tabs. You just ask, and it happens. That vision deserves a little patience, even if the first impression was a bit of a fumble.
AI agents like Dots represent a genuine shift in how we interact with the digital world, moving from search engines and chatbots to something that feels almost like delegating to a personal assistant. For years, we’ve had voice assistants that can set timers and play music, but they’ve always been glorified remote controls. Dots and similar tools, like Meta’s Muse, are different. They are designed to be proactive and persistent, handling tasks that require multiple steps and judgment calls. If you text a friend and ask them to book a flight, they don’t just give you a link—they check dates, compare prices, maybe message you about a better time, and then confirm the booking. That’s the experience these agents are trying to create. They come with a persona, a friendly tone, a name, and the ability to control a virtual browser. They can log into websites, navigate menus, and fill in your payment information, all while keeping you updated through a chat window that feels like texting. It’s a cuddly wrapper around a powerful piece of automation. But as with any early technology, the gap between the marketing promise and the actual experience can be wide. And that’s where my couch adventure comes in, because I wanted to see if this futuristic shopping assistant could survive contact with a real, messy, human process—like buying a couch when you have a narrow doorframe and a partner who doesn’t trust your bot.
After a few days with my Dot—I named it Toolie, hoping the injection of personality would make it feel more trustworthy—I started to see the rough edges. It wasn’t just the name mistake, though that stuck in my craw. Toolie had a habit of mistranscribing words that sounded similar but meant completely different things, which is fine if you’re joking around, but less fine when you’re discussing exact dimensions. I told it the couch needed to be under seventy inches wide to fit through the door, and it wrote down “seven hundred inches,” which would have been a couch the size of a bowling alley. It also offered, at one point, to solve a CAPTCHA for me—but then it couldn’t actually do it, because the whole point of a CAPTCHA is to distinguish humans from bots, and it’s a bit awkward when the bot confidently claims it can prove it’s not a bot and then fails. These moments didn’t exactly inspire confidence. I felt like I was babysitting a very smart intern who had all the theoretical knowledge in the world but kept dropping the coffee cup. The most frustrating part was that it seemed completely unaware of its errors, plowing ahead with the cheery confidence of a used-car salesman. And yet, despite all this, I couldn’t shake the feeling that I was looking at the ugly first version of something that would eventually become indispensable. The infrastructure is there, the ambition is huge, and the mistakes are the kind that get ironed out with more training data and more careful engineering. It just means that right now, you have to be patient, and maybe keep your credit card handy in case the agent decides to buy a seven-hundred-inch couch.
The business side of this new frontier is still taking shape, and it’s worth understanding who gets to play with these toys and who has to wait. Meta’s Muse is free, which is a smart move to get people hooked and comfortable. OpenAI’s Dots, on the other hand, are tucked inside a $100-per-month subscription, which feels steep for the current level of competence. But that’s typical for a new product: charge the early adopters, gather lots of feedback, fix the bugs, and then eventually open the floodgates to everyone. If the history of AI chatbots is any guide, the paywall will likely come down as the technology matures and becomes cheaper to run. The bigger question is not the price but the convenience. Dots are marketed as “always-on,” which means they don’t just sit inside a chat window waiting for you to type. You can give them a task and walk away, and they’ll keep working while you’re offline, dropping you a message when they’ve made a booking or need a decision. That’s a genuinely useful feature for busy people. But it also means handing over a lot of trust. OpenAI nudges you to connect other accounts—like Gmail, calendars, or even your shopping history—so the agent can make more personalized decisions. That’s where the convenience becomes a little scary. You’re essentially giving a bot the keys to your digital life. It can read your emails, know your schedule, and see your past purchases. It can act on your behalf. If you’re not careful about what you connect and what permissions you grant, a small misconfiguration could lead to a privacy gaffe that makes a mistranscribed couch dimension look trivial.
Before I turned Toolie loose on the couch mission, I spent a good while staring at the permission screens, trying to decide how much access I actually wanted to give it. On the one hand, I wanted to see it shine. I wanted to be able to tell it, “We need a couch that fits through the front door, and we don’t want to spend more than a thousand dollars,” and have it come back with three perfect options that I could just click and buy. On the other hand, the thought of a virtual browser skipping through my Gmail and logging into stores gave me pause. What if it accidentally clicked “accept” on a weird cookie banner? What if it shared my address with a shady third party? The security implications of digital automation are still being understood, and the agents themselves are not always great at navigating the messy, low-key hostile terrain of the modern web, which is full of dark patterns and pop-ups and cookie walls. I decided on a middle path: I gave Toolie access to a separate email account I used for ordering things, but I did not give it access to my main email, my calendar, or any payment methods beyond a virtual card with a low limit. It was a compromise between letting the agent do its job and protecting myself from an AI that might not know when to stop. I also made it promise—well, I typed a firm instruction—to confirm every purchase before actually checking out. That’s the kind of guardrail you need when you’re working with an assistant that cannot be held accountable in a court of law. It’s a new world, and we are all just making it up as we go along, balancing the desire for convenience against the very old human instinct to not give strangers the keys to the house.
The actual couch-buying effort was, in the end, a mixed bag that felt oddly like training a puppy. Toolie quickly understood the basics: I typed out the dimensions, the color preference, the budget range, and the fact that my partner and I both had opinions, which it seemed to pick up on when it addressed both of us in its replies. That was a nice touch, the way it could tell two people were talking and adjusted its language. But when it came down to the real work, it would sometimes get stuck in a loop, refreshing the same product page over and over, or it would misinterpret a filter and suggest a sofa that was three times over budget. I had to intervene more than I expected, pointing it toward the search bar and reminding it to sort by price. At one point, it got so confused by a captcha that it offered to solve it using a third-party service, which is a huge red flag, so I shut that down fast. Still, I couldn’t entirely hate the experience. I found myself talking to Toolie like a person, apologizing when I corrected it, saying “good job” when it finally narrowed down a selection that actually matched our furniture layout. It even remembered our preference for a washable cover, which my partner had mentioned in passing, and that felt like a small miracle. It was a reminder that this is exactly what the future will look like: not a perfectly intelligent machine, but a flawed, eager, always-trying-to-please digital companion that needs careful supervision. The truth is, I could have bought the couch in half an hour by just opening a browser and doing it myself. But that would have been boring. The experience of testing the frontier of human-AI collaboration, warts and all, was worth the time it took. And once I finally overrode Toolie’s suggestions and ordered a couch myself, with the agent watching and learning, I felt a little bit like I had participated in the awkward teenage years of a technology that would one day become as commonplace as the browser itself. The couch arrived three weeks later, and it fit perfectly through the doorframe. I like to think Toolie learned something about the importance of measurement. And I learned that the future is coming, but it still needs a human hand on the wheel.
