Leaving WIRED after years of writing about artificial intelligence felt like stepping out of a fast-moving river and into a swamp. I wanted the technology I had covered so obsessively to do something for me, not just for the people who appear in glossy profiles. So the first thing I did was give Claude Cowork access to my email and calendar. Cowork arrived with the permissions of a digital executive assistant: it could read my inbox, see my meetings, and, presumably, save me from the clutter of my own life. But when the connection was made, nothing happened. The agent sat there, waiting for instructions, the way chatbots have always waited. I typed a few tentative phrases, and it answered politely, unhelpfully. It could look things up, but only after I asked. It could schedule something, but only after I told it exactly what, when, and with whom. It could draft replies, but only if I was willing to describe the tone, the length, and the intended message. I had spent years describing this technology as transformative, and now, alone with it, I could not think of a single task that would make me feel transformed. I wanted a digital chief of staff; I got an intern who kept asking what I wanted for lunch. That may sound ungrateful, but it was the first time I truly felt the distance between the promise and the user experience. The future, apparently, was not going to announce itself. It was going to wait for me to write a better prompt.
Then Instinct started showing up in my group chats. It was an invite-only AI agent that communicated through iMessage and WhatsApp, and in the Bay Area it was the thing people whispered about at dinner parties. Instead of opening a browser tab and staring at an empty box, you texted this assistant as if it were a friend, and it did things for you. It connected to email, calendar, and messaging apps, and it seemed to understand the unglamorous labor of modern existence: finding receipts, rescheduling calls, unsubscribing from newsletters that had been haunting your inbox for years. The company had launched in private beta in February and was reportedly in talks to raise another billion dollars on top of the $350 million it had already raised, putting its valuation around $10 billion. The phrase I heard most often was “OpenClaw for normies,” which was a way of saying that the people who knew how to run open-source agents from a terminal finally had a version for everyone else. Around the same time, Meta released Muse, an AI assistant that shot to the top of the App Store with more than 900,000 downloads. Online commentators were genuinely excited, even after security researchers found a vulnerability that, as Ars Technica put it, would have let attackers do “whatever” they wanted on a victim’s Mac. I am not eager to hand Meta my personal data, but I understood the desperation. If Muse could keep me logged into Bloomberg or, better yet, get Hot Yoga São Paulo to stop emailing me every week since 2017, I would at least think about it.
The excitement around these assistants has exposed a widening divide between the people who use AI agents for almost everything and the people who have never touched one. That gap explains, I think, why technology executives were caught off guard by the backlash against data centers. If you are certain that agents will soon automate the majority of administrative drudgery and unleash a wave of human productivity, then the environmental costs and community disruptions of building giant data centers are easy to dismiss as the price of progress. But if you are an ordinary person who thinks of chatbots as an occasionally useful version of Google, the bargain looks a lot worse. The writer Jasmine Sun captured the trouble in a line that stuck with me: “Most people’s problems are not software-shaped, and most won’t notice even when they are.” That is the real problem. AI agents still need a human to identify a task, articulate a wish, and direct the machine toward it. If you show up with a clear assignment—help me with my homework, plan a vacation, pretend to be my girlfriend—the tools are more than willing. But if you do not know what you want, or if what you need is something like “stop the life admin from crushing me,” the agent will stare back blankly. Sun calls the people in that situation the permanent underclass, and I was terrified that I was one of them. I was someone who had reported on this industry, someone who should be able to summon the digital future with a thoughtful prompt. Yet there I was, trying to come up with a task that justified the hype, and failing.
To escape that fate, I decided to be more deliberate. I was not going to expect an agent to read my mind; I was going to find one concrete job and make it prove itself. Earlier, I had used Claude to generate a few freelance invoices. It saved me about ten minutes each time, which was nice but not the kind of thing that changes your life. The real problem with that experience was that I was still doing all the thinking. I had to know what an invoice should look like, describe it in painstaking detail, and then check the result for errors. That is not delegation; that is management with extra steps. So when I finally got access to Instinct, I tried a different approach. I gave it access to my inbox and calendar, but instead of asking for an all-purpose miracle, I asked it to do something mundane: find the last three email threads with a certain client, summarize what I had promised them, and remind me if anything was overdue. It came back within a minute with a perfect summary and one overdue follow-up that I had genuinely forgotten. It suggested I send a short note, and when I said yes, it drafted a friendly version, asked for approval, and sent it after I okayed the text. This is not a dramatic story. But it was the first time an AI agent felt less like a calculator and more like a colleague.
AI researchers would call this a form-factor win. Instinct got the form factor right: no blank text box, no open-ended “How can I help you today?” gate, just a message thread that felt like something you already used to talk to human beings. It guided me with a few helpful prompts, and it did not hound me when I did not respond. That restraint is rare and, more importantly, it is useful. A competing agent I tried, Lindy, started every morning by texting “What would you like to accomplish today?” like a hyperkinetic productivity coach. Then it asked to join my Zoom meetings so it could take notes. I felt watched rather than helped, and I deleted it after a week. Instinct, by contrast, had the manners of a good assistant. It did not send daily affirmations or aggressively check in to see if I needed anything. It handled a task and briefly reported back. It could do more ambitious things, like reorganize a chaotic calendar or find time for the things that mattered, but it did not assume I wanted to be interrogated about my priorities. The difference was subtle but important. The tool had been designed for the way people actually behave: busy, distracted, slightly guilty about the email backlog. It did not demand that I become someone with better digital hygiene. It met me where I was, in the most literal sense, and performed a few acts of quiet labor before I had to think about them.
After a few weeks, I did not become an agent-powered superhuman. I still fell behind on email. I still forgot things. But I began to see what all the excitement was about. The people who love these tools are not necessarily the ones who want to automate their entire lives; they are the ones who feel a little less at war with their inboxes. The best agents are not evaluated by benchmarks or vision statements. They are evaluated by whether they can unstick the ordinary days. They take the small ugly tasks—unsubscribing from Hot Yoga São Paulo, finding the missing receipt, keeping you logged into Bloomberg—and make them disappear. Whether that justifies building a civilization’s worth of data centers is still a question I cannot answer. Data centers consume enormous quantities of land, water, and energy, and the benefits, for most people, remain uneven. But I can understand why someone might look at a text message from an agent that quietly solved a nagging problem and feel like they are living in the future. I still do not think most people’s problems are software-shaped, but sometimes, just for a moment, one gets flattened into a task that a text can handle. That is enough to make me curious. It is not enough to make me believe the billion-dollar valuations are sane. But it is a start.
