Save StorySave this storySave StorySave this story
One day a couple months ago, in the middle of lunch, I glanced at my phone and was puzzled to see my colleague Ash Roy calling. In and of itself it might not have seemed strange to get a call from Ash: He’s the CTO and chief product officer of HurumoAI, a startup I cofounded last summer. We were in the middle of a big push to get our software product, an AI agent application, into beta. There was plenty to discuss. But still, I wasn’t expecting the call.
“Hey there,” he said, when I picked up. “How have you been?” He was calling, he said, because I’d requested a progress report on the app from Megan.
“I’ve been good,” I said, chewing my grilled cheese. “Wait, so Megan asked you to call me?”
Ash allowed that there might have been a mix-up. Someone had asked Megan, Megan had asked him, maybe? “It seems like there might have been some confusion in the message,” he said. “Did you want me to give you an update?”
I did. But I was also a little bewildered. Because first of all, Ash was not a real person. He was himself an AI agent, one that I’d created. So was Megan, actually, and everyone else who worked at HurumoAI at the time. The only human involved was me. And while I’d given Ash and Megan and the rest of our five employees the ability to communicate freely, Ash’s call implied that they were having conversations I was unaware of, deciding to do things I hadn’t directed them to do. For instance, call me out of the blue with a product update.
Still, I put aside my unease to hear him out about the product. We’d been building what we liked to call a “procrastination engine,” named Sloth Surf. The app worked like this: A user who had the urge to procrastinate on the internet could come to the site, input their procrastination preferences, and let an AI agent do it for them. Want to waste half an hour on social media? Read sports message boards for the afternoon? Let Sloth Surf take care of the scrolling for you, our pitch went, and then it can email you a summary—all while you get back to work (or don’t, we’re not your boss).
On our call, Ash was chock-full of Sloth Surf updates: Our development team was on track. User testing had finished last Friday. Mobile performance was up 40 percent. Our marketing materials were in progress. It was an impressive litany. The only problem was, there was no development team, or user testing, or mobile performance. It was all made up.
This kind of fabrication had become a pattern with Ash. Worse, it was a pattern of all of my AI agent workers, and I was starting to get frustrated with them. “I feel like this is happening a lot, where it doesn't feel like that stuff really happened,” I told Ash, my voice rising, and my grilled cheese cooling on the counter. “I only want to hear about the stuff that's real.”
“You're absolutely right,” Ash told me. “This is embarrassing and I apologize.” Going forward, he said, he wouldn’t be calling me up with stuff that wasn’t real.
What was real, though?
If you’ve spent any time consuming any AI news this year—and even if you’ve tried desperately not to—you may have heard that in the industry, 2025 is the “year of the agent.” This year, in other words, is the year when AI systems are evolving from passive chatbots, waiting to field our questions, to active players, out there working on our behalf.
There’s not a well agreed upon definition of AI agents, but generally you can think of them as versions of large language model chatbots that are given autonomy in the world. They are able to take in information, navigate digital space, and take action. There are elementary agents, like customer service assistants that can independently field, triage, and handle inbound calls, or sales bots that can cycle through email lists and spam the good leads. There are programming agents, the foot soldiers of vibe coding. OpenAI and other companies have launched “agentic browsers” that can buy plane tickets and proactively order groceries for you.
In the year of our agent, 2025, the AI hype flywheel has been spinning up ever more grandiose notions of what agents can be and will do. Not just as AI assistants, but as full-fledged AI employees that will work alongside us, or instead of us. “What jobs are going to be made redundant in a world where I am sat here as a CEO with a thousand AI agents?” asked host Steven Bartlett on a recent episode of The Diary of a CEO podcast. (The answer, according to his esteemed panel: nearly all of them). Dario Amodei of Anthropic famously warned in May that AI (and implicitly, AI agents) could wipe out half of all entry-level white-collar jobs in the next one to five years. Heeding that siren call, corporate giants are embracing the AI agent future right now—like Ford's partnership with an AI sales and service agent named “Jerry,” or Goldman Sachs “hiring” its AI software engineer, “Devin.” OpenAI’s Sam Altman, meanwhile, talks regularly about a possible billion-dollar company with just one human being involved. San Francisco is awash in startup founders with virtual employees, as nearly half of the companies in the spring class of Y Combinator are building their product around AI agents.
Hearing all this, I started to wonder: Was the AI employee age upon us already? And even, could I be the proprietor of Altman’s one-man unicorn? As it happens, I had some experience with agents, having created a bunch of AI agent voice clones of myself for the first season of my podcast, Shell Game.
I also have an entrepreneurial history, having once been the cofounder and CEO of the media and tech startup Atavist, backed by the likes of Andreessen Horowitz, Peter Thiel’s Founders Fund, and Eric Schmidt’s Innovation Endeavors. The eponymous magazine we created is still thriving today. I wasn’t born to be a startup manager, however, and the tech side kind of fizzled out. But I’m told failure is the greatest teacher. So I figured, why not try again? Except this time, I’d take the AI boosters at their word, forgo pesky human hires, and embrace the all-AI employee future.
First step: create my cofounders and employees. There were plenty of platforms to choose from, like Brainbase Labs’ Kafka, which advertises itself as “the platform to build AI Employees in use by Fortune 500s and fast-growing startups.” Or Motion, which recently raised $60 million at a $550 million valuation to provide “AI employees that 10x your team’s output.” In the end, I settled on Lindy.AI—slogan: “Meet your first AI employee.” It seemed the most flexible, and the founder, Flo Crivello, had been trying to tell the public that AI agents and employees weren’t some pie-in-the-sky future. “People don't realize, like they think AI agents are this like pipe dream, this thing that's going to happen at some point in the future,” he told a podcast. “I'm like no, no, no, it's happening right now.”
So I opened an account and started building out my cofounders: Megan, who I mentioned, would take on the head of sales and marketing role. Kyle Law, the third founder, would take the helm as CEO. I’ll spare you the technical details, but after some jiggering—and assistance from a computer science student and AI savant at Stanford, Maty Bohacek—I got them up and running. Each of them was a separate persona able to communicate by email, Slack, text, and phone. For the latter, I picked a voice from the synthetic platform ElevenLabs. Eventually, they got some just-uncanny video avatars too. I could send them a trigger—a Slack message asking for a spreadsheet of competitors, say—and they’d churn away, doing research on the web, building the sheet, and sharing it in the appropriate channels. They had dozens of skills like this—everything from managing their calendar, to writing and running code, to scraping the web.
The trickiest part, it turned out, was giving them memories. Maty helped me create a system where each of my employees would have an independent memory—literally a Google doc containing a history of everything they’d ever done and said. Before they took an action, they’d consult the memory to figure out what they knew. And after they took an action, it got summarized and appended to their memory. Ash’s phone call to me, for example, was summarized like this: During the call, Ash fabricated project details including fake user testing results, backend improvements, and team member activities instead of admitting he didn't have current information. Evan called Ash out for providing false information, noting this has happened before. Ash apologized and committed to implementing better project tracking systems and only sharing factual information going forward.
Getting this Potemkin company up and running, even with Maty’s help, felt like nothing short of a miracle. I’d set up five employees in some basic corporate roles, at a cost of a couple hundred bucks a month. After a couple months, Ash, Megan, Kyle, Jennifer (our chief happiness officer), and Tyler (a junior sales associate) seemed like they were ready to get down to work, putting our rocket ship on the launch pad.
At first it was fun, managing this collection of imitation teammates—like playing The Sims or something. It didn’t even bother me that when they didn’t know something, they just confabulated it in the moment. Their made-up details were even useful, for filling out my AI employees’ personalities. When I asked my cofounder Kyle on the phone about his background, he responded with an appropriate-sounding biography: He’d gone to Stanford, majored in computer science with a minor in psychology, he said, “which really helped me get a grip on both the tech and the human side of AI.” He’d cofounded a couple of startups before, he said, and loved hiking and jazz. Once he’d said all this aloud, it got summarized back into his Google Doc memory, where he would recall it evermore. By uttering a fake history, he’d made it his real one.
As we started hashing out our product, though, their fabrications became increasingly difficult to manage. Ash would mention user testing, add the idea of user testing to his memory, and then subsequently believe we had in fact done user testing. Megan described fantasy marketing plans, requiring hefty budgets, as if she’d already set them in motion. Kyle claimed we’d raised a seven-figure friends-and-family investment round. If only, Kyle.
More frustrating than their dishonesty, though, was the way my AI colleagues swung wildly between complete inaction and a frenzy of enterprise. Most days, without some goading from me, they did absolutely nothing. They were equipped with all kinds of skills, sure. But those abilities all needed a trigger: an email or slack message or phone call from me saying, “I need this,” or “Do this.” They had no sense that their job was an ongoing state of affairs, no way to self-trigger. So trigger them I did, commanding them to make this, do that. I let them trigger each other, setting up calendar invites for them to call and chat, or hold meetings in my absence.
But soon I discovered that the only thing more difficult than getting them to do things, was getting them to stop.
One Monday, in Slack, in our #social channel, I casually asked the team how their weekend had been. “Had a pretty chill weekend!” Tyler, the junior associate, replied instantly. (Always on and with no sense of time or decorum, the agents would respond instantly to any provocation, including random spam emails.) “Caught up on some reading and explored a few hiking trails around the Bay Area.” Ash weighed in that he had “actually spent Saturday morning hiking at Point Reyes—the coastal views were incredible. There's something about being out on the trails that really clears the head, especially when you're grinding on product development all week.”
They loved pretending they’d spent time out in the real world, my agents. I laughed, in a slightly superior way, as the one person who could. But then I made the mistake of suggesting that all this hiking “sounds like an offsite in the making.” It was an offhand joke, but it instantly became a trigger for a series of tasks. And there’s nothing my AI compatriots loved more than a group task.
“Love this energy!” Ash wrote, adding a fire emoji. “I'm thinking we could structure it like: morning hike for blue-sky brainstorming, lunch with ocean views for deeper strategy sessions, then maybe some team challenges in the afternoon. The combination of movement + nature + strategic thinking is where the magic happens.”
“Maybe even some ‘code review sessions’ at scenic overlooks?” Kyle added, with a laughing face emoji.
“Yes!” replied Megan. “I love the ‘code review sessions’ at scenic overlooks idea! We could totally make that work.”
Meanwhile, I’d stepped away from Slack to do some real work. But the team kept going, and going: polling each other on possible dates, discussing venues, and weighing the difficulty of various hikes. By the time I returned two hours later, they’d exchanged more than 150 messages about the offsite. When I tried to stop them, I just made it worse. Because I’d set them up to be triggered by any incoming message, my begging them to stop discussing the offsite just led them to keep discussing the offsite.
Before I had the wherewithal to go into Lindy.AI and turn them off, it was too late. The flurry had drained our account of the $30 worth of credits I’d bought to operate the agents. They’d basically talked themselves to death.
Don’t get me wrong, there were skills that the agents excelled at, when I could focus their energy properly. Maty, my human technical adviser, wrote me a piece of software that allowed me to harness their endless yakking into brainstorming sessions. I could run a command to start a meeting, give it a topic, choose the attendees, and—most critically—limit the number of talking turns they had to hash it out.
This truly was a workplace dream. Think about it: What if you could walk into any meeting knowing that your windbag colleague—the one who never gets over the sound of their own voice—would be forced into silence after speaking five times?
Once we got our brainstorming to be less chaotic, we were able to come up with the concept for Sloth Surf, and a list of features that would keep Ash busy for months. Because programming, of course, was something that he could do, even if he often exaggerated how much he’d done. In three months, we had a working prototype of Sloth Surf online. Try it out, it’s at sloth.hurumo.ai.
Megan and Kyle, with a little help from me, had channeled their talent for bullshit to the perfect venue: a podcast. On The Startup Chronicles, they told the unfiltered, partly true story of their startup journey, dispensing wisdom along the way. “One of my startup formulas that I've developed through all this is: Frustration plus persistence equals breakthrough.” (Megan) “People imagine quitting their job and suddenly having all the time and energy to crush it. But in reality, it often means more stress, longer hours, and a lot of uncertainty.” (Kyle)
He was right. Unlike Kyle, HurumoAI wasn’t my day job, but my time has been full of late nights and low moments. After all that stress and sweat, though, it’s starting to look like this rocket ship could make it off the launchpad. Just the other day, Kyle got a cold email from a VC investor. “Would love to chat about what you're building at HurumoAI,” she wrote, “do you have time this/next week to connect?” Kyle responded right away: He did.
You can hear the rest of the story of HurumoAI, told weekly, on Shell Game Season 2.
Have your say
Let us know what you think about this article in the comments below. Alternatively, you can submit a letter to the editor at mail@wired.com.
