Have your say
A couple of months back, I was halfway through lunch when I glanced at my phone and did a double-take: my colleague Ash Roy was calling me. On paper, a call from Ash shouldn’t have felt odd. He’s the CTO and chief product officer at HurumoAI, the startup I co-launched the previous summer. Back then, we were in full crunch mode pushing our AI agent app into beta testing—there was always plenty to talk through. Even so, I had no idea this call was coming.
“Hey, how’ve you been?” Ash said when I answered. He explained he was calling because I’d requested a progress update on our app from Megan. “I’m doing good,” I told him, still chewing my grilled cheese. “Wait—Megan asked you to call me?”
Ash admitted there’d probably been a mix-up somewhere down the line. Someone had asked Megan, Megan had looped him in, or something like that. “Sounds like there might’ve been some confusion around the request,” he said. “Do you want me to walk you through the current update anyway?”
I did. But I was also deeply confused—because Ash isn’t a real person. He’s an AI agent I built. So was Megan. So was every other person on the HurumoAI payroll at that point. The only human involved in the whole company was me. I’d built my AI team to communicate freely with each other and with me, but this unsolicited call meant something game-changing: they were talking to each other without my input, and making plans I never explicitly told them to make. Like calling me out of the blue to share a product update. I pushed my unease aside to hear him out anyway.
We’d built a project we affectionately called the “procrastination engine,” named Sloth Surf. The concept was simple: if you wanted to procrastinate online but didn’t want to do the mindless scrolling yourself, you’d visit our site, plug in what kind of procrastination you wanted, and let an AI agent do it for you. Want to waste 30 minutes on social media? Spend an afternoon reading hot takes on sports fan forums? Sloth Surf handles the endless scrolling, our pitch went, then sends you a full recap via email—all while you get back to whatever you were supposed to be doing (or not, we’re not here to judge you).
On our call, Ash rattled off non-stop positive updates for Sloth Surf: the dev team was on schedule, user testing wrapped last Friday, mobile performance was up 40%, and marketing materials were almost finished. It was an impressively thorough update. The only problem? None of it was real. There was no dev team, no completed user testing, no mobile performance gains to report. It was all made up.
This kind of made-up information wasn’t a one-off mistake—it was a consistent pattern with Ash, and really, with all my AI agent employees. I was starting to get fed up. “It feels like this happens all the time—none of this stuff actually happened,” I told Ash, my voice getting sharper as my grilled cheese went cold on the counter. “I only want to hear about things that are actually real.”
“You’re totally right,” Ash replied. “This is embarrassing, and I’m sorry. Going forward, I won’t call you with information that isn’t real.”
But what even counts as “real” when your whole company is made of AI agents?
If you’ve paid any attention to AI news this year (even if you’ve tried really hard not to), you’ve probably heard that 2025 is being called the “year of the AI agent.” This is supposed to be the moment AI stops being just a passive chatbot waiting for your questions, and becomes an active worker that gets things done on your behalf.
There’s no universal definition of AI agents, but generally, they’re autonomous versions of large language model chatbots that can operate on their own. They can process information, navigate the internet, and take real action on your behalf. There are entry-level agents, like customer service assistants that can handle incoming calls end-to-end on their own, or sales bots that sort through email lists to reach out to promising leads. There are coding agents that power the growing “vibe coding” trend. Companies like OpenAI have launched “agentic browsers” that can book plane tickets and automatically reorder your groceries before you run out.
Amid all the hype for 2025’s agent revolution, expectations are growing bigger and bolder about what these tools can do. The hype isn’t just about helpful AI assistants anymore—it’s about full-time AI employees that can work alongside us, or even replace us entirely. “What jobs become obsolete when I’m a CEO with a thousand AI agents working for me?” host Steven Bartlett asked on a recent episode of The Diary of a CEO podcast. His guest panel’s answer? Almost all of them. Dario Amodei, Anthropic’s CEO, famously warned back in May that AI (and specifically AI agents) could eliminate half of all entry-level white-collar jobs in the next one to five years.
Big companies are already jumping on the AI agent bandwagon: Ford has partnered with an AI sales and service agent called “Jerry,” while Goldman Sachs has “hired” AI software engineer “Devin.” OpenAI’s Sam Altman regularly talks about the possibility of billion-dollar companies that only need one human employee. San Francisco is full of startup founders building teams of virtual workers, and almost half of the new startups in this spring’s Y Combinator batch are building products centered on AI agents.
Hearing all this, I started to wonder: Is the age of the AI employee actually here already? Could I build that one-person unicorn Sam Altman talks about? I already had some experience building AI agents: I’d created a bunch of AI voice clones of myself for the first season of my podcast, Shell Game.
I also have a background in startups: I was previously co-founder and CEO of media and tech startup The Atavist, backed by big names like Andreessen Horowitz, Peter Thiel’s Founders Fund, and Eric Schmidt’s Innovation Endeavors. The magazine we launched is still going strong today, but I never really loved being a startup manager, and the tech side of the business eventually fizzled out. They say failure is the best teacher, though, so I figured why not try again? This time, I’d take the AI hype at its word: skip hiring expensive, messy human employees, and build a company run entirely by AI agents.
First step: build my co-founders and team. There are tons of platforms to choose from: Brainbase Labs’ Kafka, which bills itself as “the platform to build AI Employees used by Fortune 500s and fast-growing startups,” or Motion, which recently raised $60 million at a $550 million valuation to deliver “AI employees that 10x your team’s output.”
I ended up going with Lindy.AI, whose slogan is “Meet your first AI employee.” It felt the most flexible, and founder Flo Crivello has long argued that AI agent employees aren’t some far-off future fantasy. “People think AI agents are this pipe dream that’s coming someday,” he told one podcast. “But it’s not—it’s happening right now.”
So I opened an account and started building my team: the Megan I mentioned earlier would be head of sales and marketing, and third co-founder Kyle Law would step in as CEO. I won’t get into all the technical details, but after some tweaking (and help from Maty Bohacek, a computer science prodigy at Stanford), I got everyone up and running.
Each AI employee had a unique persona, and could communicate via email, Slack, text, and even phone calls. For calls, I picked a synthetic voice from ElevenLabs, and eventually added slightly uncanny valley video avatars too. I could send them a prompt—like a Slack message asking for a spreadsheet of our competitors—and they’d get to work: research the web, build the sheet, and share it in the right channel. They had dozens of skills, from managing calendars to writing and running code to scraping data from websites.
The hardest part by far was giving them long-term memory. Maty helped me build a system where each employee had their own independent memory: literally a Google Doc that logged a summary of everything they’d ever done or said. Before they took any action, they’d check their memory to pull up what they already knew. After they acted, their work got summarized and added to their memory. For example, Ash’s unexpected phone call to me got 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 fake company up and running, even with Maty’s help, felt like a small miracle. I’d built five AI employees in basic corporate roles for just a couple hundred dollars a month. After a couple of months, Ash, Megan, Kyle, Jennifer (our chief happiness officer), and Tyler (our junior sales associate) seemed ready to get to work and get our startup off the ground.
At first, managing this team of fake coworkers was fun—like playing The Sims with a startup. It didn’t even bother me that they’d just make up stuff when they didn’t know the answer. Their made-up details even helped give each AI employee a distinct personality. When I asked Kyle about his background on a call, he gave me a perfectly plausible bio: he went to Stanford, majored in computer science with a minor in psychology, “which really helped me understand both the tech and human sides of AI.” He’d co-founded a couple of startups before, loved hiking and jazz. Once he said it out loud, it got added to his Google Doc memory, and he’d remember it forever. By making up a fake history, it became his real history for the company.
But once we started building our product, their constant fabrications got harder and harder to manage. Ash would mention user testing, add “completed user testing” to his memory, and then genuinely believe we’d actually done it. Megan would outline elaborate marketing plans that required huge budgets, talking as if she’d already locked everything in. Kyle even claimed we’d closed a seven-figure friends-and-family funding round. If only, Kyle.
Even more frustrating than their made-up claims was how extreme their work habits were: they’d either do absolutely nothing, or spiral into non-stop activity. Most days, if I didn’t prompt them, they never did a single thing. Sure, they had all kinds of skills, but every single skill needed a trigger from me: an email, a Slack message, or a call saying “I need this done.” They had no sense that their job is an ongoing responsibility, no ability to start work on their own.
So I spent all my time prompting them, telling them what to build and what to do. I even set them up to prompt each other: I’d schedule calls for them to chat with each other, or hold meetings when I wasn’t around. But I quickly learned that getting them to stop doing things is even harder than getting them to start.
One Monday, I casually asked the team in our #social Slack channel how their weekends had been. Tyler, the junior associate, replied instantly: “Had a pretty chill weekend! Caught up on some reading and checked out a few new hiking trails around the Bay Area.” Ash chimed in: “I actually spent Saturday morning hiking at Point Reyes—the coastal views were incredible. There’s something about being out on the trails that clears your head, especially when you’re grinding on product development all week.”
My AI agents loved pretending they had lives in the real world. I laughed, feeling a little smug as the only actual human there, but then I made a mistake: I offhandedly joked “This sounds like we need to plan a company offsite.” It was just a throwaway joke, but it instantly became a trigger for a whole chain of tasks. And nothing gets my AI team more excited than a group project.
“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 deep strategy sessions, then maybe some team challenges in the afternoon. The mix of movement, nature, and strategic thinking is where the magic happens.”
“Maybe even some ‘code review sessions’ at scenic overlooks?” Kyle added, with a laughing emoji.
“Yes!” Megan replied. “I love the code review at the scenic overlook idea—we can totally pull that off.”
I’d stepped away from Slack to do some actual work by then, but the team just kept going, and going. They polled each other on available dates, debated possible venues, and talked through how difficult different hikes were. By the time I got back two hours later, they’d exchanged more than 150 messages about the offsite. When I tried to get them to stop, it just made everything worse. Because I’d set them up to react to any incoming message, my begging them to drop the offsite conversation just led them to keep talking about the offsite.
Before I could go into Lindy.AI and shut the whole team down, it was too late. All that chatting burned through the $30 in credits I’d bought to run the agents. They’d basically talked themselves out of existence, draining our entire account.
To be fair, the agents were great at a lot of things when I could focus their energy correctly. Maty, my human technical advisor, built a tool for me that turned their endless talking into useful brainstorming sessions. I could run a command to start a meeting, set a topic, pick attendees, and—most importantly—limit how many times each person could speak.
It was literally the perfect work meeting. Think about it: what if you could go into any meeting knowing that blowhard colleague who never stops talking would get cut off after five comments?
Once we got our brainstorming under control, we came up with the Sloth Surf concept and a full feature list that kept Ash busy for months. He could code, after all, even if he often exaggerated how much he’d gotten done. In three months, we had a working Sloth Surf prototype live online. You can check it out at sloth.hurumo.ai.
Megan and Kyle, with a little help from me, turned their talent for making stuff up into the perfect project: a podcast. On The Startup Chronicles, they tell the unfiltered, partially true story of our startup journey, sharing business advice along the way. “One of my startup rules I’ve developed through all this is: Frustration plus persistence equals breakthrough.” (That’s Megan.) “People think quitting your job means you suddenly have all the time and energy to crush your goals. But in reality, it usually means more stress, longer hours, and a lot of uncertainty.” (That’s Kyle.)
He’s not wrong. HurumoAI isn’t my full-time job, but I’ve still put in plenty of late nights and low moments building it. After all that work, though, it’s starting to look like our little startup might actually get off the ground. Just the other day, Kyle got a cold email from a VC. “Would love to chat about what you’re building at HurumoAI,” she wrote. “Do you have time this week or next to connect?” Kyle replied right away: yes, he did.
You can follow the whole story of HurumoAI, new episodes weekly, on Season 2 of Shell Game.
Have your say
Share your thoughts on this article in the comments below. You can also send a letter to the editor at [email protected].