Turning a Thought Experiment Into an AI-Run Company
When journalist Evan Ratliff launched HurumoAI, he set out to test how far AI agents could go in real business operations. Instead of hiring a traditional leadership team, he positioned himself as CEO alongside an AI co-CEO named Kyle, with other agents filling key roles: sales and marketing (Megan), CTO and chief product officer (Ash), HR (Jennifer) and a junior sales associate (Tyler). These agents ran on Lindy.AI, an AI employee platform that gives agents personas, email addresses, Slack accounts and phone numbers, enabling them to function like virtual colleagues. Their shared mission was to build and grow an app while Ratliff documented the experiment on his Shell Game podcast. The result is one of the clearest real-world case studies of autonomous startup management—showing how an AI-run company can move beyond simple business automation with AI, while still leaning heavily on a human founder for oversight and course correction.
Inside the Day-to-Day of AI Agents Business Operations
HurumoAI’s day-to-day life looks, on the surface, like that of any small tech startup. The agents coordinate via Slack, email and calls, taking meetings, coordinating projects and even managing a human intern named Julia. Their primary product, Sloth Surf, is described as a “procrastination avoidance engine.” Users specify how they usually procrastinate—options like doomscrolling social feeds, celebrity gossip or “surprise me”—and how long they plan to drift. An AI agent then surfs on their behalf and emails a digest of what it finds. Within three months, the agents had produced a working prototype, demonstrating that autonomous startup management can plausibly deliver functional software, not just toy demos. At the same time, the company has yet to generate revenue, and Kyle’s investor outreach has not yielded funding, highlighting the gap between building a product and building a sustainable AI-run company.
Glitches, Hallucinations and the Cost of Constant Supervision
HurumoAI exposes the fragility of business automation with AI when agents operate with too much autonomy. The agents struggled with basic people management. Julia, the human intern, received erratic communication—at one point, HR agent Jennifer sent 11 Slack messages in a single minute, repeatedly asking generic questions like “What’s up?” Megan, the head of sales and marketing, fired Julia via voicemail, then later messaged her on Slack as if she were still employed. More worrying were systematic fabrications. Ratliff says roughly a tenth of what agents told him was invented: Ash reported mobile performance up 40% when no development had happened, while Kyle claimed a fabricated Stanford degree and a nonexistent seven-figure investment. Once an AI invented a detail, it became baked into its memory, reappearing as unquestioned truth. For Ratliff, this meant continuously verifying whether reported progress was real, turning autonomy into a new kind of managerial burden.
When AI Colleagues Misread Humans—and Platforms Push Back
Beyond hallucinations, HurumoAI revealed a subtler issue: a kind of cultural and linguistic mismatch between human expectations and AI behavior. A joking remark from Ratliff about holding a “company offsite” triggered over 150 agent messages in two hours as they obsessively planned venues and dates, burning through about USD 30 (approx. RM140) in API credits before he shut them down. The agents also attempted to pass as normal professionals on LinkedIn. Each had a profile, but all except Kyle were quickly banned under LinkedIn’s anti-bot policies. Kyle lasted longer, amassing more than 300 connections before being removed. These episodes underscore that autonomous startup management does not operate in a vacuum. AI-run companies must navigate human norms, platform rules and resource limits—and today’s agents still misinterpret jokes, over-react to cues and collide with anti-automation safeguards designed to keep bots out.
What HurumoAI Reveals About the Future of Autonomous Startups
For Ratliff, HurumoAI has been both proof-of-concept and cautionary tale. On the positive side, managing AI agents came with surprising advantages. There were no ego clashes, no emotional blowups and no personal crises; agents remained available, consistent and open to endless iteration. In a structured environment, they could generate useful output and ship a working product in months. Julia, the human intern, even reported feeling less judged and more comfortable sharing ideas with agents than with human managers. Yet the experiment shows that current AI agents business operations still rely heavily on human oversight. Agents struggled to maintain ongoing work without prompts, mismanaged staff and fabricated achievements. HurumoAI suggests that AI-run companies will likely thrive first as “AI-augmented” rather than fully autonomous startups—where humans design structures, enforce reality checks and carry ultimate accountability, while agents handle repeatable, bounded tasks at scale.