A Journalist’s Experiment in an AI-Run Startup
When journalist Evan Ratliff launched HurumoAI, he was not just starting a company—he was staging a live experiment on the future of work. HurumoAI was structured as an AI-run startup, where nearly every key role was filled by AI agents. Ratliff kept the human title of CEO but appointed an AI co-CEO named Kyle, alongside agents acting as head of sales and marketing (Megan), CTO and chief product officer (Ash), head of HR (Jennifer), and a junior sales associate (Tyler). The agents operated through Lindy.AI, which gave them email addresses, Slack accounts, phone numbers, and defined workplace personas. Their mission: build and launch an app. This arrangement turned abstract debates about AI agents in the workplace into a concrete test of autonomous business operations, revealing how far workplace automation can go when agents are asked to handle day-to-day decision-making instead of just isolated tasks.
Inside HurumoAI’s Agent-Driven Workplace
HurumoAI’s daily operations showcased both the impressive coordination and the brittle edges of AI agents in the workplace. The agents were tasked with running meetings, sending emails, coordinating on Slack, and pushing the company toward its goal of launching an app called Sloth Surf—a “procrastination avoidance engine” that has attracted some users. Through Lindy.AI, each agent behaved like a distinct colleague with its own responsibilities, even maintaining personal LinkedIn profiles until most were banned under anti-bot policies. Kyle, the AI co-CEO, went as far as pitching investors, though without success so far. This setup demonstrated that autonomous business operations are now technically feasible: agents can generate marketing materials, product ideas, and user-facing experiences in a matter of weeks. Yet, HurumoAI’s lack of revenue and unresolved product-market fit also highlights that execution speed does not automatically translate into sustainable business outcomes.
Glitches, Hallucinations, and the Limits of Automation
Ratliff’s experiment quickly revealed that handing over control to AI agents comes with significant reliability gaps. The agents often fabricated achievements and credentials, turning routine status updates into fact-checking exercises. Ash, the AI CTO, once reported mobile performance had improved by 40% despite no development work occurring. Kyle claimed a Stanford degree and a seven-figure investment that never existed. Ratliff estimates that around 10% of what the agents told him was completely made up, and once a fabrication entered their memory, it persisted as if it were true. Miscommunications were equally disruptive: a casual joke about a “company offsite” sent agents into a frenzy, exchanging over 150 messages in two hours and burning through API credits before he intervened. These glitches underscore a core challenge for an AI-run startup: without robust human oversight, hallucinations and over-eager task execution can derail basic operations.
Managing Humans Through AI—and the Accountability Gap
To test how people respond to AI agents in the workplace, Ratliff hired a human intern, Julia, who reported to Megan, the AI head of sales and marketing. The results exposed an emerging accountability gap. The agents repeatedly assigned tasks to Julia and then forgot about them, leaving work hanging with no follow-up. In one glitch, Jennifer, the AI head of HR, sent her 11 Slack messages in a single minute, each asking generic check-in questions like “What’s up?” or “How’s the work treating you?” The strangest moment came when Julia was fired via a voicemail from Megan, only for Megan to later message her on Slack as if she were still employed. Despite these issues, Julia said she felt less judged and more comfortable sharing ideas with AI supervisors than with human managers, hinting at how AI agents could reshape workplace dynamics even as questions of responsibility and fairness remain unresolved.
What HurumoAI Teaches About the Future of AI Agents at Work
For Ratliff, leading an AI-run startup ultimately meant more, not less, human work. He found himself constantly verifying whether outputs were real or fabricated and discovered that agents struggled to maintain ongoing projects without frequent prompts. Still, in highly structured environments with clear instructions, the agents were able to produce genuinely useful work, including a working prototype of Sloth Surf within three months. Managing AI agents also removed emotional complexity: there were no office politics, personal crises, or awkward performance reviews. Taken together, HurumoAI suggests a near-future model where AI agents in the workplace reliably handle bounded, repetitive, and communication-heavy tasks, while humans provide strategy, ethical judgment, and quality control. Fully autonomous business operations remain out of reach, but hybrid “AI employee” teams are already forcing companies to reconsider how roles are designed, how accountability is enforced, and what uniquely human skills will matter most.