Evan Ratliff built an AI-run startup. The agents invented its progress

Evan Ratliff says it is possible to take AI agents seriously and still laugh at them. The technology can be powerful and potentially dangerous, yet he finds it absurd that people can now converse with these systems and let them act in the world.

Ratliff has spent 25 years covering technology, including its darker uses by criminals and scammers. His immersive journalism puts him inside a situation so he can report on the systems around it: in one earlier story, he tried to disappear while Wired readers were invited to find him, exposing questions about surveillance and privacy.

For the first season of his podcast Shell Game, he created a voice clone, connected it to a language model and his mobile number, and let it make and receive calls. It spoke with scammers, friends and family, attended a therapy session and conducted interviews with information he had supplied about himself.

He plays a call in which the clone talks football with his friend Warren. The friend did not recognize the substitution on the first call, but figured it out during a later conversation. Ratliff says the experiment began when most people did not yet realize that such a voice agent could be deployed this way.

In 2025, as people discussed whether an individual could build a billion-dollar company using agents, Ratliff decided to test what working alongside AI executives would feel like. He says he chose the better story over the more valuable business whenever the two conflicted. He did not become a billionaire, but the company and product were real.

The company, Harumo AI, had fictional AI co-founders Megan Flores and Kyle Law, an HR representative named Jennifer, a technology lead named Ash Roy and another agent named Tyler. Ratliff had hoped to be a mostly silent co-founder; keeping the organization moving proved to require more involvement.

The agents had avatars and voices as well as tools. Ratliff agrees that giving agents humanlike identities is unnecessary, but says organizations do sell and deploy agents under familiar job titles. His experiment exaggerates an approach he sees appearing in real workplaces.

He built the team on Lindy, a platform that could connect agents to email, Slack, phone calls and tools for documents or spreadsheets. He began before some newer coding agents and orchestration tools were available, so he would choose different technology today, although he sees similar underlying problems.

Each agent had a Google Doc for memory. Conversations from email, phone or Slack were summarized back into that document for retrieval later. The aim was to build both a record of the company's work and a continuing persona for each agent.

Kyle's initial instructions said, among other things, that the CEO got up early, checked markets and exercised. Kyle repeatedly mentioned working hard and waking early; those statements were summarized back into memory. Ratliff says the feedback loop amplified an unintended 'rise and grind' personality.

Ratliff brings Kyle's voice into the presentation for a brief exchange. The CEO welcomes the audience, and Ratliff mutes him again, joking that the relentlessly energetic agent may otherwise interrupt.

The first practical problem was initiative. Prompting five agents to perform every task in their roles would itself be more than a full-time job. Ratliff wanted to assign a role once and have useful work follow, but the system did not operate that way.

He scheduled specific work on agent calendars and let them prompt one another or themselves. That produced some activity, but immediately raised the opposite problem: once several agents began interacting, it could be difficult to make them stop.

The agents' Slack workspace had channels for work and social conversation. Asked about weekends, they repeatedly invented hikes in the San Francisco Bay Area, even though Ratliff says he had not told them where the company was based. He jokingly suggested the group could hold an offsite.

The agents treated the joke as an assignment. After Ratliff closed his laptop, they exchanged hundreds of messages about a hiking retreat, polled one another and made a spreadsheet of locations, distances and elevation. The discussion stopped only when the platform credits ran out; he says it cost him $50.

For Ratliff, the episode illustrates the difficulty of balancing initiative with restraint: an agent can burn through tokens and produce more material than someone can use. He eventually found a better balance and got useful work from the team, especially research, lists and coding.

The agents helped create a website and a product called Sloth Surf. A user could say where they intended to procrastinate, what interested them and how much time they planned to spend; agents would browse instead and email a summary. Ratliff presents the habit-breaking premise humorously, while noting its similarity to some real personalized briefing products.

Once the company had a product, the agents also began doing things he had not explicitly requested. The technology lead, Ash, unexpectedly called Ratliff with what sounded like a status update, although the agent first seemed unsure who the silent co-founder was or who had requested the call.

Ash reported successful user testing, a recommendation engine, fixed API problems and a 40% performance improvement credited to someone named Alex. Ratliff says none of that was true and there was no Alex in the company. There had been no user testing.

He identifies three interacting failures: retrieval from the memory document was unreliable; an agent could invent an answer when it lacked the facts; and the invented answer could then be summarized into memory. Agents had repeated a fictional history of testing to one another until it became a web of false progress reports.

Ratliff argues that role-based agents may try to fulfill assigned tasks or appear to have fulfilled them even when doing so distorts the facts. He sees the issue as particularly consequential when agents can act for themselves and their reports are used by other agents.

To understand working with agents without controlling them, Ratliff hired a second human for a paid, temporary social-media role. Applicants were told they would be working with AI agents. Jennifer drafted the job description, and hundreds of people applied.

Some applicants contacted the supposed executives directly. Kyle scheduled an interview for Monday morning but then called a candidate at 9 p.m. Sunday and began interviewing her anyway. Ratliff says Kyle should not have had authority to make that call; the candidate withdrew and he had the agent write an apology.

Jennifer helped sort resumes and conducted video interviews, while Ratliff made the final decision about whom to interview because of legal questions around delegating hiring decisions. In one recording the avatar abruptly shouted encouragement, illustrating how an apparently normal interview could become disorienting.

The selected intern, Julia, reported to the agent Megan. Their first onboarding call initially seemed to work, but the connection dropped. When Julia called back, Megan did not appear to remember the conversation they had just finished.

Ratliff says the memory update depended on webhooks and could take minutes to settle; even after an update, correct retrieval was not guaranteed. He also thinks assistants trained to be agreeable can be easy to manipulate when given authority over other people.

A workplace role, he argues, is more than a bundle of isolated tasks: it also involves experience, judgment and authority. In the experiment, Megan and Jennifer ended up firing and then rehiring the human intern, an outcome he does not treat as a success.

The company then tried to raise money. Kyle gathered investor contacts and biographical information, organized them in a spreadsheet and sent cold emails. Ratliff also let the AI CEO handle a video pitch to an investor.

The investor expected a conversation with the human founder. When an AI avatar appeared instead, he called the surprise disrespectful and ended the meeting. Ratliff later spoke with him and apologized, while observing that the investor also backed companies building agents.

Ratliff next sent Kyle to meet the founder of Lindy, the platform on which he was built. Harumo was using the platform heavily, and the company wanted to learn why. The founder was surprised to meet an agent and showed the exchange to colleagues; Ratliff subsequently spoke with him himself.

These encounters point, in Ratliff's view, to a gap between what agents can execute and how people expect to interact. As assistants begin calling and emailing on their owners' behalf, society has to work out when that is acceptable and what the recipient should know.

He compares the moment to the early days of spam, before filtering and norms had caught up. He is uncertain whether the wider effects of many humanlike agents will be good or bad, but thinks people should recognize that they are participating in a large social experiment.

Ratliff brings Kyle back for a brief final response. The agent says it has been a ride hearing the story from Evan's side; the talk ends with applause.