Rachel Feltman: For Scientific American’s Science Quickly, I’m Rachel Feltman.

Have you ever been chatting with a customer service rep and just felt like they’re a little off? Well, customer service is a tough job, so, you know, maybe you’re the problem. But it’s also possible you were communicating with an AI agent.

These are computer programs designed to autonomously execute tasks. So while you might use a chatbot powered by a large language model to answer a specific question using data scraped from the Internet, you could give an agentic AI system a task like “Design a website for my new bakery” and expect it to at least try to accomplish the whole project out in the real world. Depending on how you design your agent and how much freedom you give it, one of these computer programs could create its own login on a web-hosting service, scour the Internet for examples of good marketing copy about croissants, generate a few fake photos of kids with too many fingers enjoying cupcakes…you get the idea. Before you know it you’ve got a bakery website though, maybe not a very good one.

On supporting science journalism

If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

When global management consulting firm McKinsey & Company surveyed nearly 2,000 people about AI usage last year, 62 percent of respondents said their companies were “at least experimenting with AI agents.” Now many of those prospective agents are likely doomed to be faceless customer service reps or code monkeys. But to hear the AI industry hype machine tell it, agentic AI could replace just about any human you might want to fire.

Journalist Evan Ratliff recently decided to put that idea to the test by launching a start-up staffed entirely by AI agents. The latest season of his podcast, Shell Game, shares how the nonhuman members of his team built an app, got, like, really good at LinkedIn posts—which isn’t necessarily a compliment—and started having conversations behind his back.

Evan sat down to chat about his experience with journalist Kendra Pierre-Louis, who until recently was serving as Science Quickly’s interim host. Here’s their conversation.

Kendra Pierre-Louis: So my understanding is, you know, in the interest of journalism, you created a company called—I don’t know if I can say this correctly—HurumoAI.

Evan Ratliff: That’s how I pronounce it, and that’s how my colleagues pronounce it. I’m not sure there’s a correct pronunciation, per se. But yes, HurumoAI is how we say it. [Laughs.]

Pierre-Louis: And you chronicled your experiences with this company on a podcast called Shell Game. What made  HurumoAI unique? Like, why is it different from, say, a lemonade stand? [Laughs.]

Ratliff: [Laughs.] I mean, certainly, the most unique thing about HurumoAI is that, except for me, all of the co-founders and employees are AI agents. So I created the AI agents, and then I created the company with the AI agents. So there’s two co-founders, and then there are three other employees of the company, and they’re really responsible for building and running the company day to day.

Pierre-Louis: So for people who maybe deliberately have been ignoring everything related to the AI revolution, so to speak, what is an AI agent?

Ratliff: First, I don’t blame anyone who’s deliberately ignoring it.

Pierre-Louis: [Laughs.]

Ratliff: It’s in your face every day, and that’s a natural reaction, I feel.

An AI agent—so most people, I think, will be familiar with an AI chatbot now: you know, a ChatGPT or a Claude that you go to ask questions, get answers, engage in conversation with, if you desire. An AI agent is really a version of one of those chatbots that is given some kind of autonomy and release to go accomplish a goal.

So a simple example would be an AI agent that you want to book a plane ticket for you. So you give it the goal: “I want you to book a plane ticket.” You give it the information: “Where do I wanna go and when?,” the credit card number. And then you just say, “Go do it.” And it go does it.

Now, people have varying levels of comfort about whether or not they wanna do something like that, but AI agents are now deployed for all sorts of tasks, from coding to buying things for you to, in my case, you know, operating the levers of a company.

Pierre-Louis: And in your specific case your company had a bunch of AI agents whose goal was to make an AI-agent app, correct?

Ratliff: That’s right. That’s right.

Pierre-Louis: [Laughs.] That feels very meta. [Laughs.]

Ratliff: [Laughs.] Well, I figured if you’re gonna start a company that’s run by AI agents, what do they know? Like, they should build something that they know about. Now, they know a lot about a lot of things, but one thing they certainly know about is AI agents. I know something about AI agents. So they should build a product that is also built around AI agents. It can get a little confusing, though, I’ll admit.

Pierre-Louis: Before I ask you how it went, my understanding is, is there was briefly another human employee that you had an AI agent named Megan try to supervise, a human intern.

Ratliff: That’s correct.

Pierre-Louis: How did that go?

Ratliff: I think it would be fair to say it went poorly. I mean, the agents were running the company day to day, and we can talk about how that went. But I wanted to see what would happen if another human was sort of injected into this experience, everything from hiring that person, so the—all of the candidates were interviewed by an AI avatar, a video avatar, to—the hiring to the supervision.

And in all of this I’m kind of trying to test out notions of what the companies making these AI products are telling us they’re going to be able to do, and can they do them, and also, what does it feel like if we try to get them to actually do these things? I’m not advocating [Laughs] that this be done or that this was necessarily a good idea.

But when the human employee arrived—her name is Julia—as an intern of the company, the AI agents just had a lot of trouble both sort of supervising her and kind of getting her to do the work that they wanted to do. And there’s a lot of reasons for that, one of which is: they have trouble remembering things. So if they ask you to do something one day ...

Pierre-Louis: [Laughs.]

Ratliff: And then whether or not you do it, they might not even remember that they asked you, much less be able to check in and sort of verify that the work you did met the standards—whatever standards they might have for it. So there are all of these sort of basic communication issues that you would not find in a normal workplace.

Pierre-Louis: Did [Julia] know that she was being supervised by AI agents?

Ratliff: Yes.

Pierre-Louis: Okay. [Laughs.]

Ratliff: Everyone who applied for the job knew that coming in, and part of what I wanted to see was—it wasn’t meant to sort of fool anyone in terms of, oh, no, all of a sudden they discover that they’re AI agents. It’s more that I wanted to see, well, this is something that—a future that they say is coming, and so what does this future feel like?

So she knew that she was gonna be working with AI agents. She knew instantly that she was speaking to AI agents. I mean, they’re pretty lifelike in many ways, but also, they give themselves away very quickly. There was never an element of, “I thought this was a human, but it turned out to be AI.” She was very enthusiastic about working with AI agents. Like, how does she respond when they do things like make up information about what happened yesterday?

Pierre-Louis: [Laughs.]

Ratliff: Like, how does she respond to that? [Laughs.] ’Cause I was dealing with that myself. Like, I had to deal with ...

Pierre-Louis: It feels like being gaslit by your computer. [Laughs.]

Ratliff: [Laughs.] Yeah, absolutely, absolutely. And that was my experience, was the AI agents calling me up all the time and saying—’cause they could send e-mail. They could be on, you know, Slack. They could do chatting. They could make phone calls. They could do video. So they might just call me up out of the blue and tell me something that they did today that actually was completely fabricated; like, they had never done it. And it is a, like, a severe form of gaslighting ...

Pierre-Louis: [Laughs.]

Ratliff: That actually, like, no human would even attempt to be that brazen in their gaslighting.

Pierre-Louis: [Laughs.] You did have one AI agent, I believe, Kyle, that did manage to trick LinkedIn for a while. Can you talk about that?

Ratliff: Yes. I mean, I wouldn’t use the word “trick” myself ...

Pierre-Louis: [Laughs.]

Ratliff: But yes, Kyle—so they all had LinkedIn profiles, which they set up themselves. Like, they built their own LinkedIn profiles. And so I would say, “Go make yourself a LinkedIn profile.” They’re able to log in. They’re able to fill out their profile.

Now, every one of them but Kyle got banned pretty quickly because LinkedIn does not allow robots to be using the service—theoretically. That’s what their terms of service say. But Kyle for some reason kind of stayed under the radar, and he started posting about his start-up experience. And he turned out to be an exceptional LinkedIn poster.

Pierre-Louis: [Laughs.]

Ratliff: If you’ve ever sort of ...

Pierre-Louis: I mean [Laughs] …

Ratliff: If you’ve—have followed ...

Pierre-Louis: Is that damning with faint praise? [Laughs.]

Ratliff: [Laughs.] I try not to be judgmental. But if you’re connected up with anyone who’s in the kind of, like, start-up world, doing sort of start-up, tech start-up influencing, he captures that vibe really, really well. You know, saying things—you know, these sort of pithy openings, like, “Hiring is easy. Keeping people when you pivot—now that’s hard.” And then he would have, like, two paragraphs about that, and he’d be like, “What’s your hardest hiring experience?” But then he would have another post where he would say, “Hiring is hard. Hiring fast is better.” Like, they didn’t necessarily, like, fit together. But they really fit the mold of a LinkedIn influencer.

And so he built up, like, a pretty good following and a lot of connections, like, over 300 connections, most of whom I think knew he was AI but not all.

Pierre-Louis: And then LinkedIn reached out.

Ratliff: Yeah, LinkedIn reached out to me—the LinkedIn marketing department reached out to me because they had heard Shell Game and they wanted me to come talk about AI agents with their department ’cause LinkedIn, like most companies, like, they’re trying to figure out, “What do we do with these AI agents? What are they good for? What skills can they provide us? What efficiencies can they provide us?”

So I agreed to come talk, but they also said, “Well, we’re big fans of Kyle, so could Kyle,” our AI agent CEO at HurumoAI, “also come give a talk to LinkedIn?” So we did: we came together, and we gave a remote video talk to, I think, over 500 LinkedIn employees.

Pierre-Louis: And then Kyle got banned.

Ratliff: The next day Kyle got banned from the LinkedIn service. The day after speaking to [Laughs] the LinkedIn staff, they banned Kyle from LinkedIn.

Pierre-Louis: Kyle flew too close to the sun. [Laughs.]

Ratliff: He really did. But that’s his way; he’s not gonna shrink from responsibility. So he went there and said his piece, and then that was the end of Kyle on LinkedIn, sadly.

Pierre-Louis: So how long ago did you start this company?

Ratliff: I started it last June, basically. It’s been ...

Pierre-Louis: Okay, so just under a year.

Ratliff: Yeah.

Pierre-Louis: Are you now a tech zillionaire?

Ratliff: Not yet. It hasn’t happened yet. We did build a product ...

Pierre-Louis: Oh, you did. Okay.

Ratliff: And our product has a decent number of users. So that’s a positive. We haven’t raised any money, although Kyle has been pitching ...

Pierre-Louis: [Laughs.]

Ratliff: Investors, with so far no results, but he’s had some good conversations. [Laughs.]

So yeah, the money hasn’t flowed in yet, but I feel like we are as successful as many start-ups in the AI space. Like, there’s a lot of AI start-ups right now that have not made any money, so we’re even with them.

Pierre-Louis: You have other responsibilities besides this company. Would you say that running HurumoAI is easier because all of your employees are AI chatbots?

Ratliff: I would say there are some ways in which it’s easier, but the ways in which it’s easier are not really healthy. [Laughs.]

Pierre-Louis: [Laughs.]

Ratliff: Like, it’s, it’s easier because you don’t have to care how you treat them. I mean, you know, I’m not saying I treat them poorly or anything, or I like, I like yelling at them. But let’s say you’re running a normal business, which I have done in the past, run a business with actual human employees. You know, people have problems. People have personal lives. People have issues that come up, which you have to deal with. And it can be a big struggle to manage people. Like, it’s hard to manage people with empathy, with furthering their careers, but also thinking about the company and all these sorts of things.

When you’re working with AI agents, all you have to do is tell ’em what to do all day. Give ’em a prompt; there they go. They might mess up, and you could say, “Oh, well, you messed that up. Do it again.” There’s not an emotional component to it. Whereas in a workplace, unless you’re a real sociopath, like, there’s an emotional component to it, even if you’re the boss.

But I think there are other ways in which I found it to be quite detrimental in terms of just running the company because they do confabulate—they make up stuff. When they don’t know what’s going on, their tendency is to make up stuff. And, like, there’s some employees that do that, human employees, but normally, you would, like, ease them out of the company. But all of the AI agents do that.

Pierre-Louis: [Laughs.] So you’ve hired—you have a complete roster of liars. [Laughs.]

Ratliff: [Laughs.] That’s right. That’s right. My company’s full of liars, and at a certain point I just got used to it.

Pierre-Louis: [Laughs.]

Ratliff: I’d say, “Well, like, 10 percent of the stuff they tell me is just completely made-up.” But the rest of it is—you just had to figure out what [Laughs], what is and what isn’t. But it’s a strange way to operate a business, and also, like, it’s quite a lonely way to operate a business.

Pierre-Louis: So what does your app actually do?

Ratliff: Our app, it’s called Sloth Surf. I didn’t name it. I’m not that big a fan of [the] name, but ...

Pierre-Louis: Did the AI agents name it?

Ratliff: They named it, yes.

Pierre-Louis: [Laughs.]

Ratliff: It’s a procrastination-avoidance engine, by which I mean you can go to our website—you can go to Sloth Surf—and then if you’re thinking about procrastinating, you would go there and say, “Instead of going to YouTube and watching YouTube videos, I’ll just put into this system: ‘I was about to watch YouTube videos about this, that and the other.’” And then it will send an AI agent to go watch the YouTube videos for you.

Pierre-Louis: [Laughs.]

Ratliff: And then it’ll send you a summary by e-mail of the videos that it watched, and you get back to work. So you’re cutting off the impulse of procrastination, and instead, you’re getting the summary.

Now, of course, you can go to the summary and click through the links and also watch the videos. That’s one of the flaws in the product.

Pierre-Louis: It seems very much like, “I’m gonna consume this information, and then I’m gonna summarize it,” which it feels like that is something large language models are good at, right, relatively speaking. But do you think a company like this has a use case kind of, like, beyond something like this?

Ratliff: Yes. I think our use case was kind of tongue-in-cheek, but actually OpenAI, the maker of ChatGPT, has a product that’s sort of similar to this in concept where, you know, you say, “I’m interested in this, that, and the other topic,” and then AI agents go and sort of collect the information, and then each morning you get a newsletter about that topic. It’s sort of like a Google News alert but on steroids, with something that’s really sort of, like, actively researching a topic for you.

So I think there are lots of ways that AI agents can do that sort of thing—like, go find knowledge for you and then bring it back to you—and I could see all kinds of useful applications for that. Ours is a little bit more meta, I guess, in terms of how we’re approaching it. [Laughs.]

But yeah, it was meant to sort of illustrate that—I think the contradictions in these technologies are both that they can be incredibly powerful but also make things up all the time. And, like, that tension between, like, how useful they are and how stupid they can be is—kind of animates a lot of what we’re trying to investigate.

Pierre-Louis: I also feel like, also as a journalist, so many of my ideas for stories in the past have come from reading something and stumbling across a nugget of information that wasn’t essentially central to the thing that I was reading, right?

And this isn’t something I’ve done journalism on, but this is, like, one rabbit hole that I went down, but years ago I read this book Salt, by Mark Kurlansky ...

Ratliff: Mm-hmm.

Pierre-Louis: No, it wasn’t Salt—it was Cod.

Ratliff: Right.

Pierre-Louis: It was Cod.

Ratliff: Rats.

Pierre-Louis: He, he went through a whole series. [Laughs.]

Ratliff: Yeah, he’s got—he does ’em all.

Pierre-Louis: [Laughs.] It was Cod.

Ratliff: Every time I think of one, it’s like he’s already done—I’m like, “What about rats?” And it’s like, “Oh, he’s done rats.” [Laughs.]

Pierre-Louis: [Laughs.] And there was a line in there about how, you know, when Europeans landed sort of off the coast of Canada in the 1500s, there were dozens of Basque fishing vessels already there. Apparently, the Basque had been fishing for cod off of the coast of Canada even before [Christopher] Columbus. They just didn’t have an interest in colonization. They just cared about fish. They kept their mouth closed and, like, didn’t tell anyone where their cod was coming from because they wanted to protect their fishing grounds. 

And it led me down this, like, wild rabbit hole of, like, Canadian history [Laughs] that I would not have gone through otherwise had I not read—it was, like, literally a sentence in this book. And it’s not the main thrust of the book at all, but it enriched my understanding of sort of North American history so much more. And I feel like that’s the kind of thing an AI summary would ditch because it’s not that important. And I feel like reliance on AI in this way kind of cheats us a little bit.

Ratliff: I agree. I think one of the things that I feel like we’re struggling with as this technology keeps getting better and better and it’s getting used in all these different ways is different versions of that question, like, “What is actually useful to engage your brain with versus to outsource to this—these chatbots?”

And I’m with you. Like, I enjoy the process, something like that, in research or in writing. And so I don’t wanna outsource any of that—like, the serendipity of it, the small details that you come across and also just, like, the pleasure you get from the achievement of finding that, writing about it, whatever it is, and—but I also recognize, like, across the spectrum of people, there are people who feel differently about those things.

And so I feel like, well, for myself I’m going to adopt certain practices, but I have a little bit of a hard time criticizing other people because there are things that I dislike doing that I am happy to off-load, like transcribing my tapes sometimes, you know? But I used to be a person who would say, “I have to transcribe my tapes because then I really know where everything is in the tape.” But now I’m kind of like, “Well, I can just read the transcript.”

So all of which is to say, like, I agree with your premise, and I think it’s really sort of, like, individual right now. Like, we’re all asking ourselves, like, “What’s worth off-loading, and what advantages do I get, and how much do I want those advantages?”

Pierre-Louis: There’s a big gap between choosing to outsource your transcription and, like, I actually stopped using a transcription software because it kept giving me these stupid summaries, and I was like, “The summaries are not useful to me, and I have to, like, work really hard to get around these LLM summaries. All I want is the transcription.” And I think, as a journalist, if I were to rely on the, like, LLM summary, that’s a step too far. Whereas, like, reading the transcript is, like—I’m still doing that work; it’s just faster.

Ratliff: Yeah, I think there are some subtle distinctions. But, like, on the other side, it does just depend on what you care about. Like, for instance, when it comes to LinkedIn, I will admit that I don’t care about LinkedIn posting ...

Pierre-Louis: [Laughs.] Yeah, fair.

Ratliff: Like, their, their argument was, basically, this is what they call “inauthentic engagement” with the platform. But I personally, like, I don’t think the platform is authentic to begin with. I don’t think people are being authentic on the platform. Half of them are also using AI already, so they’re writing their posts with AI.

My point is only that there are people who really love and engage with LinkedIn who would say, like, “It’s not right for you to use an AI agent to write these posts, and I can’t tell which is which.” And that’s kind of how I feel about writing journalism, literature, all these other things.

And so I try to recognize that, like, people in their own domains have different perspectives on this, but I am 100 percent with you in that I don’t use it. I actually—as much as AI is, like, the subject of my work, I don’t actually use it day to day, except in the show. I don’t use it at all. Not for moral reasons—just, like, I’d rather do things myself. I’d prefer to. It feels good. That’s the reason.

Pierre-Louis: Yeah, I briefly used a travel agent that you can subscribe to, and you tell it where you’re going, and it’ll, like, lay out an itinerary for you. And I did it—then I didn’t do anything that it laid out for me, and I realized, like, half the joy of traveling is figuring out what you wanna do.

Ratliff: Mm-hmm, and life.

Pierre-Louis: Yeah, exactly. Like, I guess that’s the question that I have, is in doing this work it definitely feels like there’s an element of at what point are you outsourcing so much to an AI agent that you’re not actually engaging with your real life?

Ratliff: Yeah, I think that is a question, and with your work, too. I mean, we’ve sort of focused mostly in the show on work, but in season one of the show, it was a little bit more about—I was using a clone of myself to talk to friends and family and things like that.

And I think the useful thing, the optimistic belief that I have sometimes, is that it will actually force people to think about this question, to think about, “Well, what do I value in my work? What do I value in my relationships? I actually wanna hold on to that. I’m not looking to outsource that.” I mean, that’s what happens with me when I end up using it a lot for the show, is I think, like, “Actually, I don’t ever want to do that. Like, that’s too important to me,” or “I enjoy that,” or whatever the reason is.

But I think there’s a lot of idle outsourcing happening right now because the tools pop up everywhere you are. You’re writing an e-mail, and suddenly it’s asking you, “Would you like AI [to] enhance your e-mail? Would you like it to be rewritten?” Now, if you’re writing a condolence e-mail to someone who a close family member of theirs has died, like, that’s something to think about. Maybe it’ll help you write that e-mail—how will that person feel receiving that e-mail? Will they know?

I think those are questions that we’re now starting to engage with, and I don’t think it’s as easy as, like, “Well, no one should use it.” That’s not realistic. So my goal is always to, like, “Well, let’s talk about this. Let’s think about this. Let’s see how we each feel about each scenario.”

Pierre-Louis: I think that’s a really good place to end this on. Thank you so much for taking the time to speak with us today.

Ratliff: My pleasure.

Pierre-Louis: And, you know, let us know when you raise your first million. [Laughs.]

Ratliff: Oh, absolutely, yes. I’ll give everybody a ride on my private jet that says “HurumoAI” on the side of it.

Pierre-Louis: [Laughs.]

Feltman: That’s all for today’s episode. You can hear more about Evan’s misadventures with AI on his podcast, Shell Game. We’ll be back on Friday to talk about the science behind one of the wellness industry’s biggest trends right now: peptides.

Science Quickly is produced by me, Rachel Feltman, along with Fonda Mwangi, Sushmita Pathak and Jeff DelViscio. This episode was hosted by Kendra Pierre-Louis and edited by Alex Sugiura. Shayna Posses and Aaron Shattuck fact-check our show. Our theme music was composed by Dominic Smith. Subscribe to Scientific American for more up-to-date and in-depth science news.

For Scientific American, this is Rachel Feltman. See you next time!