Replit president Michele Catasta predicts the first one-person $1B company by end of 2025 and the end of the prompting era

Absolute pleasure. Thanks for having me.

We'll talk soon, Daniel. Cheers.

Goodbye.

Our next guest is the president and head of artificial intelligence at Replet. Michaela is in the waiting room and we're excited to talk about all things live coding. The shift from Vive coding to Vive building. break it down for us. Welcome to the show. How are you doing?

Hey guys, describe me. How are you doing?

Thanks so much for taking the time.

We're doing great.

Uh

uh three-day weekend. What did you build?

Ooh, good question.

I built a sunburn on my back.

I worked hard.

I was flying to Paris.

Oh, okay.

I'm calling you from Paris right now. Yeah. I'm going to be seeing race summit.

Well, thanks for staying up late. Uh so uh why don't you set the table for us, introduce yourself a little bit and uh explain how you fit into the organization and what the top projects are forlet in your world they are these days.

Yeah. So as I said I'm president and head of AI basically means everything technical the company currently rolls up to me.

I'm the person who you know came up with the push of creating rapid agent and led the project from day one and this happened back in September 2024. If you recall, back coding didn't even exist as a term back then. It was coined in February 2025, if I recall correctly. So,

we are the first one to launch in that space and we we basically pioneer them, you know, we chart this new market that keeps expanding like crazy. We had four major agent releases in the meantime. And if anything, I'm more exciting about the mods that are about to come rather than the crazy roller coaster we went through in the last couple of years because the capabilities are expanding so much that naturally the scope we have as a company will just extend like crazy.

So pitch me like it's 3 months from now. It's September and October. Uh and I'm non-technical, someone who has, you know, heard about the AI boom but uh doesn't really know where to start. What are you instructing someone in a few months from now? How should they be interacting with Replet and Replet agents to do something? What's a good hello world on absolute steroids project for them to get started with?

Think about it this way. For the last two years at least, we've been talking about agents all the time.

But it's still extremely difficult for everyone to build their own agents. Mhm.

And in a sense, you know, we've been lucky because, you know, a company like Rapid could grow so much because we know we pioneered the space,

but now I want everyone to be capable of creating their own agents.

And the reason is fairly simple. If I can make a prediction, I do believe that sooner than later, we're going to be seeing far fewer apps being used rather than agents. And the reason is we we built UIs, we built computers in this way in the last decades because the only way we had to interact with them was keyboards and mouse and click and data input which is very much errorprone. The way you work instead is you talk to your colleagues, you discuss, you decide what to do, you express in natural language your intent and then you either delegate or you do you get the job done yourself. Well, guess what? Agents do exactly that. You know, the more autonomous we make them, the more capable we make them, the more they're going to be able to complete tasks end to end and I want everyone to be capable of building their own customized agent. So,

three months from today, if we chat again, I will be mostly talking about that and our product will be pushing that direction very much.

Okay. Talk to me about the split between these three use cases that I see uh coding agents generally being used for. One is um the ad hoc go and do something. Pull a bunch of information. Uh I did this most recently uh looking around my neighborhood looking at different lots and registrations of the government. Go pull a bunch of data, organize it into a spreadsheet, turn that spreadsheet into an infographic, work through all of that like you were a a super powerful executive assistant who knew how to write Python and scrape the web to produce a single report. basically deep research on steroids. Then the second bucket that I see a lot is I want a daily summary. I want you to do something for me that requires AI and LLMs to summarize information, put something together, something that runs continuously. And then the third is go build me a piece of software one time uh deterministic software and then I will interact with that software on a regular basis. So uh Ben Thompson just went on a vibe coding journey. He built a inventory management system for his house. Uh it is a piece of deterministic software that runs uh on his local network and then once he is interacting with it is pure deterministic software. There is no AI inference that's happening really. Maybe it calls out to an API but those three buckets is that a good way to think about the way people are using AI today? What's the breakdown and how do you see it evolving? I think it's a great way and the last bucket you mentioned the deterministic software the fact that you can build a lot of internal tools and you see a lot of solopreneurs creating software 0ero to1 on replet that's where most of the effort went in the last couple of years

I think we're going to be seeing a shift very quickly away from that

apps are not going to be disappearing overnight they still play a role s will play a role in companies for a long long time

y

but it's also through that sound that software can be replaced by agents accomplishing tasks hand to end and that's I think if we

yeah and is that is that is that the fat models like you used to go to an LLM and ask it to do 2 plus two and it would guess four then it was able to write a little bit of Python but and and that's just expanding to the point where the actual LLM itself the the API for the model can do more and more capabilities and the agent harness will be able to do more and more as well. So you essentially instantiate whatever software you need ad hoc on the fly and you expect that to grow.

Exactly. Right. LM have become much better at longer horizon on tasks and you see them working like even for hours in a row.

Yep. The additional layer that we put on top of them the agent harnesses are such that you can also introduce verification criteria so that for example you come up with a high level goal say you built an app on rapid and then you want to optimize the conversion rate and you set a 5% as a goal but guess what you can have an agent doing that for you don't have to prompt it on the specifics on how to make that happen. It will connect to your app. the agent will decide to improve conversions, maybe I need to change the copy of the website, maybe I need different images, maybe I need like a different pricing page

and then you keep measuring until it accomplishes the goal. That's what you have been hearing more and more being talked, you know, in the industry as loops.

And I would call this like the postprompting era.

We don't put much effort in prompting anymore. Even if you hear entropic researchers and engineers explaining how to use table 5.

Mhm.

The main insight is stop giving it prescriptions, stop giving it rules, just tell the model what you want to accomplish and then trust the process.

Yeah.

Well, that's the type of product I think we're going to be seeing build more and more in the future.

Mhm. The goal finding Jord finding. Exactly. K, you you mentioned going going back to your earlier call out on uh c you know people building customized agents. What uh are are these are these is this like entirely personal software or do you see this as like one person companies where they'll build an agent and sell it to other people? Is it both? I'm having trouble uh I've been just experimenting lately and trying to every time I think oh I should try to use oh I need I need some information from an app. So like I use surf report as an example

like I'll just ask chat hey what's the surf report for this spot.

Yeah

and it does it quite well. You can do it for the weather. You can do it for so many different apps. And I expect that agents as they get better and better will just be able to

anything that I could imagine a custom agent to do personally,

the agent

the models can do.

And I've seen this a lot where people will go and build they'll vibe code a solution to something that the model can do natively. And then you'll be like, "Yeah, I have a piece of deterministics software that does all it summarizes all my emails." And it's like, well, you can actually just integrate your email directly and ask.

Yeah. And so part of me part of me is like I just I don't know if I'm I have I'm being creative enough to think of use cases uh or

or what I think the rule of thumb to be as creative as possible is you need to think in general terms.

You don't have to narrow down on I want an agent that asks me to come back to the few applications that I use on a daily basis. You need to start to think in goals. Mhm.

And the moment you have a goal and you want to get some type of work done, then allow the agent to decide what are the steps necessary and allow the agent to create the code in order to make that happen.

Now there's one advantage only if you actually create a piece of deterministic software is that you trade that in flexibility for the cost of tokens.

Yeah. So every time you just allow the model to do the majority of the work, you're actually spending quite a bit of money, especially if you need to use frontier capabilities. So a lot of the tools, the internal tools built on rapidly today, they tend to be very deterministic also for the reason because if they start to be used by hundreds of thousands of people in a company, you don't want them to, you know, be hitting the frontier models otherwise it's, you know, financially irresponsible to do that. Uh, last question for me. Uh, get me up to speed on the state of soloreneurship and I want a prediction from you on when you think we'll see the first oneperson 1 billion dollar company.

It doesn't really happen. You probably read about that on the New York Times. Controversy.

I don't think that one counted. I don't think it doesn't count. Okay. I I it was close, but I it there were too many too many complaints from uh customers. It seemed like they bent the rules of the marketing and they were getting hit with some FTC FDA uh issues. Uh I want I want a clean fing$1 billion valuation.

Let's give you another shot. I would say by the end of the year we have several

Yeah.

Yeah. We have several nine figures company built on rabbit. So, he's going to help them.

Yeah,

we're almost there.

I love it. I love it. Well, uh yeah, definitely shoot us a note when it happens. If you see it inside rapid,