So I think now we are just truly seeing this unlock where people who were really close to the problem, domain experts, but have been blocked by, you know, the technology barrier to sort of really express themselves, are using Emergent to sort of build these things out.
There's just so much focus on AI is going to replace jobs, knowledge work is going away, like what's that going to mean for employment and civil unrest, but like no one's really talking about the fact that actually like if you have like some agency of interest, you want to start your own business and have autonomy over your life, like you are empowering that at scale.
Welcome back to another episode of the Light Cone. Unfortunately, Garry got called to jury duty and can't be here with us today. But we are really excited to be joined by Makund and Madav Jha. They're both twin brothers and founders of Emergent, which went through YC in summer 2024. Emergent is a platform that lets anyone build and ship production-ready software using AI agents. You guys are actually one of the fastest-growing companies I believe YC's ever funded. I mean, the statistics you were telling us were mind-blowing. You have in 8 months since launch 7 million apps have been built with Emergent. Walk us through this like incredible growth you're seeing. Actually, when did that hit a real inflection point and how did that feel for you guys?
So we both are twin brothers. We actually, you know, started programming when we were age 12. Both of us came to the US to do our PhDs. I dropped out of the PhD program, joined Google, and Madav went on to — was in Zens, then went on to start the deep learning team at Amazon. And we've been meaning to do a startup together for a long time. And before this, I was running a startup in India called Dunzo, which was a hyperlocal quick commerce company.
And Dunzo was a big company actually, right?
Yeah, it was really big, and we were almost a verb in India. So when people ship things, they say 'Dunzo it.' And I was managing a really large team of 300 engineers. And we had been sort of watching the deep learning field for a while, and we knew an inflection point is coming. One of the things that I observed when I was running this large engineering team was that software testing was the biggest bottleneck in shipping fast. So when we started looking at, you know, what we want to build in AI, that was the first idea we actually...
This was '23 end. Yeah. And so when we applied to YC, like we applied with this idea of automating software testing. That was the first idea. In fact, we went to a lot of VCs with this idea. They thought it was too crazy. You know, and now looking back, it almost looks funny. And so we applied to YC with this idea, and when we were building these testing agents, we realized that if you can solve for verification, which is essentially, you know, you can solve the testing part, you can actually automate all of software engineering. That was sort of our key insight, that like, you know, verification is the loop which sort of keeps the agent running for a longer period of time. And that's when we pivoted to looking at general coding agent as a space, and we started building a general coding agent.
And this takes us into 2024.
Yeah, tell us what the landscape looked like. Like, how big was Lovable at this point, and just...
I mean, nobody had started — Lovable had not started. I think Cursor was just getting started, and very, very early. I think Devin had just come out. So really, really early. And we looked at this benchmark called SWE-bench, which is essentially a benchmark — now it's saturated, but at that point of time, that was the benchmark where all of the coding agents were getting measured on. And we took on this challenge of becoming number one on that benchmark. And like, we sort of packed ourselves in a room, four of us, and said, 'Okay, let's just look at this benchmark, how do we crack it?' That sort of set the foundation for Emergent. And we built, you know, SOTA coding agents, which became world number one on SWE-bench, you know, in two months of time. And that was the time when we sort of discovered a lot of the fundamental truths about building with LLMs, building with agents.
Your intended users at this point were, we're presuming, engineers.
Yeah. At that point we were like purely just a research company, just building coding agents. We were not thinking about a product. There was a time when we sort of invented the multi-agent system. We invented memory. We invented like, how do we do agent-to-agent communication? How do you scale up test-time compute? A lot of those things which like were sort of coming out — like we would discover something and we'll see three months later something come out in a paper. You know, and that sort of set the foundation for us to...
So we were like Claude Code before Claude Code was a thing.
Yeah, a bunch of the paradigms like multi-agent orchestration, how do you use like different routings, a lot of those things we sort of discovered.
I definitely want to come back to that. I'm curious at this point in the story though, when did you sort of pivot into becoming a tool for non-technical users?
Yeah. So we actually, like, once we had this coding agent, we actually went the enterprise route. That was the common wisdom at that point, that hey, like, go to enterprise, build for enterprise. And we spent like two, three months trying to, you know, make our agents work within enterprise. We found that it was too slow. And at the same time, we internally started using Emergent's platform to build internal tools and internal software. And at that point, you know, we saw like Lovable was growing like crazy, Bolt was growing like crazy. So we thought, hey, why don't we have this, you know, really strong coding agent, how do we sort of package it and bring it out in the world? And we launched a very like small beta pilot almost in June last year, 2025, and that really took off. And since then, you know, like we've been just focused on solving problems for non-consumer. We in fact thought a lot of technical people use us, but today 80% of users who are on the platform are non-technical users with zero programming knowledge. And they're building like apps that run real businesses on top of today. So it's almost...
And they're based all around the world, right? Like, how many countries?
Yeah. So they're all global audience, 80% — 70, 80% are in US, Europe, over 190 countries right now.
Something that we have talked a bunch about at YC internally is just how does first-mover advantage versus second-mover advantage play out in the AI world. Certainly something that we've noticed, like if we look at some of our companies, like Lora entered the legal AI space after Harvey but is like growing incredibly fast. So there clearly wasn't maybe as big of a moat around being a first mover as you traditionally think there is in software. When you guys made that sort of pivot or the slight change in direction into non-technical users at a time where Lovable and Bolt are growing really, really quickly, how did you think about that?
There are like two, three different threads I would want to pull. One essentially is that I think the model — every new model generation actually is presenting a new opportunity of looking at the world. Like for example, when we started, GPT-4 was the first model that we sort of started looking at. And at the end of the day, the biggest problem that everybody was trying to solve was JSON parsing, like, hey, structured output format. And we thought, okay, like, the next model is going to solve for it, you know, like, let's not spend time on that. And I think with every new model, what's happening is that you need to start reimagining the world. For example, like, Opus is a different class of model right now. It's going to enable extremely long-horizon tasks. It's going to enable like multiple agents coordinating together. And so I think like one of the advantages of starting second, right, is that you can actually, one, learn from what is not working for the current competition, right? And also I think you fundamentally start from a different starting point, right? Like where your aperture of the world is like very different, like your imagination is really big, right? And I think when we were starting Emergent, we realized that like a lot of the users that were going to, you know, some of these apps, they wanted to actually really build an app that works, right? And most of these were actually like really, really optimized for front-end prototyping at that point. So we started fundamentally reimagining that, okay, what would the world look like if you could actually ship things to production. And our key insight was that to automate all of software engineering, you will have to build a platform that replicates what the best engineering teams do, like code reviews, automated testing, debugging, deployment, security, hosting. So we reimagined the entire platform from ground up, saying what would an end-to-end platform look like, and the real user need was actually to ship the product, not just the front-end prototyping. I think the second thing is like how do you sort of get the distribution because you're coming from behind, right? So even if your product is really, really strong, fundamentally I think you'll have to enter the market with a really, really strong product which is, you know, head and shoulders above what exists in the market today for people to take notice. We were very confident about the product, and so a lot of our focus, like in early days once we sort of launched, was on how do we sort of rapidly scale up distribution. We built out a large influencer network, and that was our initial sort of starting point for us. Like we used TikTok, Instagram, and partnered with a bunch of influencers to really, really spread the word out, and that sort of kickstarted the whole thing for us.
To me, so building the influencer marketing engine is like tactics to land grab. Like, were you also thinking about just focusing on personas and specific subtypes of users you wanted to go after that weren't being targeted by Lovable or others, or Emergent was a better fit for them?
I mean, our thesis was that like there are a lot of users who would want to build serious applications, right? And that was our sort of target audience. And a lot of our marketing, a lot of our initial messaging was around that, like, hey, come and ship real software. What we did was like a little bit broad-based marketing, but users that, you know, were coming to the platform that we would convert were users who actually wanted to ship a real app on the platform.
And was that in the messaging then?
It was in the messaging, yeah. So we would say, come and build real apps. We would also use the common errors that you would see on other platforms, you know, like, hey, don't face this error on Emergent.
It seems like a key insight for you. Basically, you went very hardcore in terms of being maximalists in engineering from your experience having run large engineering teams of 300 engineers, having worked on deep learning teams at Amazon, you really knew how to architect the systems. Can you maybe share a bit how you built it? One of the cons of all these other big products like Lovable or Bolt is just that it's difficult to get those into a fully usable state. You can get to a prototype very quickly, but yours — you went zero to 100% very quickly. And that takes finesse. It's almost like that 20% gets 80%...
effort, like the Pareto principle, but you did more than that. The last 20% of that engineering to production was a lot of work. And that's a lot.
Yeah. And I think like the last mile that you mentioned, right, is always what people neglect, that hey, you need to make sure that not only the app gets built, it also gets deployed. And this is one of the conscious reasons why we chose to build our own infra on which the agent is running. So like we provide, you know, cloud sandboxes, we don't outsource it to like some third-party sandbox provider, which was also pretty popular at that time, right? So we built our own Kubernetes tech stack from ground up, the container tech stack. And one of the insights here is that if you give your agents the same infra during the build time and the same infra during the deploy time, then during this deployment phase you don't encounter those many problems, right? And the fact that we have our own infra also allows us to give rapid feedback to the agent. So your agent is only as good as the feedback that you provide. So we built this infra and agent sort of co-built it together from day one. And to your point, right, because we focused on building ship-ready apps which are production-ready, which come with backend and frontend and everything, the tech stack we chose was also pretty unique to us. We have a Python backend server. We have a React frontend server. Like most people would typically go with a much more Node-focused, Node-heavy tech stack, right? And this server-client architecture where you can have background jobs if you want to have background queues. So we knew that users who would use this app, their ambitions are going to go bigger and bigger, right? Hey, I want to run a job which can do this asynchronous video processing, you know, and they're going to prompt it, and we wanted to support it from day one, right? And so it's the same tech stack on which Emergent is built is what we expose to our end users, is what we expose to our agents. Okay. On the agent side, we were very early on the multi-agent architecture. So we knew that you want to be very frugal about your context management. So what you do is, hey, let the main agent, the driving agent, handle the main routine. But any delegated task that you want to delegate, you delegate to a sub-agent. Be it like testing, be it like, hey, I want to do a design search, or I want to do like, you know, integration search, like how do I integrate this unique API. And along the way, when we were doing all of this, we were able to figure out, okay, all the trajectories that we are generating, we can kind of aggregate over time and sort of build in a long-term memory for the agent, which is very unique in the sense that your agent learns not just from your own session, it learns across the sessions. This is something I would say is one variant of continual learning that people are interested in now. You would have noticed that people are interested in skills, like people create skills, and there's a new benchmark called SkillsBench which shows like agents with skills outperform agents without skills. And interestingly, like those skills cannot be generated by agents themselves. Like if you generate those skills by agents, they don't match up to the performance. So we were able to do it in a way where the skills get auto-generated based on previous trajectories, and we run it through a CI/CD process and then add it to the long-term memory. So all of that compounds for us, right? So if your agent was struggling to do a calendar integration three weeks ago, today it is no longer struggling thanks to the previous session where it was able to make it happen.
So fascinating. So it learns on its own because I think one of the challenges of all these vibe coding app platforms is at some point the applications would get so complex that if you build it very simply you would run out of the context window for all the models because that seemed to be the bottleneck. And I think you guys architected your way out. So you kind of built a lot of what the state-of-the-art is now, but way back a year before.
Our coding agent is so powerful that we basically internally use it as a replacement for Claude Code as developers, right? So we are so proud of that. But yet we don't want to expose that sort of power tool to our end non-technical user. And so even though we have this VS Code editor, we kind of hide it because what we have noticed is that non-technical users, they even get panicked as soon as they see a diff, you know. We had a fairly technical PM in our team and like he doesn't like JSON, you know, he's like, no, don't show me, you know, I get intimidated.
So building that user empathy, where you have that user empathy and building that agent empathy, you also have to empathize with your agents. What is the agent feeling like, right?
We internally have a term called agent experience, right, that we measure, how is the agent's experience on the platform.
Actually a really important point I think people don't realize is you guys actually — you actually started out essentially as sort of Devin, Cursor in like the actual coding agent world for engineers. You just made the choice to package it up for non-technical users. So you're sort of like moving almost in the opposite direction from like Lovable, Bolt. Like you have the power, you have all of the actual power. You just need to simplify the user experience, whereas they sort of start with the user experience and they're going to have to develop the power over time.
Right. And I think fundamentally it's like unless you start from a starting point which solves all of these problems along the line, the whole software development life cycle, it's actually really hard to come from the other side and solve these problems because you'll make some architectural choices which are very hard to reverse.
Do you have any more — I'm really curious, like any more examples of where sort of as you were engineering the system you just trusted the model? Like you mentioned JSON parsing, but was there anything else where you were like, let's not invest time in that because like Opus 4.5 will solve it?
I mean, some of them has been, for example, you know, like library definition, some of the integrations that we have sort of built, like, you know, we think that, you know, next sort of models are solving for us. Similarly, like how do you generate unit tests, some of those things that we would have heavily prompted before. And the other thing that we are very conscious of is how do we give more and more autonomy to the models as the next generations come out. And the more autonomy you're able to give to the models, the better they perform. Like initially, like our harness was very strict and, you know, like we would tighten it up. And slowly, like what we were observing is that as these models are getting larger and larger, more efficient, like, you know, like the more control you give to the model, the better the harness gets.
If we extrapolate that out or sort of like really far out, are you worried about where that sort of leaves you as a company versus the models themselves and the models get more powerful?
Yeah, I think there is this underlying current right now, right, in the industry that, hey, like, is Anthropic going to eat everybody up.
Yeah, I mean, our view is that I think the coding aspect is only 20% of the job, right? I think like taking an app to production is really, really hard. And I think what matters is how closely are you working with the user? How well do you understand their needs? And I think as the models are going to get more and more capable, I think the human desire is also continuously growing at the same rate. So I think people are going to want to build more complex apps on the platform. The other thing is that at least with our harness, we're able to extract 20, 30% more on top of these models. And essentially, like, we can use multiple foundation models together to sort of extract more. And I think we'll have to keep continuing, you know, like delivering more and more things to our users. For example, now we're thinking about a lot of our users who have built the app now want help with distribution, now want help with growth, now want help with like how do you sort of manage users and things like that. And I think for us the spectrum sort of keeps growing on that side.
I agree with it. I mean, there's another graph that I shared recently, it's just like the number of software engineering positions available is actually going up, right? And I feel like at least internally at YC we're experiencing this. It's like the more powerful the tools get, the more ideas you get, and the more work you want to do. And it just feels like everyone here is working like more hours, doing more stuff, and it's just the rate of like software that you're expected to ship per week just keeps going up and up and up. It's...
accelerating. Yeah. It's a hedonistic adaptation to, you know, like, hey, oh, this is more powerful now, I can do more work.
Yeah, it is really Jevons paradox at play. And I think there's a lot of concerns like, oh, the software engineering jobs will be gone. I don't think that's the case. I mean, based on everything that you're telling us and what we're experiencing.
I mean, I think we're in an expanding market, right? Like we are letting non-developers be developers, right? I think, you know, that market is expanding. We also are internally seeing like the roles sort of combining. So like a PM, a designer, an engineer, like a single person is doing, you know, like the work of all three together, right? So like we have a PM who's vibe coding internally things. And recently, like we — so we are seeing this internally right now where a lot of the work that was done by like a five, six people team can now be just done by like a single engineer or a single PM.
YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com/apply. It's never too early, and filling out the app will level up your idea. Okay, back to the video.
Could we see a demo of Emergent?
Oh yeah, sure. Yeah. So this is what Emergent's interface looks like, and I'm going to put a prompt where, like, because we were coming for this podcast, I thought like, you know, there should be an app which lets you practice, you know, podcast questions, or maybe you are going to a job interview and you want to practice questions, right? So you can build a full-stack app on Emergent, you can build a mobile app. Our prompt engine is smart enough that once you give it a prompt, it will figure out that this is talking about a mobile app. So it'll figure out like, hey, the right agent to use is a mobile app builder.
So even though you have like selected the wrong tab, it's just like...
Yeah, behind the scenes auto. Yeah, I got you. Right. So while this is running, let me quickly also show you a few user apps. So this is by somebody based out of Illinois. He sort of has a business of audio-video setup that they do like on-site manually, right? So basically whatever this kind of intake form they would have taken through spreadsheet and other calls, they basically built this out without any coding background knowledge, right? Like, hey, this is the kind of AV setup I want. So you go and you build your room and then you get — it's a lead-gen sort of a form, but this is a fairly full-stack app.
One thing I noticed about that is like the design is really good, like the icons, like it just looks like a well-designed app.
So we have actually spent a lot of time on like making sure the design is actually good. And like, so earlier there used to be a big trade-off between design and functionality, like if you're optimizing for design, like your functionality would not be that strong. And so we had to figure out like how do we sort of share the context in a way where design also gets better.
There's another sort of person based out of Norway. He sold his previous business to a PE and realized how much lawyers have to struggle with spreadsheets and other things. So he...
built a CRM for lawyers. He describes himself as like a business developer — I like the word he used, 'I'm a business developer.' He doesn't have a programming background. So a lot of CRM-related apps we are seeing for small businesses, it's your second monetization avenue, right? And so one of the unique things about Emergent is that before the agent goes off to build things, it asks you for some clarification because the agent wants to make sure that it understood your requirements properly. And another thing is that non-technical users probably don't know the concept of an API key. 'How do I get an OpenAI API key?' So in this particular case I can just say, 'Hey, use the Emergent LLM key,' so you don't have to worry about getting an API key from a third party.
This feels like a good example of what you were saying, because this is sort of like the ask-a-question skill, include code, but you just abstract that away and build it into the experience for someone who had no idea about it.
Absolutely, I can be very casual here. I can say, 'Hey, for the first one use the Emergent API key, for the rest assume good defaults and then go.' This is the first time I hand off the agent, and at this point I can just close my laptop. We also have a mobile app, so you can on the go keep prompting the agent if it requires additional things. Once it's done, you see a preview of your app. So here, for example, in this case I can practice, what is my origin story? I can record what my origin story is and I can keep going to various questions eventually.
