How Emergent is making app building more accessible with Claude

1. 0:00 – 0:21

   Emergent’s mission: AI-built software for the “million niches” of small business

   1. MJMukund Jha0:00

      We are building for like small businesses today, right? Small businesses globally account for seventy percent of employment, accounts for like fifty percent of GDP of the world, right? But they've never had the tools to sort of really express themselves because, like, there are a million niches in small businesses. But for the first time with AI, you can actually serve all of these million niches at zero marginal cost, and that's what we're trying to enable.

2. 0:21 – 1:22

   From sibling founders to YC: the original bet on automating software testing

   1. CRCarly Ryan0:21

      Hi, my name is Carly, and I'm on the Applied AI team here at Anthropic, specifically working with startups. Today, Mukund has joined me. He is the CEO and co-founder of Emergent. So Mukund, let's bring it back to the beginning. Emergent's growth has been one of the standout stories of this year. Why don't you walk me through your journey with YC, some of your pivots, and how you landed on the product that you have today?

   2. MJMukund Jha0:42

      Thank you for having me here. Super excited to be here. So I'm doing this startup with my twin brother. Madhav and I actually started programming at age twelve.

   3. CRCarly Ryan0:49

      Love it.

   4. MJMukund Jha0:49

      And we have been-

   5. CRCarly Ryan0:49

      We love, we love sibling founders here.

   6. MJMukund Jha0:51

      [laughs] Totally. And we actually, like, have been obsessed with this problem of automating software engineering from the beginning. Before this, I was running a startup called Dunzo where I had a five hundred people engineering team. And when we entered YC, our first sort of insight was to automate software testing, because I'd seen in, in at Dunzo that how software testing was a bot-bottleneck for shipping software fast. And we started with this problem of like how, hey, how do we automate all of the software testing, uh, including mobile apps, web apps? And when we sort of started building that out, we realized that to automate testing,

3. 1:22 – 1:56

   Pivot to general coding agents: verification loops, multi-agent systems, and long-running autonomy

   1. MJMukund Jha1:22

      we had to build, you know, our own container technology, our own coding agents.

   2. CRCarly Ryan1:25

      Yeah.

   3. MJMukund Jha1:25

      And we stumbled upon this insight that if you can actually solve for the verification loop, you can actually make agents run longer. On the day one, we told our YC partners that, "Hey, like testing is great, but you know, we want to now build general coding agents." We were like almost like a applied research lab working, just building, you know, uh, high production grade quality agents. Invented a bunch of things like test time compute, how do we sort of scale the memory across agents, how do we solve our agent-to-agent communication. We were one of the first team to productionize multi-agent systems. When we launched, our first approach was to go to enterprise.

   4. CRCarly Ryan1:56

      Yeah.

4. 1:56 – 2:14

   Enterprise wasn’t the wedge: internal usage revealed the real opportunity—democratization

   1. MJMukund Jha1:56

      We had an enterprise customer, started working with them, but realized that the adoption in enterprise is gonna be slow. At the same time, we started using these internal coding agents to build everything internally. In fact, like there were like few non-technical people on the team who were using it more than developers were at that point.

   2. CRCarly Ryan2:09

      Hmm.

   3. MJMukund Jha2:10

      And we truly realized that the power is actually to democratize software engineering for everyone.

   4. CRCarly Ryan2:14

      Yeah.

5. 2:14 – 3:10

   Beyond prototypes: production-ready app building with testing, security, deployment, and reviews

   1. MJMukund Jha2:14

      And that's when we, in June, launched this almost like a research preview, and that time agentic coding was not that popular. And when we launched, it just took off and, you know, we were actually surprised. You know, everybody else at that time was building for front-end heavy applications.

   2. CRCarly Ryan2:28

      Hmm.

   3. MJMukund Jha2:28

      They were building for more demo kind of things. And we had fundamentally approached the problem thinking that what would it take for actually anybody to ship production-ready software that will actually have real use cases, will have business-critical apps running. And our insight was to sort of automate all of this testing, linting, deployment, security, um, how do you solve for like code reviews. And that actually like really, really resonated with users because most of the users that who are coming to Emergent today are serious builders. These are small businesses, entrepreneurs who actually want to see an economic value from the software they're getting built.

   4. CRCarly Ryan3:01

      Yeah. That's awesome. So it's like basically democratizing, it's taking natural language, it's for non-software engineers, but it's also building durable products.

   5. MJMukund Jha3:10

      Right.

6. 3:10 – 5:20

   Why Claude (Sonnet/Opus): instruction-following, coding strength, and outcome-first quality

   1. CRCarly Ryan3:10

      It's not just for prototyping. It's really cool. So you've been building on Claude since the really early days.

   2. MJMukund Jha3:15

      Right.

   3. CRCarly Ryan3:15

      We met about a year ago. What made that the default choice, and what's kept that all the way through? Why have you stayed on Claude?

   4. MJMukund Jha3:21

      Yeah, I mean, I think we were almost lucky to get started when Sonnet came out.

   5. CRCarly Ryan3:24

      Yeah.

   6. MJMukund Jha3:25

      Uh, and, um, and I think in AI, I think every- everybody needs to sort of imagine the world six months ahead and build for that versus building for what's available today. Imagine the world from the lens of what is available and what, what's gonna come next, right? So, uh, people who would have started with the previous generation of models-

   7. CRCarly Ryan3:39

      Yep

   8. MJMukund Jha3:40

      ... would be solving a different problem versus us, because for the first time, like, I think Sonnet had really good instruction following so that you could actually really imagine what long-running agents would look like. It was really good at programming. It was really good at front end as well. Um, and so that actually like really gave us, you know, a little bit of a futuristic view of what the world of programming could look like and what could agentic programming look like. And our users are actually like comparing us against a dev shop, right?

   9. CRCarly Ryan4:03

      Yeah.

   10. MJMukund Jha4:03

       So they're not comparing us against like an IDE or a tool. And to them, outcome matters a lot, right?

   11. CRCarly Ryan4:09

       Yeah.

   12. MJMukund Jha4:09

       And so we are really, really focused on what would give us the highest quality, what would give us the highest probability of getting to the right outcome versus, uh, speed. Most of the other players who are focused on front end or design, to them, speed matters a lot. But for us and our users, what is important is that, that are they able to reach the outcome that they want. And that's why, like, Opus has been sort of, you know, our, our workhorse and, and it's really good, great at instruction following. And if you have the right tight feedback loops that we have been able to build using our multi-agent systems, you can actually keep them running for a much longer time.

   13. CRCarly Ryan4:37

       Yeah.

   14. MJMukund Jha4:37

       Today, our agents can run for hours trying to build the software. And, uh, we, we run like very tight evals internally. And, and of course, like you've been really helpful in-

   15. CRCarly Ryan4:45

       [laughs]

   16. MJMukund Jha4:45

       ... helping us.

   17. CRCarly Ryan4:46

       I know about your evals, for sure. [laughs]

   18. MJMukund Jha4:47

       Yeah. Yeah, and I remember like, you know, when we were about to launch, we had this like two AM meeting just-

   19. CRCarly Ryan4:51

       Yeah

   20. MJMukund Jha4:51

       ... to get started on things.

   21. CRCarly Ryan4:53

       Totally. Well, I mean, so, so a lot of your engineers are in India. So one day when I was like, it was like Friday at two PM for, for us, and it was the middle of the night for you guys.

   22. MJMukund Jha5:00

       Right. Right.

   23. CRCarly Ryan5:00

       And I was like, "Oh, of course, they're gonna answer the next day." And then one of your engineers was like, "Oh yeah, let's like, let's get on a call in fifteen."

   24. MJMukund Jha5:06

       Right.

   25. CRCarly Ryan5:06

       And I was like, "Is it not the middle of the night?"

   26. MJMukund Jha5:08

       [laughs]

   27. CRCarly Ryan5:08

       [laughs] So your team during YC-

   28. MJMukund Jha5:10

       Right

   29. CRCarly Ryan5:10

       ... built this internal coding agent. Maybe you didn't know what it would become. What did you learn from that, and like what shows up today in the product, and then maybe what are learnings that like you really scrapped? [laughs]

7. 5:20 – 6:28

   Owning the entire stack: containers, harness quality, and tight feedback loops

   1. MJMukund Jha5:20

      Yeah, I think our first insight was to, you know, to really, really build an automated software engineering platform-

   2. CRCarly Ryan5:25

      Mm-hmm

   3. MJMukund Jha5:25

      ... you will have to own every layer of the stack.

   4. CRCarly Ryan5:27

      Yep.

   5. MJMukund Jha5:27

      Right? And you have to almost co-build the infrastructure to support your agent and an agent to sort of support the infrastructure.

   6. CRCarly Ryan5:32

      Hmm.

   7. MJMukund Jha5:33

      And at that time, everybody took the shortcut, you know, hey, what is the third party that is available? Let's use that. Uh, when we started, like there were no good container technologies.

   8. CRCarly Ryan5:40

      Yeah.

   9. MJMukund Jha5:40

      We had to actually really invent and build sort of our own container technology on top of Kubernetes, which does this snapshotting, memory snapshotting, so you could actually save the states and run parallel agents. And for us, what's been sort of really, really insightful has been that agent is the product, right, in many ways.

   10. CRCarly Ryan5:54

       Yep.

   11. MJMukund Jha5:55

       And like the harness quality really, really matters.

   12. CRCarly Ryan5:57

       Yeah.

   13. MJMukund Jha5:57

       And we sort of built this multi-agent system. We were one of the first teams to sort of productionize it, and many learnings, including like how do you sort of save for memory, how do you do agent-to-agent communication, how do you do better context management, how do you effectively use caching to keep the cost down, um, and tightening all the feedback loops from the container and all of the infrastructure that you've built, like is really important. The other thing that really was important to us from day one was to own the entire stack, because the last mile is where like most solutions trip. We spend a lot of time just to make sure that our production quality is really

8. 6:28 – 7:19

   Production data flywheel: deep logs, long-term memory across apps, and faster learning

   1. MJMukund Jha6:28

      high, and what you can see in the development environment is something that replicates in the production. And-Today, I think, I think we're probably one of the deepest logs in app building space, and that allows us to sort of compound our learning. For example, like a lot of the errors actually don't show up in the dev side. Like, they'll only show when your app is live and real users are using it. We're able to pipe all of that back to our development environment and allow our agents to learn. We also have built this, um, what we internally call long-term memory, where an agent is not just learning from within the user session, but is learning across all of the apps that are getting built. So first time it sees an error, a new error, or a new library upgrade, it's able to learn that and, and sort of, you know, do that in the next iteration in much less tokens with much more accuracy. So I think, like building this co-agent and infrastructure together, like has been really, really helpful for us.

   2. CRCarly Ryan7:10

      And I guess because you own the whole stack, like up to the hosting as well, that gives you kind of all this proprietary data that you can feed into the next parts of the agent and the future of the agent.

   3. MJMukund Jha7:19

      Right. Right.

9. 7:19 – 9:01

   Model selection when users can’t debug: prioritize reliability, not latency or cost

   1. CRCarly Ryan7:19

      How do you think about model selection, especially when your users can't debug the output?

   2. MJMukund Jha7:23

      I mean, a lot of the burden lies on us today, right? Like to make our users successful, and we take that very, very seriously and, and that's what we are sort of com-- like extremely obsessed about like, how do we get users success? A couple of months back our like, you know, deployment rates were eighty-four percent, now it's closer to ninety-eight percent.

   3. CRCarly Ryan7:36

      Cool.

   4. MJMukund Jha7:37

      So ninety-eight percent of people on the platform were able to build something, are able to deploy to productions. That's, that's been amazing. And that-that's just tightening all of these feedback loops, making sure that, you know, your agents are extremely reliable and self-learning. In terms of model, like the way we think about it, we are getting compared to a dev shop, right? So when a user comes to us, they're comparing us to like a two fifty thousand dollar price point, uh, that they would have paid if they had gone and, and hired a dev shop. So they are not price sensitive. They're also not like latency sensitive that much because they're comparing us to a three months project that would have been executed outside. Uh, so we are very, very obsessed on the quality of the product and not so much about the cost as much. And so for us, like being on the frontier, the, the highest reasoning, uh, possible is, is something that we sort of index a lot on. Internally-

   5. CRCarly Ryan8:17

      Yeah

   6. MJMukund Jha8:17

      ... we index a lot on like just highest quality output versus, you know, the fastest output.

   7. CRCarly Ryan8:21

      Yeah.

   8. MJMukund Jha8:21

      And we are not trying to one-shot a solution. We're trying to, you know, build really, really reliable production-ready system that can be iterated upon. Instruction following should be really good-

   9. CRCarly Ryan8:29

      Yeah

   10. MJMukund Jha8:29

       ... because the error sort of compounds and even small sort of errors as you run ten thousand steps, like they, they compound very quickly.

   11. CRCarly Ryan8:35

       Totally.

   12. MJMukund Jha8:35

       And because the outcome matters the most, right? So we sort of, you know, rely on the best possible models for reasoning and coding abilities.

   13. CRCarly Ryan8:42

       Yeah. Yeah, I think that's been a through line that I've observed with you guys where it's like you really put accuracy top-

   14. MJMukund Jha8:48

       Mm-hmm

   15. CRCarly Ryan8:48

       ... and like that's why you've had so much success. So this is a question we ask our founders often. The AI builder space has gotten a bit crowded. So curious to hear from you, what has Emergent built that's genuinely hard to replicate?

10. 9:01 – 10:58

    What’s hard to replicate: customer focus, proprietary data/infra, vertical stack, and full lifecycle delivery

    1. MJMukund Jha9:01

       Yeah, I mean, I think first of all, like I-I think we are beginning of a very, very large market that's opening up, right? I think the market is actually growing faster than all of us can serve right now, right? And-

    2. CRCarly Ryan9:09

       The pie is getting bigger.

    3. MJMukund Jha9:11

       Way bigger, way faster-

    4. CRCarly Ryan9:12

       Yeah. Yeah

    5. MJMukund Jha9:12

       ... than all of us can serve right now. So we are actually not so, so much worried about like, hey, what is competition doing? You know, what are labs doing? We just generally think that the market is so big and it is growing so fast that us being able to serve the customer that we want to serve is, is generally sort of now really, really important. For us in-- on the technical side, I think like, you know, our approach has always been to sort of really, really build very closely with the customer. So we understand our customer really, really well. You know, what their needs are. For example, they really don't care about, you know, demos, front-end prototypes. They care about like real production use cases.

    6. CRCarly Ryan9:39

       Mm. Mm-hmm.

    7. MJMukund Jha9:39

       They care about third-party integrations working. They care about payments getting through. They care about like, you know, authentication working. A lot of our users, you know, early on started uploading like really large files. So we had to sort of think about like, hey, how do we support large file systems on our production systems? They care about scalability. So when we write code, we make sure that these code are written in a horizontally scalable manner.

    8. CRCarly Ryan9:59

       Yeah.

    9. MJMukund Jha9:59

       So as traffic grows, the infrastructure sort of grows with them, right? For us, like I think the technical moat is, one, our sort of agent quality, the harness quality is really, really strong, and it's continuously evolving. Second is all of the provided data that we are collecting right now, right? I think all of that feeds into our long-term memory, feeds into, you know, like our self-learning agents, feeds into our RL systems. And because we are solving only for one stack, we are almost like a vertical agent for coding because our users don't really care about what technical choices is being made. So they leave that onus on us, and we make the right choice for them, and we are solving for just one kind of tech stack. So that allows us to sort of really, really accelerate our sort of learnings on that tech stack. And lastly, I think the code generation is only twenty percent of the problem, right? Like the, the eighty percent of the problem is actually how do you take it to the, uh, deployment, how do you make sure it's maintained in production? How do you make sure that security is really, really high?

    10. CRCarly Ryan10:46

        Yeah.

    11. MJMukund Jha10:46

        And we are solving for all of those deep infrastructure problem where, you know, not only your code is written, but does it work in the dev environment when it works in dev? Does it replicate in the production? Do you have the feedback loop? And solving for the full stack is like really, really important for us.

11. 10:58 – 13:32

    Multi-agent workflow for quality: refactoring, pre/post-deploy checks, and security scanning

    1. CRCarly Ryan10:58

       Yeah. Really cool. Yeah, I think one of the like early differentiators I saw from you guys is kind of that third point, building durable, reliable apps that can be built on later. It's for actual like business owners and not necessarily just people prototyping. And so I remember in one of our early calls you were talking about you have your agent that initially builds the prototype or, or the product, but then you later have another agent that comes back-

    2. MJMukund Jha11:21

       Right

    3. CRCarly Ryan11:21

       ... and cleans up the code and makes it such that people can build on it later.

    4. MJMukund Jha11:24

       I mean, we, we take like code quality very seriously. So we have like refactoring agent that'll come and refactor your, your app. We have a post-deployment agent and a pre-deployment agent.They'll check for all security flaws. They'll sort of make sure that there are no key leakages-

    5. CRCarly Ryan11:36

       Yeah

    6. MJMukund Jha11:36

       ... bunch of those things. And sort of approaching the problem from, you know, the outcome first perspective, right? Like, allows us to sort of really design the system for that. What do you think our most defensible moat... Like, you know, what, what are you seeing you guys?

    7. CRCarly Ryan11:47

       Yeah. This has always been true for startups. It's like deeply understanding your users and deeply-

    8. MJMukund Jha11:52

       Right

    9. CRCarly Ryan11:52

       ... like building a product that you can iterate on quickly that really captures your users' needs and problems. And then kind of related to that is like building a brand and a go-to-market strategy and a user experience like around those people that you understand so well. The second thing is like moats that intelligence alone can't overcome.

    10. MJMukund Jha12:10

        Right.

    11. CRCarly Ryan12:10

        And so that's like proprietary data or infrastructure, building in a highly regulated space. That's like something that like AI is not gonna automate, you know, compliance regulation fully soon. The final one is like human trust.

    12. MJMukund Jha12:22

        Right. Mm.

    13. CRCarly Ryan12:22

        And so like this is also in regulated spaces, but anywhere, like human trust is not something that's gonna be automated soon. So kind of these moats that model intelligence can't bridge that gap yet. And then the final thing, which I, I see you guys doing, it's the most exciting for me to see a customer do, is build a beautiful product right now, but then also build a product that's looking towards the future. We are seeing models consistently get better at longer horizon tasks-

    14. MJMukund Jha12:45

        Totally

    15. CRCarly Ryan12:45

        ... more autonomous tasks. And you know, a year and a half ago, someone probably didn't think Emergent's product could exist. You know, text-to-app builder, truly democratizing software, and then continuing to push that forward of like, okay, today we know a product like Emergent can exist. What's the next thing ahead of the frontier for the models?

    16. MJMukund Jha13:03

        Right. I mean, internally, we have actually rewritten our systems like four times over nine months.

    17. CRCarly Ryan13:07

        Yeah.

    18. MJMukund Jha13:07

        Right.

    19. CRCarly Ryan13:07

        Yeah.

    20. MJMukund Jha13:08

        And we almost feel that with every new model launch, we have to sort of delete everything that we've learned so far-

    21. CRCarly Ryan13:12

        [laughs]

    22. MJMukund Jha13:12

        ... and reimagine the world. For example-

    23. CRCarly Ryan13:13

        And I work with you on it. [laughs]

    24. MJMukund Jha13:15

        Yeah, I mean, for example, Opus like really feels like a different class of model-

    25. CRCarly Ryan13:18

        Yeah

    26. MJMukund Jha13:18

        ... and much more capable of long horizon tasks. Especially like now we can have multiple agents coordinating on the same task.

    27. CRCarly Ryan13:24

        Totally.

    28. MJMukund Jha13:24

        My belief has always been that, okay, every time like a new class model shows up, let's delete everything we have done. Let's reimagine the world from this lens of the new model. So I think, I think that's been really, really helpful.

12. 13:32 – 15:26

    Who builds with Emergent: domain experts and a standout user story (EquiMind)

    1. CRCarly Ryan13:32

       I'm curious to transition into like your users. So we've talked about how most of them have never written code. Who are these people, and like what are they building? I'd love to hear some like user stories too.

    2. MJMukund Jha13:42

       Yeah, I mean, when we started, like, you know, like I... We thought that we were building, um, for mostly like semi-technical users, right?

    3. CRCarly Ryan13:48

       Mm.

    4. MJMukund Jha13:48

       It's gonna be mostly program managers, you know, product managers, designers who are gonna use us initially. But we were surprised to see when we launched that, like a lot of our users were actually business operators who were using us to build serious applications. And essentially we're actually built for like domain experts, people who really, really understand their domain really well and but never had the tools to express themselves, can now just come in and describe the problem in depth, and like agents will go and build the application for them. The gap between communication has been reduced, and build out and, and have this tight feedback loop to sort of, you know, iterate on the software themselves. And we have users across the globe, 190 countries, almost seven million users. My favorite example is, uh, this user Christie. She's in Alaska, and she is a clinical psychologist and also an equestrian coach, and she had been waiting for 10 years to sort of marry these two fields. And she had gone to, you know, a dev shop in Nova Scotia and got a quote of $15,000 to build this app, and tried it with them, but it didn't sort of work out. Discovered Emergent, started building on it, has her app live on the App Store today. It's called EquiMind. And for the first time she was able to marry these two fields of psychology, and her sports coaching has hundreds of users using it right now. And she was telling me that like, you know, like, "I've been waiting for 10 years for this opportunity to sort of really build something." So we are building for like small businesses today, right? Small businesses globally account for 70% of employment, accounts for like 50% of GDP of the world, right? But they've never had the tools to sort of really express themselves because, like there are a million niches in s- small businesses, right? And traditionally when you had SaaS world, they were forced to move upmarket for economics reason. But for the first time with AI, actually you can actually serve all of these million niches at zero marginal cost.

    5. CRCarly Ryan15:21

       Mm.

    6. MJMukund Jha15:21

       And that's what we are trying to enable, like, you know, really empower these small businesses to sort of accelerate their businesses.

13. 15:26 – 16:35

    What’s next: longer-horizon autonomy and ‘Wingman’ to automate business operations

    1. CRCarly Ryan15:26

       What do you think Emergent looks like a year from now? What are you most excited to build next?

    2. MJMukund Jha15:31

       Yeah, I think a year is a long time.

    3. CRCarly Ryan15:33

       Yeah. [laughs] In this world it definitely is.

    4. MJMukund Jha15:34

       Yeah, yeah. I think all of us have seen that meter chart where, you know, the long horizon agents are gonna be... For us, the market segment is like really clear.

    5. CRCarly Ryan15:41

       Mm.

    6. MJMukund Jha15:41

       We want to serve these small, medium businesses. We want to serve new entrepreneurs. So we are very, very focused on making them successful.

    7. CRCarly Ryan15:46

       Yeah.

    8. MJMukund Jha15:47

       We are coming out with a new agent which we think is significantly better than existing ones. One of the other things that we are really, really excited about is with especially long horizon agents coming online, we especially think that there's a way for us to automate all of the businesses' operations-

    9. CRCarly Ryan16:00

       Yeah

    10. MJMukund Jha16:00

        ... for small businesses. We want to move from software to sort of actually automating their entire business and allow for these autonomous businesses to happen on the platform. So we are very soon launching a new product. We're very excited about it. It's called Wingman. It's kind of agents for businesses.

    11. CRCarly Ryan16:14

        Cool.

    12. MJMukund Jha16:14

        And it's gonna automate all of your business processes, including finances, including operations, sales, marketing. 2026 is gonna be a year where people start automating large part of their businesses, and Emergent wants to be the platform for businesses to come and automate all of their businesses on, on top of us.

    13. CRCarly Ryan16:28

        Well, thank you for being here today, Mukund. It's been such a pleasure having you.

    14. MJMukund Jha16:31

        Thanks for having me. It's been a great chat. Thank you so much. [outro music]

Episode duration: 16:38