- Carly (Applied AI team at Anthropic) interviews Mukund, CEO and co-founder of Emergent, about how the company democratizes production-grade software building for small businesses via natural language.
- Emergent pivoted from automating software testing at YC to building general coding agents, owning the entire stack — its own container technology, multi-agent systems, and long-term memory — to deliver durable, deployable apps rather than just prototypes.
- The company runs on Claude (Sonnet, then Opus as its workhorse) because outcome and accuracy matter more than speed for its users, who compare Emergent to a $250K dev shop, not to an IDE.
How Emergent is making app building more accessible with Claude
- Emergent serves "serious builders" — small businesses, entrepreneurs, and domain experts who want real economic value and business-critical apps, not front-end demos, distinguishing it from prototype-focused competitors.
- The team chose Claude early (Sonnet 3.5 in 2024) for strong instruction following and coding/front-end ability, which let them imagine long-running agents; Opus became their workhorse because accuracy compounds over thousands of steps.
- Owning the whole stack — custom container technology on Kubernetes with memory snapshotting for parallel agents, plus hosting — generates proprietary data that feeds long-term memory, self-learning agents, and RL systems.
- "Long-term memory" lets agents learn not just within a user session but across all apps built on the platform, so a new error or library upgrade is handled in future iterations with fewer tokens and more accuracy.
- Tight feedback loops and reliability raised deployment success from 84% to nearly 98%, with agents now able to run for hours; errors that only surface in production are piped back to the dev environment for agents to learn from.
- Code generation is only ~20% of the problem — the other 80% is deployment, maintenance, security, scalability, and integrations (payments, auth, large files); refactoring, pre-deployment, and post-deployment agents enforce code quality and security.
- Emergent's defensible moats: deep customer understanding, agent/harness quality, proprietary data, and solving a single vertical tech stack; they plan to move beyond software to automating entire small-business operations with a new product called Wingman.
Emergent — A company building an AI platform that lets non-developers create production-grade software from natural language.
Dev shop — An outsourced software development agency; Emergent's users compare it to a ~$250K, multi-month engagement.
Multi-agent system — A system where multiple agents coordinate, communicate, and share state to complete tasks.
Verification loop — A feedback mechanism that checks an agent's work so it can run longer and more reliably.
Container technology — Custom infrastructure (here on Kubernetes with memory snapshotting) that runs and saves agent states in parallel.
Long-term memory — Cross-session, cross-app learning so agents improve over time using less tokens with more accuracy.
Instruction following — A model's ability to faithfully follow detailed instructions, critical when small errors compound over many steps.
Zero marginal cost — Serving each additional niche customer at essentially no extra cost, enabled by AI.
Wingman — Emergent's upcoming product: agents that automate business processes like finance, operations, sales, and marketing.
We are building for small businesses today. Small businesses globally account for 70% of employment and 50% of GDP of the world. But they've never had the tools to really express themselves because there are a million niches in small businesses. For the first time with AI, you can actually serve all of these million niches at zero marginal cost. And that's what we are trying to enable.
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.
Thank you for having me here. Super excited to be here. So I'm doing this startup with my twin brother, Madhav, and we actually started programming at age 12. We love sibling founders here. Totally, and we've 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 500-people engineering team. When we entered YC, our first insight was to automate software testing, because I had seen at Dunzo how testing was the bottleneck for shipping software fast. We started with how to automate all of the software testing, including mobile apps and web apps. When we started building that out, we realized that to automate testing, we had to build our own container technology and our own coding agents. We stumbled upon this insight that if you can solve for the verification loop, you can make agents run longer. On day one, we told our YC partners that testing is great, but we want to now build general coding agents. We were almost like an applied research lab, just building high-production-grade quality agents. We invented a bunch of things like test time compute — how do we solve scale, the memory across agents, agent-to-agent communication? We were one of the first teams to productionize multi-agent systems. When we launched, our first approach was to go to enterprise. We had an enterprise customer but realized the adoption of the enterprise is going to be slow. At the same time, we started using these internal coding agents to build everything internally. In fact, there were a few non-technical people on the team who were using it more than developers were. We truly realized the power is actually to democratize software engineering for everyone. That's when, in June, we launched almost like a research preview. At that time, agentic coding was not that popular. When we launched, it just took off, and we were surprised. Everybody else at that time was building for front-end-heavy applications and demo kinds of things. We had fundamentally approached the problem thinking, what would it take for anybody to ship production-ready software that will have real use cases and business-critical apps running? Our insight was to automate all of the testing, linting, deployment, security. How do you solve for code reviews? That really resonated with users because most of the users coming to Emergent today are serious builders — small businesses, entrepreneurs who want to see economic value from the software getting built.
That's awesome. So it's basically democratizing, taking natural language, for non-software engineers, but also building durable products — not just for prototyping. So you've been building on Claude since the early days. We met about a year ago. What made that the default choice, and what's kept you on Claude?
I think we were almost lucky to get started when Sonnet came out. In AI, everybody needs to imagine the world six months ahead and build for that versus building for what's available today. So people who started with the previous generation of models would be solving a different problem versus us, because for the first time, Sonnet had really good instruction following so you could imagine what long-running agents would look like. It was really good at programming and front-end as well. That gave us a futuristic view of what agentic programming could look like. Our users are comparing us against a dev shop. They're not comparing us against an IDE or a tool. To them, outcome matters a lot. We focused on what would give us the highest quality and the highest probability of getting to the right outcome, versus speed. Most other players focused on front-end or design care a lot about speed. But for us and our users, what's important is, are they able to reach the outcome they want? That's why Opus has been our workhorse — it's really great at instruction following. If you have the right tight feedback loops that we've built using our multi-agent systems, you can keep them running much longer. Today, our agents can run for hours trying to build software. We run very tight evals internally, and you've been really helpful in helping us. I remember when we were about to launch, we had this 2:00 AM meeting just to get started. A lot of your engineers are in India, so it was Friday 2:00 PM for us and the middle of the night for you. I expected they'd answer the next day, but one of your engineers said, "Yeah, let's get on a call in 15."
Your team during YC built this internal coding agent. What did you learn from that? What shows up today, and what did you scrap? Our first insight was that to build an automated software engineering platform, you have to own every layer of the stack. You have to co-build the infrastructure to support your agent and the agent to support the infrastructure. At that time, everybody took the shortcut of using a third party. But there were no good container technologies, so we had to invent and build our own on top of Kubernetes, which does memory snapshotting so you can save states and run parallel agents. For us, the agent is the product in many ways, and the harness quality really matters. We built this multi-agent system, one of the first teams to productionize it. Many learnings — how to save memory, agent-to-agent communication, better context management, effectively using caching to keep costs down, and tightening all the feedback loops. The other thing important from day one was owning the entire stack, because the last mile is where most solutions trip. We spend a lot of time making sure production quality is really high — what you see in development replicates in production. Today, we're probably one of the deepest logs in the app-building space, and that lets us compound our learning. A lot of errors don't show up on the dev side; they only show when your app is live and real users are using it. We pipe all of that back to our development environment and let our agents learn. We also built what we internally call long-term memory, where an agent learns not just within the user session but across all the apps being built. The first time it sees a new error or a new library upgrade, it learns that and does it in the next iteration in fewer tokens with more accuracy.
I guess because you own the whole stack, up to the hosting, that gives you proprietary data to feed into the next parts of the agent. How do you think about model selection, especially when your users can't debug the output? A lot of the burden lies on us to make our users successful, and we take that very seriously. A couple of months back, our deployment rates were 84%, now it's closer to 98%. So 98% of people on the platform were able to build something and deploy to production. That's just tightening all these feedback loops and making agents extremely reliable and self-learning. In terms of the model, a user comparing us to a dev shop is comparing us to a $250,000 price point they would have paid. They're not price sensitive, and not latency sensitive, because they're comparing us to a three-month project. We're very obsessed with the quality of the product and not so much about cost. Being on the frontier, the highest reasoning possible, is something we index on a lot. We're not trying to do a one-shot solution; we're trying to build reliable production systems. Instruction following should be really good because errors compound — even small errors over 10,000 steps compound very quickly. Because the outcome matters most, we rely on the best possible models for reasoning and coding.
The AI builder space has gotten crowded. What has Emergent built that's genuinely hard to replicate? First, we're at the beginning of a very large market. The market is growing faster than all of us can serve. So we're not so worried about what competition or labs are doing. On the technical side, our approach has always been to build closely with customers. We understand our customer really well — they don't care about demos or front-end prototypes. They care about real production use cases, third-party integrations working, payments getting through, authentication working. A lot of our users early on started uploading really large files, so we had to think about supporting large file systems on production. They care about scalability, so when we write code, we make sure it's horizontally scalable. So the technical moat is, one, our agent quality and harness quality, which is continuously evolving. Second is all the proprietary data we're collecting, which feeds into our long-term memory, self-learning agents, and RL systems. Because we solve only for one stack, we're almost a vertical agent for coding — our users leave the technical choices to us. That lets us accelerate learnings on that tech stack. Lastly, code generation is only 20% of the problem. The 80% is how you take it to deployment, maintain it in production, make sure security is high. We have a refactoring agent that refactors your app, a post-deployment agent, and a pre-deployment agent that check for security flaws and key leakages.
I'm curious to transition into your users. Most of them have never written code. Who are these people, and what are they building? When we started, we thought we were building for semi-technical users — program managers, product managers, designers. But we were surprised that a lot of our users were business operators building serious applications. We're built for domain experts — people who understand their domain really well but never had the tools to express themselves. They can now describe the problem in depth, and agents build the application for them. We have users across the globe, 190 countries, almost seven million users. My favorite example is Christie. She's in Alaska, a clinical psychologist and also an equestrian coach. She had been waiting for 10 years to marry these two fields. She had gone to a dev shop in Nova Scotia and got a quote for $1,000, but it didn't work out. She discovered Emergent, started building, and has her app live on the App Store today. It's called EquiMind. For the first time, she was able to marry psychology and her sports coaching, and it has hundreds of users right now. We're building for small businesses today. Small businesses globally account for 70% of employment and 50% of GDP, but they've never had the tools to express themselves because there are a million niches. Traditionally with SaaS, they were forced to move up market for economic reasons. But for the first time with AI, you can serve all these million niches at zero marginal cost.
What do you think Emergent looks like a year from now? What are you most excited to build next? A year is a long time in this world. The market segment is clear — we want to serve small and medium businesses and new entrepreneurs. We're coming out with a new agent that we think is infinitely better than existing ones. Especially with long-horizon agents coming online, we think there's a way to automate all of the operations for small businesses. We want to move from software to automating their entire business, allowing autonomous businesses to happen on the platform. We're very soon launching a new product called Wingman. It's agents for businesses, and it's going to automate all your business processes, including finances, operations, sales, and marketing. 2026 is going to be a year where people start automating a large part of their businesses, and Emergent wants to be the platform for that. Well, thank you for being here today, Mukund. It's been such a pleasure. Thanks for having me. It's been a great chat.
TL;DR
- Carly (Applied AI team at Anthropic) interviews Mukund, CEO and co-founder of Emergent, about how the company democratizes production-grade software building for small businesses via natural language.
- Emergent pivoted from automating software testing at YC to building general coding agents, owning the entire stack — its own container technology, multi-agent systems, and long-term memory — to deliver durable, deployable apps rather than just prototypes.
- The company runs on Claude (Sonnet, then Opus as its workhorse) because outcome and accuracy matter more than speed for its users, who compare Emergent to a $250K dev shop, not to an IDE.
Takeaways
- Emergent serves "serious builders" — small businesses, entrepreneurs, and domain experts who want real economic value and business-critical apps, not front-end demos, distinguishing it from prototype-focused competitors.
- The team chose Claude early (Sonnet 3.5 in 2024) for strong instruction following and coding/front-end ability, which let them imagine long-running agents; Opus became their workhorse because accuracy compounds over thousands of steps.
- Owning the whole stack — custom container technology on Kubernetes with memory snapshotting for parallel agents, plus hosting — generates proprietary data that feeds long-term memory, self-learning agents, and RL systems.
- "Long-term memory" lets agents learn not just within a user session but across all apps built on the platform, so a new error or library upgrade is handled in future iterations with fewer tokens and more accuracy.
- Tight feedback loops and reliability raised deployment success from 84% to nearly 98%, with agents now able to run for hours; errors that only surface in production are piped back to the dev environment for agents to learn from.
- Code generation is only ~20% of the problem — the other 80% is deployment, maintenance, security, scalability, and integrations (payments, auth, large files); refactoring, pre-deployment, and post-deployment agents enforce code quality and security.
- Emergent's defensible moats: deep customer understanding, agent/harness quality, proprietary data, and solving a single vertical tech stack; they plan to move beyond software to automating entire small-business operations with a new product called Wingman.
Vocabulary
Emergent — A company building an AI platform that lets non-developers create production-grade software from natural language.
Dev shop — An outsourced software development agency; Emergent's users compare it to a ~$250K, multi-month engagement.
Multi-agent system — A system where multiple agents coordinate, communicate, and share state to complete tasks.
Verification loop — A feedback mechanism that checks an agent's work so it can run longer and more reliably.
Container technology — Custom infrastructure (here on Kubernetes with memory snapshotting) that runs and saves agent states in parallel.
Long-term memory — Cross-session, cross-app learning so agents improve over time using less tokens with more accuracy.
Instruction following — A model's ability to faithfully follow detailed instructions, critical when small errors compound over many steps.
Zero marginal cost — Serving each additional niche customer at essentially no extra cost, enabled by AI.
Wingman — Emergent's upcoming product: agents that automate business processes like finance, operations, sales, and marketing.
Transcript
We are building for small businesses today. Small businesses globally account for 70% of employment and 50% of GDP of the world. But they've never had the tools to really express themselves because there are a million niches in small businesses. For the first time with AI, you can actually serve all of these million niches at zero marginal cost. And that's what we are trying to enable.
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.
Thank you for having me here. Super excited to be here. So I'm doing this startup with my twin brother, Madhav, and we actually started programming at age 12. We love sibling founders here. Totally, and we've 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 500-people engineering team. When we entered YC, our first insight was to automate software testing, because I had seen at Dunzo how testing was the bottleneck for shipping software fast. We started with how to automate all of the software testing, including mobile apps and web apps. When we started building that out, we realized that to automate testing, we had to build our own container technology and our own coding agents. We stumbled upon this insight that if you can solve for the verification loop, you can make agents run longer. On day one, we told our YC partners that testing is great, but we want to now build general coding agents. We were almost like an applied research lab, just building high-production-grade quality agents. We invented a bunch of things like test time compute — how do we solve scale, the memory across agents, agent-to-agent communication? We were one of the first teams to productionize multi-agent systems. When we launched, our first approach was to go to enterprise. We had an enterprise customer but realized the adoption of the enterprise is going to be slow. At the same time, we started using these internal coding agents to build everything internally. In fact, there were a few non-technical people on the team who were using it more than developers were. We truly realized the power is actually to democratize software engineering for everyone. That's when, in June, we launched almost like a research preview. At that time, agentic coding was not that popular. When we launched, it just took off, and we were surprised. Everybody else at that time was building for front-end-heavy applications and demo kinds of things. We had fundamentally approached the problem thinking, what would it take for anybody to ship production-ready software that will have real use cases and business-critical apps running? Our insight was to automate all of the testing, linting, deployment, security. How do you solve for code reviews? That really resonated with users because most of the users coming to Emergent today are serious builders — small businesses, entrepreneurs who want to see economic value from the software getting built.
That's awesome. So it's basically democratizing, taking natural language, for non-software engineers, but also building durable products — not just for prototyping. So you've been building on Claude since the early days. We met about a year ago. What made that the default choice, and what's kept you on Claude?
I think we were almost lucky to get started when Sonnet came out. In AI, everybody needs to imagine the world six months ahead and build for that versus building for what's available today. So people who started with the previous generation of models would be solving a different problem versus us, because for the first time, Sonnet had really good instruction following so you could imagine what long-running agents would look like. It was really good at programming and front-end as well. That gave us a futuristic view of what agentic programming could look like. Our users are comparing us against a dev shop. They're not comparing us against an IDE or a tool. To them, outcome matters a lot. We focused on what would give us the highest quality and the highest probability of getting to the right outcome, versus speed. Most other players focused on front-end or design care a lot about speed. But for us and our users, what's important is, are they able to reach the outcome they want? That's why Opus has been our workhorse — it's really great at instruction following. If you have the right tight feedback loops that we've built using our multi-agent systems, you can keep them running much longer. Today, our agents can run for hours trying to build software. We run very tight evals internally, and you've been really helpful in helping us. I remember when we were about to launch, we had this 2:00 AM meeting just to get started. A lot of your engineers are in India, so it was Friday 2:00 PM for us and the middle of the night for you. I expected they'd answer the next day, but one of your engineers said, "Yeah, let's get on a call in 15."
Your team during YC built this internal coding agent. What did you learn from that? What shows up today, and what did you scrap? Our first insight was that to build an automated software engineering platform, you have to own every layer of the stack. You have to co-build the infrastructure to support your agent and the agent to support the infrastructure. At that time, everybody took the shortcut of using a third party. But there were no good container technologies, so we had to invent and build our own on top of Kubernetes, which does memory snapshotting so you can save states and run parallel agents. For us, the agent is the product in many ways, and the harness quality really matters. We built this multi-agent system, one of the first teams to productionize it. Many learnings — how to save memory, agent-to-agent communication, better context management, effectively using caching to keep costs down, and tightening all the feedback loops. The other thing important from day one was owning the entire stack, because the last mile is where most solutions trip. We spend a lot of time making sure production quality is really high — what you see in development replicates in production. Today, we're probably one of the deepest logs in the app-building space, and that lets us compound our learning. A lot of errors don't show up on the dev side; they only show when your app is live and real users are using it. We pipe all of that back to our development environment and let our agents learn. We also built what we internally call long-term memory, where an agent learns not just within the user session but across all the apps being built. The first time it sees a new error or a new library upgrade, it learns that and does it in the next iteration in fewer tokens with more accuracy.
I guess because you own the whole stack, up to the hosting, that gives you proprietary data to feed into the next parts of the agent. How do you think about model selection, especially when your users can't debug the output? A lot of the burden lies on us to make our users successful, and we take that very seriously. A couple of months back, our deployment rates were 84%, now it's closer to 98%. So 98% of people on the platform were able to build something and deploy to production. That's just tightening all these feedback loops and making agents extremely reliable and self-learning. In terms of the model, a user comparing us to a dev shop is comparing us to a $250,000 price point they would have paid. They're not price sensitive, and not latency sensitive, because they're comparing us to a three-month project. We're very obsessed with the quality of the product and not so much about cost. Being on the frontier, the highest reasoning possible, is something we index on a lot. We're not trying to do a one-shot solution; we're trying to build reliable production systems. Instruction following should be really good because errors compound — even small errors over 10,000 steps compound very quickly. Because the outcome matters most, we rely on the best possible models for reasoning and coding.
The AI builder space has gotten crowded. What has Emergent built that's genuinely hard to replicate? First, we're at the beginning of a very large market. The market is growing faster than all of us can serve. So we're not so worried about what competition or labs are doing. On the technical side, our approach has always been to build closely with customers. We understand our customer really well — they don't care about demos or front-end prototypes. They care about real production use cases, third-party integrations working, payments getting through, authentication working. A lot of our users early on started uploading really large files, so we had to think about supporting large file systems on production. They care about scalability, so when we write code, we make sure it's horizontally scalable. So the technical moat is, one, our agent quality and harness quality, which is continuously evolving. Second is all the proprietary data we're collecting, which feeds into our long-term memory, self-learning agents, and RL systems. Because we solve only for one stack, we're almost a vertical agent for coding — our users leave the technical choices to us. That lets us accelerate learnings on that tech stack. Lastly, code generation is only 20% of the problem. The 80% is how you take it to deployment, maintain it in production, make sure security is high. We have a refactoring agent that refactors your app, a post-deployment agent, and a pre-deployment agent that check for security flaws and key leakages.
I'm curious to transition into your users. Most of them have never written code. Who are these people, and what are they building? When we started, we thought we were building for semi-technical users — program managers, product managers, designers. But we were surprised that a lot of our users were business operators building serious applications. We're built for domain experts — people who understand their domain really well but never had the tools to express themselves. They can now describe the problem in depth, and agents build the application for them. We have users across the globe, 190 countries, almost seven million users. My favorite example is Christie. She's in Alaska, a clinical psychologist and also an equestrian coach. She had been waiting for 10 years to marry these two fields. She had gone to a dev shop in Nova Scotia and got a quote for $1,000, but it didn't work out. She discovered Emergent, started building, and has her app live on the App Store today. It's called EquiMind. For the first time, she was able to marry psychology and her sports coaching, and it has hundreds of users right now. We're building for small businesses today. Small businesses globally account for 70% of employment and 50% of GDP, but they've never had the tools to express themselves because there are a million niches. Traditionally with SaaS, they were forced to move up market for economic reasons. But for the first time with AI, you can serve all these million niches at zero marginal cost.
What do you think Emergent looks like a year from now? What are you most excited to build next? A year is a long time in this world. The market segment is clear — we want to serve small and medium businesses and new entrepreneurs. We're coming out with a new agent that we think is infinitely better than existing ones. Especially with long-horizon agents coming online, we think there's a way to automate all of the operations for small businesses. We want to move from software to automating their entire business, allowing autonomous businesses to happen on the platform. We're very soon launching a new product called Wingman. It's agents for businesses, and it's going to automate all your business processes, including finances, operations, sales, and marketing. 2026 is going to be a year where people start automating a large part of their businesses, and Emergent wants to be the platform for that. Well, thank you for being here today, Mukund. It's been such a pleasure. Thanks for having me. It's been a great chat.