Hey everyone. Super excited to be here and talking to my friend and founder and CEO of LangChain, Harrison. Before we do that, for those that are less familiar, I'm Rajee, an investment partner with Sapphire Ventures. And our mission statement at Sapphire Ventures is to partner with enterprise AI founders that are building companies of consequence. Those are companies such as LangChain, Glean, Clay, Sierra, Temporal, WorkOS, and a host of others that we've been partnering with over the last 10-15 years. So now let's get on to Harrison. You know, the world has obviously shifted a lot yet again. I would say probably from mid-December with Opus 4.6, now there's obviously 4.7 and OpenClaw and now long-running agents, and there's no one that's been more at the center of it than Harrison and LangChain. He'll tell you a little bit about what they do, but effectively helping developers everywhere, frankly, from Fortune 500 companies like Cisco and Workday to AI-native startups like Harvey and Clay and Replit, helping these developers build and monitor and observe and evaluate agents. And the more complex the better, and his platform can kind of do it all. So I'm excited to welcome him, Harrison. Now, it's a $1.25 billion valuation company and leading the agent revolution here. So, welcome Harrison.
I'm excited to be here. It's going to be a fun conversation.
Of course. Yes. So, take us back to the beginning. Now, I think you're very well known amongst the community at large and being a leading voice in AI. But you started the company in 2022. What did you see maybe in your prior experiences or then that led you to start it? And then give us the quick snapshot on how the LangChain story itself has evolved.
Yeah, so my background is in ML and MLOps, and I worked at a number of startups before this. Did some early stuff with early BERT models in like 2017, 2018. Then went and did some other things not related to NLP. I was at my previous startup and knew I was going to leave, didn't know what I was going to do. Was going out talking to a lot of folks in the startup community at events like this and saw some people doing really cool things with early versions of these LLMs. So the Da Vinci models from OpenAI I think were the only main usable ones at the time, but I thought those were so cool because they were general purpose and you could just change what they did by prompting them, and they were so much better than the BERT models I'd used earlier. So basically saw some patterns in terms of what people were working on, put them into an open source package and released that in October of 2022, so a month before ChatGPT. Fantastic timing. And you know, ChatGPT comes out, everyone wants to build that but for their data and their APIs. And that was a lot of what LangChain helps with. It was patterns for connecting these LLMs to other sources of data and computation. So kept on building it, kept on tweeting about it, kept on talking to users about it. Started a company around it in early 2023. In summer of 2023, we released LangSmith, which is our platform for observability and evals. The core insight there is basically it's pretty easy to build a prototype that you can share on Twitter. It's hard to get that reliable enough to be something that can run in production. And so we released that in summer 2023. Since then, we've worked on a number of other open source frameworks because the space just moves so fast. So LangGraph and Deep Agents are two of the newer frameworks we worked on there. And we've added more things into our platform. So we've added in deployments, we've added in sandboxes, and we're continuing to add things into this like really what we think of as like the whole agent development life cycle.
Amazing. Yeah. You know, this reminds me, given your guys's evolution, at least for an investor, the formula for a great founder CEO 10 years ago or just generally through the cloud revolution was you back someone who has this very discrete, concrete sort of vision and they have a very defined sequence to get to that vision. It's steadfast. It never changes. A pivot was almost looked down upon. And now I feel like I'm meeting these founders and I'm sort of evaluating them on how hard they are pivoting and moving every three months to keep up. It's like it's a whole different sort of thing. So talk to us about some of those decisions that you've had to make at LangChain, but moreover, what advice would you give some of the younger founders and folks in terms of how to almost lead your company through these pivots? Because it's how do you get everyone sort of marching to that same beat if that beat may change every three months?
Yeah, I think for us, we've always had kind of like the same high-level vision statement. So I went back and looked at our pitch deck from the seed round and the vision was to make intelligent agents ubiquitous. Like that was the goal. And when we think about a lot of the tools that we're building, it's still towards that goal. It's how can we make like as many agents out there that are doing work and being intelligent as possible. What do the agents of the future look like and what are the tools that are needed to build them? The thing that changes is like what the agents of the future look like. No one knows what that means. Not even today. And I think we've changed our belief on what the agents of the future look like over the past three years and we'll probably continue to change it in some ways. Some ways are similar. So I think like when we think about there's two big parts of what we do. You think about like the open source frameworks and then LangSmith, which is again observability, eval. Open source changes very, very quickly. So we launched, we did LangChain and then LangGraph and then LangChain 1.0 and then Deep Agents, and all of these are like fundamentally different ways to build and architect agents. On LangSmith, observability and eval, and so that's actually been more stable. And so the only thing that I'd add, or my main answer to you, like I think like adaptability is like the most important thing. I think it's great to have a, we do like, I think it's great to pivot. I think it's also great to have like a consistent north star. I think even if you have that north star, you're going to have to be incredibly adaptable as the models change, as the ecosystem changes. Even if that north star is the same, the way to get there, I think, changes a bunch right now because the space is just so fast moving.
I guess do you also then adjust your hiring? The types of folks that you want to hire, you want to make sure they can be also agile and adaptable as well. And you're sort of quizzing for that.
Yeah. I mean, I think there's something really interesting happening with coding agents right now where like in order to build kind of like software, it's a different type of building than before. I think like generalists or builders or people who kind of combine product sense with some engineering knowledge with some design taste as well. Those types of generalists are super impactful in terms of being able to just drive projects with a really, really small team around them. Like previously you might need to have a PM and have three engineers and a designer and that would be like the smallest team you can have. Now it's maybe like two people, like one really product-minded engineer and then one maybe like architect to help scale that and bring that into kind of like the production system. And so I think that type of profile, like I don't really know what to call it besides like a builder. That's super valuable in this day and age when you need to be able to, yeah, adapt and change directions.
I think that's well said. There was actually, I don't know if I sent it to you and Ankush, but an article Brett Taylor for the audience wrote recently on the AI-native interview. Did I send that to you as well?
You didn't send that to me, but I read that.
Separate to, yeah. So, you just, everyone has to build in an interview and explain why and show some design taste and sort of this intuitive knowledge of what they're doing. Anyways, let's go back to the, at least what I view, maybe I'm just in my own investor world, as sort of the elephant or two elephants in the room with OpenAI and Anthropic. There's so much development there all the time. You've written a lot and talked a lot about how owning the memory and context layer and all the scaffolding and harness work you're doing, like every enterprise should own that and not give that up to the foundation models. Can you unpack that? What exactly does that mean and what do you do there?
Yeah. So I think like it's actually quite easy to kind of like change models these days. There's some prompting that you have to do to go from one family of models to another family of models, but by and large it's pretty easy to change models. They all have a similar sort of API. They're all largely kind of like stateless. And so to go from one model to another is actually pretty easy. I think what creates more lock-in is when there's state that starts to accumulate behind these APIs. And so that could take the form of kind of like short-term memory. So like what happens inside of a conversation. So conversational history, which we see some of the model labs starting to do. And or it could take the form of long-term memory and learning kind of like preferences over time. I think that creates a lot more lock-in and I think that's something that everyone should be, especially enterprises, should be pretty wary of. It makes it a lot harder to switch. That, you know, I think again, adaptability is a key thing that you want in this day and age. A new model comes out that's better for coding or better for writing, you want to switch to that as quickly as possible. If some of your like memory or state is locked into a particular platform or provider, that prevents that kind of like adaptability and keeping up with the most recent model. Another way to kind of like frame it, I think long-term memory will be one of the big kind of like topics that we talk about as an industry this year or next year. Or both. I think when we think about building agents that actually work, there's a step one of getting kind of like the functionality correct. So like tools and basic instructions, but then I think there's a step two of personalizing it for the user or company that's kind of like building or interacting with that agent. So I remember I chatted with a CIO of a bank in New York, and they were talking about how they wanted to basically take all the knowledge that was in some of their traders' heads that you can't really like write down or capture anyway. And they wanted to look at basically all of their interactions and turn that into a form of memory that they could use if this trader was to leave and go to another shop down the street. And I think that's what a lot of people want. They want memory. Like a lot of people use agents to interact across multiple different tools or multiple different platforms, and that's replacing humans who used to do that. And those humans had some implicit knowledge of a job that required some intelligence and some decision-making that interacted across all of these. And I think memory is our attempt, the industry's attempt, to capture that implicit knowledge. And I think that's really, really valuable because that's what makes humans really good at a lot of these jobs is knowing when to use this tool and when to use this other tool and when to go deep and what to do in this scenario. And so I think memory has a lot of gravity. And I think as an enterprise or as a startup, you want full control over that memory so that you can switch it out and use it with different models or use it in different scenarios as much as possible. And so I think everyone kind of recognizes this at this point. I think if you look at a lot of the recent releases that a lot of the model labs are doing, I think they know that a lot of the models themselves are commodities and are getting commoditized. And I think a lot of what they're doing is trying to capture more state because they recognize that that's where a lot of the stickiness will come from.
