Transcription
All right, everybody, welcome back to Startup Basics. This weekinstartups.com/Basics is the URL where you can find these videos and the series that we've been doing for over 5 years now. What do we do in Startup Basics? We look at things that Founders, you know, need to get right and that they might make a mistake on sometimes, or they might not know about an opportunity. That's why we call it the basics. We do legal with our friends over at Wilson Cini, we do accounting with our friends over at Cruz, and we're really excited because one of the basics you have to get right today is AI.
Every startup I see, whether they're an AI startup or not, they're using AI to run their companies. And one of the topics we talk about here on This Week in Startups all the time is static team size. A lot of folks are sticking with five or 10 or 100 people at their startup, and instead of hiring people just running up the head count, they're deciding, hey, maybe I could automate stuff, maybe I can use AI to figure things out. And so it is now one of the best practices that you got to get right, just like you got to get your legal, just like you got to get your accounting right, you got to get your AI right inside your startups.
Don't I know it? When people email us their decks and they apply for funding from us, you know what we do? Zip, zip, zip. People didn't know this—I don't want to say because I'm gatekeeping here—one of the things we do is we send all that information to Gemini. I don't know if you know Google's really amazing large language model in service, and then Gemini spits out this great report based on our criteria and this analysis of the startup, so we can just read that short summary based on all the information we have. And you know what? I have some researchers and analysts who write short summaries, and I get the Gemini one. I think the Gemini one's a little better—to be totally honest—than the humans. Then the humans can go do more important work. So there's like an example for you of how, you know, this stuff is impacting everything, and we're so excited we have a partnership with Google Cloud for this series. They just published an amazing report; it's titled The Future of AI: Perspectives for Startups. Hey, that is really on-brand and on-target for us.
So what are we going to do here on this series? In other series, we would have one guest and maybe we do four or five topics. Here we have so many great opportunities to have a rotation of guests on who are building important tools in AI and, uh, you know, who our friends at Google may or may not have partnerships with. And today we're very lucky to have Harrison Chase, the CEO of LangChain, on the program. They provide a framework for building and designing an LLM of your own, which sounds like something, Harrison, I need for maybe startups and pitch decks. Welcome to the program, Harrison.
Thanks for having me. It's great to be here. Tell me a little bit about LangChain, and then let's get right into it. What do you think startup founders should be thinking about when it comes to using AI inside their companies to get an advantage, to save money, and to create better products?
Yeah, the types of applications that we see people building, they're starting to be ones that do the work of what humans would do in the past. So if they're kind of like functions inside a company that you would hire—what I like to say a smart intern to do—um, those are now functions that can kind of be automated by some of these AI systems. So, for example, within LangChain, we have a few places we use this. I have an email assistant that helps respond to all my emails; we have a customer support bot that helps with some of the customer support issues; we have a marketing bot; and, uh, we have an SDR bot. And so all these are places where we'd hire maybe, like, in the past an entry-level intern and entry-level person. And, and you know, because we build tools, we like to dog food them, and so we're dog fooding our own tools by trying to automate some of these processes away.
The interesting thing about this dog fooding you're doing is the positions you talk about are not positions people want to stay in for a career; they're entry-level, they're the first rung of a career ladder. And you know what? There used to be—when I was coming up, I'm a Gen X—you're like a—I think you're Gen Z or Millennial—you're Gen Z, I think, right? Millennial. Millennial. You're Millennial. Okay, great. You're Team Millennial. No judgments there. You know, Gen X, we're like the last free-range generation; we're a little crazy on the margins. But when I was coming up, the reception desk, working in the mail room, working in the typing pool, working as a runner—which is basically somebody who would run packages of paper around—those were the entry-level jobs. You know what happened to those jobs? AI, email, the internet. You didn't need a receptionist; you put technology in the front, people badged in, they pressed a number, and whatever, somebody came and got them. You didn't need all of humans doing those, and now we have another series of them that are entry-level jobs. SDR is a super fascinating one—Sales Development Rep. For folks who don't know, they find leads, they get those leads, they warm them up perhaps, and then they hand them off to an account executive, a salesperson.
Uh, in plain English, tell us a little about about the agent that you created for the SDR role; what do they do, and um, how well does it work, and how long have you been deploying your SDR agent? That's probably one of the newer ones. Um, basically what it does is we get a lot of inbound leads; it does some research on who the people are, um, and it actually drafts an email to them, um, if if it thinks they're interesting. So does it's, you know, it uses the reasoning of the models to determine whether it's kind of like an interesting prospect for us, um, it does some research on events that have happened to their company recently, and then it will draft an email, and, and uh, notably for for all of these positions, you're absolutely right that they're kind of like entry-level positions, but I want to call out that we have a sales team; we have a head of customer support; we have a product marketer. It's not like we're eliminating these functions completely; it's rather like these are doing some of the parts of the job that people don't want to do. They're not the creative part; they're not the kind of like the value-add part, and then they're hooking in; they're they're communicating with the kind of like the experts when needed. So when it drafts an email, we have a human in the loop that will go in and kind of like approve the email or something like that. So these these like, you know, we have we have a really good sales team; I think they can go in and basically talk to this junior intern and say, no, this is the wrong email; like, don't send it to these types of people in the future. So there's still this human in the loop component.
I think that's really important for enabling a lot of these applications. I think this is critically important at this stage in 2025 when we're recording this, because we do see on the margins a hallucination here or there, and um, you know, you don't want to have a hallucinated mistake in an outbound email to a prospective customer, nor do you want it to make a mistake and say this person doesn't need the product; we're not going to email them. So I like this, you know, taking those emails that are outbound, maybe putting them in the, you know, in your drafts box, you take a look at those 10, you just read them, okay, maybe we shouldn't talk about—I don't know—it pulls their high school or something and mentions their high school in the email, and that's like the super important part. Human in the loop, reinforcement learning is a very important piece of this as well, because over time, you know, these things could take on more and more work. Maybe, you know, you look up, hey, this person's company has 10 employees; this other person has 10,000; maybe the one with 10,000 we should just book a Zoom; the person with 10 people, hey, maybe it's okay to send that one automated. You know, it could depend on what you're doing there.
How hard is it to create these agents, and then are there situations where these agents have, you know, gotten a little bit out of control? Maybe they jumped the fence. How do you protect against that? Because that's everybody's concern, right? They may not say it to you, but people are like, oh my God, I don't want an agent to go wild, just like back in the day we wouldn't want somebody to spam, you know, and send a hundred accidental emails. Could be embarrassing, could be annoying to our partners and customers.
Well, that's exactly why the human-in-the-loop stuff is so important, and I'll I'll get to that after I answer your first question. I mean, we still see that it's still it's still pretty hard to create these agents. So we build developer tooling to help people build these agents. We see that most of these agents are still being created by developers. There's a lot of integrations to figure out; there's a lot of uh, what we call kind of like the cognitive architecture of the agent, like what information is it looking at, how is it processing that information? It's still a lot of work to get these agents to work, and the ones that we see work, um, some of the ones that have been built with our tools—Replit, LinkedIn, Uber, Clara, GitLab—these are like vertical agents; they're not like fully autonomous ones; they're vertical ones doing kind of like, you know, specific domain tasks. And then for the for the question around how do you keep these on the rails, this is this is why the human-in-the-loop stuff is super important as well. And I think there's two there's two big benefits to human in the loop. One is what you talked about, like it keeps them in check; it basically doesn't let them go off the rails. You have people not at every step. Like I think part of the benefit of having agents running in the background is you don't have to be involved at every step; you can be involved at the most important step. So, for example, like you can be involved right before an email is sent, because that's more important than before a Google search is done, like, you know, it's it's kind of like a read versus write operation. So it's more, you put them in at kind of like the crucial steps where it actually could do things that would not be good. But the second underrated part of human-in-the-loop is what you were talking a little bit about earlier is basically aligning the agents with what you want them to do. So when they first start working, there's probably some prompt, um, and that prompt is, you know, like I I have like I I think I'm relatively good at prompting; I wrote the prompt for my email assistant; I still forgot a ton of edge cases about who I would want to respond to or what emails I would want to ignore, just like didn't come to mind as I was writing that prompt. And I don't I don't think it's realistic to ever write like a perfect prompt right off the go. And so this human-in-the-loop helps you kind of like, if you set up the proper kind of like systems, it helps you update that prompt and update instructions and basically align these agents with what you actually want them to do. And so I think there's two really important benefits to human in the loop.
Let's talk about where this will be next year. So we're referring to AI as interns, and we probably referred to AI three years ago, you know, as if you're—in Gmail, uh, you know, guess the next word, and then it was like guess the next two words, you know, we were kind of in that nascent phase. If you said, hey, I'd love to invite you to, then it said to lunch, and then it said to lunch to discuss, and whatever, you get the idea. And now here we are saying, hey, read my email and draft something, put it there. Where would we be next year, uh, and then the year after? So let's talk about 2026, 2027. If these agents do a good job in 2025, hey, they go from being interns, maybe, you know, they get the full-time job, entry-level job, maybe to the next job. And, of course, we're giving this a caveat of this is the exoskeleton. If you think about this like an Iron Man suit, you still need to have humans at your company, but they're going to be able to take the grunt work, have AI do it, or do 80% of it; you're going to get that those superpowers, as it were, right? So maybe talk about what your predictions are for 26 and 27.
Yeah, I'd say, um, within a year we'll probably still have still have interns; they'll just be smarter. I think the models will get better; I think we'll get a little bit better at hooking them up to stuff, but I think they'll still be kind of like smarter interns. After that, I think there's like two two kind of like steps that will happen: one is this like memory component, so interacting with these agents and having them learn from from your feedback, um, I think that'll be really important for aligning them, because it doesn't matter how smart the intern is if it doesn't know how you like to do things at your company. Like you have—if you can write down a standard operating procedure for the role, that's fantastic, um, and we we don't have that for all roles, and I don't think it's realistic to ask that, but people do pick up those those processes that they should be following through memory, um, that's what we do as humans. So I think that'll be something that we start to work on probably towards 2027. And then I think the other thing will be—right now these interns are pretty independent; they just work by themselves. So the agents I talked about, like Replit has its agent that's pretty separate from Clara's customer support agent. What happens when these start being able to talk to each other and hand off things? And so multi-agent systems are probably something that will also pop up in like 2027. Multi-agent. So you got the SDR, you know, processing the inbound leads, drafting the emails, and then you're going to have a CRM agent cleaning up the database over there and saying, hey, we just updated everything over here about our customer, and let's say the customer was I don't know, McDonald's or Starbucks, and it's like, oh, if you see anything from Starbucks or McDonald's on the inbound, please take the account executive listed in our Salesforce, HubSpot, whatever, and um, check with them first or CC them or put it in their outbox. Wow, that's kind of dope when you start thinking about how these things might work together, uh, it could become really interesting.
When will they be sort of working next to you? I've always envisioned like these things having a bit of a persona; maybe we give it a name. Hey, this this is JCal, my SDR, and uh, you know, this is Harrison or this is Chase, my, you know, CRM manager, and they keep the database up to date. It'd be kind of cool if they were like in the Slack or they were in your Teams or sitting in a little window here while we're on this Zoom call and maybe listening in, contributing on the margins. Hey, you know, I was on the sales standup; we heard you talking about Starbucks, and so we wrote a little update on the latest news from Starbucks; there's a new CEO; here's what's going on there. So we just took the liberty of writing a dossier to uh, educate everybody, and then we did a quiz where we quizzed all the sales team who are associated and the customer support people on the history of Starbucks so they know they have a little bit of small-talking banter they can do.
Why aren't they hanging out with us yet, and when will they hang and be like peers in these spaces? So we call our customer support bot Carl, and Carl hangs out in our Slack.
Um, in the Slack now? Carl's in the Slack? Yeah, he's not sending dank memes, right? You talked to him about the dank memes; do not send—don't bring up politics at work; tell him we're focused. I don't—that's when you know we hit the singularity; Carl starts sharing memes.
Carl's the only one that's in the Slack, so so there's four—Carl's the only one that's in the Slack. Why is that the case? I I think like the big thing, um, or a big thing to figure out is like what these human-agent interaction patterns look like, um, and I think we have some idea, and I think the idea of treating them as like a coworker in in Slack or Teams or something like that makes a lot of sense, but it's still really early, and so I think I I think one of the best spaces that companies can be spending time is thinking about what does this like human-agent collaboration pattern look like. If you look at like a lot of the companies that have kind of taken off, I mean, like ChatGPT, ChatGPT changed the UX that we used to interact with LLMs, uh, you know, turned it into a chatbot—doesn't seem like a big thing now, but like that was a—in the UX. I think Cursor for coding has done a fantastic job at nailing the UX for developers in the IDE, or Google Search has the snippet up top, and I have to say exactly changed my behavior really, because now I get—my behavior was, you know, bifurcating. Okay, I I want to talk and do a chat interface on an LLM sometimes, and other times I kind of like the presentation of, let's say Google Flights or Google Local or Shopping, like there's like a lot of like intricate uh things that Google provides—Maps, uh, etc., images. And now you kind of have both, and so that's become super powerful. Sometimes they go to do a search, and the snippet up top or whatever they call that—it used to be called the one-box snippet; I don't know what they call the little chat window up there—but boy is that helpful because you get both. And I was wondering when they would do that because that would take a lot of servers, but yeah, I do believe the UX is going to be quite interesting.
Final question for you: Used to have to hire a developer to do anything, and maybe a script kiddie on the margins or whatever. Now I'm seeing a lot of people using, you know, pick a platform—Notion, Coda, Slack—and then they use something like Zapier or If This Then That, kind of glue some workflow together, and I think some of those other products are starting to add a little bit of workflow here on the margins, you know, simple stuff. But when will a non-developer be able to do the coding for agents? Because we are seeing, you know, in the startup community, I've had three or four startups come to me with no developer, and they built MVPs, and I'm like, well, that's pretty impressive. So can they do—do you do you have that on your roadmap? English language agent creation, is it on your roadmap at LangChain?
I think um, it's it's not super close on our roadmap. The agents that we see being built that are the most like um, intern-like, they're all built by pretty strong kind of like developer teams. So so Replit has a very strong developer team; GitLab does this well; Clara does this well. And I think the reason for this is is a fewfold. One, I think the best practices for building these agents, it's still super early on. Like LLMs have really only been a thing for—in the public's mind—for about two years, um, and agents for maybe like a year. And so we're still figuring out what the best practices are, and so there's a lot of control that you want to be able to have. And then another big part is giving these uh systems access to all the uh, you know, other systems that exist within a company, and this is very heavy on integrations, and and that's a place where there's a lot of need for coding and data engineering at the moment. So at the moment, to be honest, I'm a little bit skeptical that we'll see that anytime soon. Most of the most of the most impressive agents we see are being built by strong developer teams still. Small price to pay; good use of developer hours to make an agent that then, you know, takes out—I don't know if it's like 2 hours a day, and you're working 50 weeks a year times 5, you know, you're talking about 500 hours. One of the nice things, too, is these things can be working 24/7, uh, that's why they're agents, and they're running in the background. So I think it's like uh a super fascinating concept, uh, so well done.
Where can people find out more about your company if they want to use your solution? You can find us at LangChain.com or on Twitter or LinkedIn. Awesome, everybody go check out LangChain and uh see if that's the tool right for you if you want to save uh, you know, a couple thousand hours of work every year in uh the intern jobs and not have interns doing grunt work; have them do something more interesting in your company. All right, thank you to Harrison Chase for joining us here on the AI basic series on This Week in Startups. You can see all the This Week in Startups basic series at thisweekinstartups.com/Basics. It's a long URL, I know. And if you want to check out Google Cloud's awesome The Future of AI: Perspectives for Startups report, go to go.gle/futureofAI. That'll be in the show notes as well, for predictions, real-world examples, and tons of startup advice. Once again, the URL—you can write it down right now—go.gle/futureofAI. No spaces and dashes in future of AI. Discover what uh all these AI leaders have to say about the future of AI and its impact on your business. Thanks again for listening, and we will see you next time on This Week in Startups. Bye-bye.