Langbase Founder & CEO Ahmad Awais and GitHub's Andrea Griffiths explore State of AI Agents report. The report features insights from 3400+ professionals on AI agent adoption, and challenges. Together, they explore the ecosystem of building and scaling AI agents in the real world.
Resources
Transcript
Andrea: Hello my friends, welcome! Good morning, good afternoon, good evening - wherever in the world you're joining from. Welcome to the special edition of Open Source Fridays, but not quite Friday - it's Open Source and AI!
It's my pleasure to be here with you today. I'm Andrea Griffiths, aka Columbia Dev, and all places. We have a super special guest today. We're going to get to talk about some really fun, exciting things that he's working on. We're going to go over the State of the Agent - AI Agents report, which is super relevant to all of you who are out there building things. And if you have not considered using agents in what you're building, you're going to learn why - a bit about the state of the industry.
But before we get to work, I want to make sure that you got to see a few of the announcements that we shared a moment ago. More importantly, my very, very serious announcement about our Valentine's Day little giveaway opportunity for you to get into Spark preview. And relevant since we're going to be talking about agents and Sparks - say that's agentically, right? It can create a whole app for you just by you inputting natural language. It's right now on a closed preview - you have to apply. And I will share the link to that little context later on today so that if you feel poetic and want to enter for a chance to get a preview slot, you go ahead and do that.
Again, my name is Andrea Griffiths, and we're here to talk about the AI Agent State of report by Langbase. My guest today is Ahmad Awais, who is an award-winning open source engineer, leader, Founder and CEO of Langbase. You can follow Ahmad on X - but I'm pretty sure you probably already do - or on GitHub.
Ahmad, thank you so much, welcome!
Awais: Yeah, thank you for having me here. Am I audible?
Andrea: I think you're muted. It's been a while since we've gone live, so systems are a bit wonky, so bear with me, friends. But welcome, thank you so much for being here.
Awais: Awesome, awesome, awesome to be here. Andrea: Where are you joining us from? Awais: San Francisco Bay Area. Andrea: Okay, nice, nice. And we met a bit ago - we were actually, we met in person a long time ago, but then we got a really good opportunity to chat last October, I believe? Yes, yes, yes. And I believe the report was not out yet at the time.
Awais: No, no, no. We were still, you know, gathering data at that time. You know, I think at that time, a lot of frontier startups and anybody, you know, who has a lot of data or scale was looking at, you know, where the space is going. I think at the end of November we started compiling the results as well. So the report actually - you know, the State of AI Agents - it's a combination of two things. So it's about 3,500 people who contributed data to this, which is completely manual - you know, you submit a survey or whatnot. But there are like 850 million agent runs, almost 300 billion tokens that we analyzed apart from their submissions as well. So it was, you know, both of those things.
Andrea: Awesome. I want to make sure that it's just me who's listening - I can hear myself twice, and I think it might be on your end. Are we open in more than one screen? Awais: No, I think I can probably connect some headphones. Maybe that will help. Andrea: Yeah, let's try that. Thank you. And sorry folks about that. I'm like, "Well, wait a minute, it's me and now I can't hear myself twice." Thank you so much, Fabio, I appreciate the check-in. And I'm thinking because I keep - listen, friends, we got to do it, this is production of one - so I always keep another laptop open so I can talk to you, and I'm thinking, "No, I probably have a window open there that I haven't found."
So thanks for your patience. Again, for all of you who heard me twice, I'm sorry. But let's get into this report. I want us to talk about - well, of course I want to talk about all of it.
Defining AI Agents: Layman's Terms
Andrea: First things first, I think we should share the link to the report so folks who are a little bit more visual can go ahead and take a look at it. And this is, folks, this is completely free - you don't have to pay for anything to access the report. I'm going to go ahead and share the link to the report on the screen. And I think there is a bit of a queue - it does ask you if you want to subscribe to a newsletter or something like that - but you can access the report, you can see the whole thing. And then at some point, I think I'd love to ask you to share your screen as well so that we can go over it.
But before we do that, I do want to understand a bit - I think it's important for folks, especially who are not very familiar with what we're talking about, to get a bit of that concept of what is an AI agent to begin with. So I'm going to do my super layman terms explanation, and then you're going to do the real professional and scientific one. The really right, because this is the way that I -
Awais: and we're going to agree with what you said, right?
Andrea: Yeah, well, I mean, I'm bringing you like the best kind of example, right? So for those - I mean, I'm sure all of you are familiar with ChatGPT. I mean, I'll be surprised if someone here has not used an LLM in one way or another since this gigantic boom. But I think 2024 was the year of agents, right? And 2025 is going to be no different. I mean, we just - GitHub just launched their very own Copilot agent what, two weeks ago? And it's been unbelievable, the response of the community and how they're using it to just build things.
The Beyoncé Ticket Example
Andrea: But from the beginning, in essence - and this is where I'll come up with my non-scientific example - let's say you want to buy Beyoncé tickets, which is something that is the reality of a lot of folks, especially folks that are near and dear to my heart. Shout out to my friend Lady Deb! So you're ready to go to the Cowboy Carter concerts and you're looking for Beyoncé tickets. If you had an AI agent, right, that's going to help you accomplish this, it could be a lot easier. It could be a much better experience for you to do it, right? Because instead of just going into Ticketmaster or whatever it is that they're selling tickets - I don't know, I can't afford them - you're going to ask your agent to do that work for you.
So it's going to help you look at many more things more contextually than just saying, "Okay, I'm going to do a Google search for where can I buy these tickets." So if I'm your AI agent, I'm going to ask you questions like - I have the intelligence, the context, and the layers to be able to answer questions like, "Okay, what city do you want to go? What's your budget?"
So this is a little bit of a different experience versus you manually filtering and buying or adding those parameters yourself, because the agent already has this information from you. So I will tell it, "Okay yeah, I'm in Florida, I want to spend no more than $200." The agent is going to laugh in my face and say, "You're not going to get to see Beyoncé for that," but we're going to stay on and try, right? So maybe we'll stretch our budget a bit.
And so if you're using an agent to do this, it's going to automatically maybe recognize the location, check to see if there is a concert in the area, check to see if there is something within the parameters of my budget constraints, check the inventory, compare those prices, consider seating options. It's going to take all those actions simultaneously - so it's a sequence of actions that are going to act intelligently, one versus the other, instead of giving me a linear response that maybe will say, "Yes, I can find your tickets in Chicago," but I'm not in Chicago - I'm in Florida, I told you that.
So taking all those actions continuously, helping - the difference between like what an AI agent in this made-up scenario would do and, I say, a simple search tool, is that it's going to keep that context, right? So I don't have to go back and explain to it that I'm broke, I can't afford any more than $200. And it's also going to have the intelligence to give me proactive suggestions, continue to handle the complex requests, right? So I'm saying I don't want to sit that far, I want to sit that close to the stage, etc., etc. And then adapting to those changing circumstances that I'm giving it.
So that's my layman terms explanation of how an AI agent can do a function for folks. Now let's hear your very technical one, I'm sure, Ahmad.
Technical Definition of AI Agents
Awais: Yeah, I think - so I have a very different take on this. I think everybody has a different example in their head of what AI is, what AI agents are. And last four years - in fact, you know, when I started building Langbase - so Langbase is a platform where we help you build agents, deploy them, scale them without having to worry about infrastructure, right?
So agents - you can think of agents as - I think everybody has kind of understood that this is the kind of thing that we need from AI, right? So AI is you give it a bunch of data and it gives you some - you know, it generates some data out, right? So it's data in, data out algorithm, right? But how do you use that data with a system design, with actions, just like you said? I think that becomes much more phenomenal and useful, right?
So for example, you telling ChatGPT something and it just sending you back some text was a simple AI use case. But when you're giving ChatGPT an Excel file or a Google Sheets file to analyze, and it's analyzing it, it's creating an environment, and then it's doing what you want it to do - it's taking actions - that becomes agentic, right? And that is a lot more helpful than just generating some data, right?
Inspired from the Human Brain
Awais: So I think of agents as - all of the space is actually inspired by a human brain, right? We're trying to map how our brain works with AI, right? So you can literally think of an agent as someone that has a lot of knowledge but doesn't really know how to do things. And if you give it access to your systems, it can maybe reliably - sometimes, and sometimes not - take those actions, right?
I can't finish this definition without talking about workflows, though. I think a lot of people confuse this, right? Like there's a lot going on in the industry. You know, there are a bunch of research frameworks where you can pick something up, deploy it - like install it in your computer and build something - but it takes a whole lot more to actually deploy something useful, right? Something that lives online, can scale, can deal with your parameters, your rate limits, and still be cheap enough that you can afford it. You know, you don't spend $200 on that.`
I think that space is - imagine like this: you know, you are a software engineer, you know how to code, you've studied or you're self-taught or you've gone to university. Now you join Company A and they have a different way of doing things, so you have to learn their ways to work at that company. And then you join Company B - all of a sudden you have to learn their ways of doing things, even though you had this entire knowledge and you were in the field for quite a while, right?
I think workflows and agents work together, right? So agents are these super smart AI bots or whatever. The moment you train them into the workflow you want them to go through - like "do this for me" - all of a sudden it becomes a more reliable, more trusted thing you want an agent to do. And it's a little bit different than automation - it's generally we call this "augmentation," right? So you're augmenting an agent in a workflow which can or cannot be very well defined, right?
But this is the very advanced definition, right? So at a very simple level, I think everybody now calls anything that you do with AI an "AI agent," right? And I think that is fine as well, right? If you're getting any kind of value from using any kind of AI, any kind of LLM, I think that is the agentic nature of AI, right? So I agree to both sides of things.
Andrea: I appreciate that because it does make sense. But yes, going hand in hand with workflows. And listen, we're here about learning, and if you want to call it an agent or not call it an agent, that's totally fine, right? And I'm sure there is plenty of discourse about that elsewhere.
So awesome. So you set out to do the survey to obviously understand - you have the data, you have like, I don't know how many users you have using Langbase, but I'm sure the survey was open. You had over 3,400 developers who responded to the survey and then analyzed - and correct me if I'm wrong - 184 billion tokens across 786 million API requests. I mean, the sheer volume of the data pool that you have to create a real comprehensive report is unbelievable.
Out of all the findings that you had from this data, what was one thing that personally surprised you the most?
The Story of Corona CLI
Awais: I think everybody is experimenting, right? That is most surprising to me. You know, when I started this company, I was like, "Okay" - like there's a fun little story there. You know, in 2020 I wrote this CLI software, open-sourced it on GitHub called Corona CLI, went absolutely viral - about 10 billion API calls in two months, right? Because COVID happened, and that was probably the only developer tool you could use to track data. And then there were a lot of copies as well, but it became our to-do list project for developers.
And right around that time, I got in touch with Sam Altman and Greg Brockman, and I think GPT-3 was just like one month old, and I got access to GPT-3 and I was blown away. I was like, you know, like I don't know - I don't want to train a model. I have background in AI and ML, I did electrical engineering back in the day. But when I found out that, "Oh, this is the first time actually AI and ML is coming to the web" - it's similar to the transition that we saw between Photoshop and Figma, right?
So design was like always - you know, you had to download Photoshop and you know, use that software, you had to have the right system specs or whatnot. And then Figma thought, "Well, web is cool enough now and has enough capability that you can design everything on a web platform," right? And they pulled it off.
I think that is what I saw happening in 2020 - that ML required a lot of training and this and that, and now it's just an API call, right? And then we waited like for three years for ChatGPT to come out, and nobody had access, and I was like, "Oh man, this is super helpful, and maybe someone can - how soon can people get access to this technology?" Right? I remember when GitHub started building Copilot early on on that same tech, right?
Enterprise is Experimenting Plenty
Awais: So what I didn't expect was the sheer volume of people and enterprise who would start experimenting. And it's beyond me that I feel like enterprise is doing a lot more experimentation than even individual developers at this time, right? Which is like, I don't know how to say this out loud. Like, this was very surprising.
And I think it's not that these companies lack ambition. I think they're experimenting a lot. I think the fundamental tension is between curiosity and scale. So they are curious about how this can help our business, people are curious about how this helps me, how I can stay up to date. "I don't want to be, you know, ousted by an AI bot. I want to stay in the workforce." And the tension is between, you know, when you start tinkering, you learn something and you figure out like a bunch of this has old knowledge, you have to have a PhD or something in - a bunch of it is, even if you do implement something, all of a sudden you have a lot of scale problems, right?
And it's not - I think it's not the lack of ambition. It's I think everybody's sort of waiting for tooling and trust, right? Most importantly, trust, right? To catch up and build and share these agents, you know, in their personal lives or, you know, in their businesses, right? And because the trust factor is honestly not there yet.
So a lot of people are not using AI - a lot of their use cases are not "spend money, go buy a ticket for me" or this and that. Maybe that is in a not-so-distant future. Right now it's "go analyze this thing for me," right? And "go build this thing for me and I'll see what you did," right?
Scaling Challenges and Growth Patterns
Awais: I always had these loads of ideas listed in Google Keep list I keep, and now I started giving it to an agent to build, and every now and then it builds something that I'm like, "Oh, this is really good," right? It works. Yeah, it works. It's like, you know, so that was the most surprising factor for me - you know, the sheer scale of how quickly we are growing.
Like when customers we got, they were like, "You know, you're going to do 3,000 to 5,000 requests per agent throughout the month." Within three months they were doing more than three billion requests per week, right? And we were like, "How do we scale this?" So every now and then we have moved from one database to another, from one technology to another.
And ultimately we figured out like, "Oh, maybe the thing that is stopping everyone is, you know, like it's the lack of declarative approach," right? Like for example, the biggest declarative language I know of is SQL. You don't know behind the scene what it is doing, but you just give it some command like "select this from this users table and only people who are above 50" or something, and it goes ahead, selects some algorithm that you don't even know about, right? And figures out what data you need without having to figure out, you know, what network latency, what disk I/O it needs to look at, or how quickly, or which - I don't know, which sort algorithm it needs to go through, right?
I think that is what is missing in this space, right? A lot of what I see are research frameworks, right? Which are again algorithms. So if you have a PhD, then you know, you're more than welcome to do that research. But a lot of people are like, "Can I just implement this in a declarative way?"
Andrea: Who has the time to do that. And yeah, I think maybe the accessibility of it made people want to be experimental with it and actually get in and work because it's not as daunting. I don't think there are very many people that can explain to you how a SQL query works. I mean, I'm sure there are many, but that's not something you're going to be talking about your day-to-day. That makes sense.
Yeah, so looking at the declarative aspect of that - I'm sorry, I cut you off, go ahead.
Component-Based Development Evolution
Awais: Oh no, no. I think the same thing happened with web as well, right? So for I don't know, 15, 20 years we had different ideas - PHP, JavaScript, Node.js and whatnot - and then React comes along and says, "Well, here are components. You build your web with components. They are similar to div tags in HTML, right? You create your own." And all of a sudden everybody agreed - even in different, you know, Vue.js community or Svelte - and everybody like, "Okay, so we build this in a composable way with components," right?
The same thing happened on the backend as well, right? Everybody was like, "You know, this works on my machine. Why isn't it working for you?" And literally, that is the problem that Docker solved, right? Like, "Okay, we will ship your machine to production and it will work the same way, and we will figure out whatever the changes we need to in the infrastructure, in the, you know, literally bare metal layer or whatever." And all of a sudden it became really easy to ship really complicated code with literal siloed teams, right? Somebody - you know, 40,000 engineers, you know, 1,000 are working on this backend API that these 30 don't even know about, and they're still able to work together with different systems, right? It became all of that became composable, right?
Now everything is like a mixed salad of sorts in AI, right? And everybody is experimenting. And I think that is what my initial thought was when I thought like, "Okay, maybe we need to build something that is composable, which is component-like React component-like, or Docker-like container-like," right? I chose to call it "pipe," but then industry kind of pivoted towards the word "agent," right? So agent is that thing that you can compose, right? You can create small agents, very complicated agents, they can be working together while accomplishing the task you want them to, right? People Prefer Composable Primitives
Andrea: Yeah, and your report and your report is something like 73% - no, 76% of developers prefer composable primitives over pre-builds. And obviously that's influenced your product strategy, right? Just as a builder yourself, why do you think developers are choosing to build composable parts? Because everybody's used to working with React now, we're all used to like the web is composable, so why shouldn't this be? But why do you think that's - I mean, 76% is a big number for adoption, like for a preference, right?
Awais: Yeah, like you have to think it this way, right? So actually, I have a graph that I like. So there's a painful way of doing things and there is an easier way, right? The painful way is where development is extremely fragmented, there are complex research frameworks, there is lack of reusable primitives, and these frameworks obscure abstractions like - and they become like very hard to debug, right? You've seen this over and over again with a bunch of technologies. Like GraphQL, I think, kind of failed because of this, right? Every time you had a bug in GraphQL, you're like, "How do I debug this thing?" right?
On the other hand, if you build with composable primitives, they're simple, they are reusable. You want an "agent runtime" of sorts where it's - imagine a way like this: you know, basically the way I can talk to you in English language - English is the primitive we use to, you know, express whatever we want to say, right? And if I can't speak English and you can't speak English, or I can't speak another language that you can speak, all of a sudden we cannot work together, right?
The Building Blocks of AI
Awais: So that is why an industry pivot towards primitives makes a lot of sense. There are not many primitives in this space. You need vector data, you need embedding, you need an LLM, you need to process data, you need to chunk data - all of a certain eight or nine primitives and you're done. And then you need a runtime that runs that thing, right?
And once you have done that, now what you have done is you have kind of built an API of sorts like, "Oh, so this is how we do things here." No matter if you're a backend developer, if you are machine learning engineer, if you're a full-stack web developer, software engineer - I think a lot of these people are turning themselves into AI engineers, right?
And AI engineers are generally, in my opinion, more than 70% of them are full-stack developers like web developers. They not had a PhD in machine learning, they will never learn machine learning algorithms. They need things to just work like when I install a database, it just works, right? I don't need to do the research behind, you know, how this database works or whatnot. Maybe give me some serverless database and I can scale it, or maybe let me host it in a VM and I'll do what I need to do with it, right? So first thing is primitives so everybody can build. And the same thing applies to the people who have these PhDs, by the way, right? Like they've only done research, they've not built products, they've not scaled systems. They have had defined primitives, they have defined criteria of doing a research. All of a sudden now they're like, "Oh, I know so much and I can ship an agent, but man, how do I charge for it?" right? And "How do I build an app around it?" right?
So again, these primitives win because of that. If you want to do - even if you are somebody from academia or somebody who is now becoming an engineer because this is where I think the money is - a lot of money is being spent in enterprise - in building some form of, you know, AI moat, right?
Scaling and Cost Challenges
Awais: So the second thing becomes, you know, I think it's one thing to actually experiment with these primitives, and the second thing is actually scale with it, right? Like I'm not going to name names, but one of my friends built this really complicated agent architecture which he's really proud of. The only problem is he can only run it once in 10 minutes, right? Because it's like 100 calls and whatnot, and he's willing to pay like 50 bucks for it. But the next level for it - when, you know, if he wants to sell that complicated agent architecture, as we call it now, he needs to upgrade to $120,000 per year package with that company. And all of a sudden he's like, "I don't want to build this company. I just wanted to build this agent. You made it so easy for me to build this with your framework, but now I want to do this," and like, that is such a big jump, right?
So deploying serverless - I think serverless is where serverless comes in, where you one-click push, deploy, and the heavy lifting of scaling it is easy. It's not your concern. Like, for example, we are processing billions of tokens every day. We've gone through so many architectural pivots like - at one time I remember sitting down with head of - I think he was head of Azure databases, I don't remember his exact title - and figuring out like, "I need this type of data which is very, very read-heavy and should be able to write this amount of 5MB rows. I don't want to shard these rows because now Google supports 2 million contacts and that could end up in one thread to about 1.5 to 5MB." And you're pulling here like, "There's not a single database that supports this right now." I'm like, "Who do I go to after - like if I can't find the solution to that at GCP, AWS, or Azure, what am I supposed to do?" Because I have a lot amount of people who want to scale this, right?
So I think it's again, it's a lot of the community here are - I think they're waiting for tooling to catch up, right?
Andrea: Right, right.
Awais: And there's an unbelievable amount of interest, curiosity, and experimentation happening, right?
Andrea: On the experimentation - because one thing is, and I love this story you told about your buddy that built the thing and now is realizing that maybe it's not scalable, right? Like "How do I even afford to make this into a company?"
Part of the research that you did also showed that effect of complexity when it comes to scaling deployment, right? I think like 59% of the respondents said that that's like their top concern. I'd love to know because this is what you do, right? Like this is what Langbase does. Like what are you doing to approach this challenge for users? And then more abstractly, what do you think - is there ever going to be a time where this is not a concern?
I'd love for us to talk about cost as well. I know you and I had a really good conversation about that before. Awais: I wish we were recording that, by the way, right? That was like so high energy. Andrea: Right, so yeah, yeah, yeah. But talk to me about okay how do I not become your friend. How canI use Langbase to make sure that you know scaling, deploying is not a concern. Building these great things and now I want to send them into the world.
Awais: Mind if I share my screen? That might -
Andrea: Please do, yeah, yeah, yeah. Let's do it, let's do it. And hey friends, feel free to drop your questions and I'll go ahead and read them out for you. Perfect, this is the page, thank you. Live Demo: Google's Rise and Model Switching
Awais: Yeah, so I think everybody can probably read through the report. Some of the surprises we saw was like Google scaling really, really well. When you sign up on Langbase, you create an agent that we call "pipe," right? But the default model was GPT - version of GPT, it was 3.5 and then 4o-mini. But then more developers using Gemini Flash model on Langbase than any other model, right? And they're manually switching it, right?
Andrea: Interesting.
Awais: So that was like Google finally catching up, I guess, right? That was like, "Oh, okay, maybe it's cost or maybe the model is really good," or whatnot, right?
One more thing we do is like we make it very, very simple. Like for example, this is the UI part, right? Let me actually share one thing, right? So Langbase is a serverless platform to help you build composable agents, and composition comes in many, many ways, right?
Email Agent Architecture Demo
Awais: So like, let's give - let's take this very simple example. Let's say you're building an AI email agent that, you know, helps you respond to people. And if you get a spam email - like this is a particular flow I actually have. Each box here is an agent. One agent is just figuring out what the sentiment of this giant email is - this could be a lot bigger, right? The other one is summarizing, right? They're doing this in parallel. And then there's the decision maker who decides if I need to respond to this or not. This is clearly spam, and it figured out the category for this as well, right?
And then if I get a valid email, again, one agent to figure out sentiment, one agent to summarize. The decision maker actually decides that it needs to respond today, right? Because this is live, this is not a dummy example, right? And then it sends all of this information into an email picker. So I have eight different types of email writers, so I respond to my investors differently, customers differently, enterprise customers differently, right? And figures out the right email writer for me, which has the right amount of context about the customer who is sending us that data, right amount of memory, and sends that email, right?
So this composition is very, very powerful. If you try to do a bunch of these things one-shot with just one model, you will realize that, you know, the amount of times you will be successful in that is very, very low compared to this approach, right?
I can literally jump into one of these agents and change everything about it. Come on, internet. So right now, I'm basically just explaining, you know, what kind of decisions it needs to make. It has access to the summary and sentiment that previous agents sent it, right? And this is what I want in response. I can go down and change the model to 250+ models without actually changing any API code in this architecture. So that's the one thing - first thing we offer in pipe, right? It's a unified API layer over 250+ LLMs, right? Open source or proprietary from Google, from Anthropic, everybody, right? And nothing else changes here, right? So this composition is really, really helpful in developers just seeing that, "Oh, Google released this, let me see how this does on my current workflow" - flick, flick of a switch, you know, or an API code change.
The Langbase Platform
And that brings me to this thing, right? So we basically do everything, right? We have a studio for anybody who is non-technical or a stakeholder that you - you just, you know, you just saw here where you can basically use UI to create agents, right? We have an SDK which we are very proud of - it's built for TypeScript, not Python. I have roots in web development like I helped React get open sourced, I was part of the core team at WordPress, Node.js, and a bunch of, you know, web roots where I come from. And I think this is the biggest beneficiary in this space - the people who are coming who are full-stack developers who are now becoming AI engineers. They don't know Python and they don't know how to deploy Python, right? They don't want to learn too much to do what they want to do. They want to keep doing things with their stack, right?
So and then we have an API so, you know, API can be used with any language, right? You, Python, Go language, or whatnot. Then we have a framework, so if you know - we can touch on this a little bit later because if you want to do a lot of experimentation and you want to do it locally with something like Ollama, you can just use our framework and do it locally. And when you're done, you can then deploy. All of that experimentation remains to be free because it's local, right?
Memory and Agents
Awais: So now, a very basic example of this would be - let's say this is a simple chatbot and let's say this is embedded in a documentation, so you can ask questions from it, right? And I'm right now building it, and I'm asking it to ignore all of its previous training, and then the user is coming in and asking a lot of questions, right?
So, "What are the features of Langbase? When was it created? Who was the founder? How tall is the founder?" - that's the fun one, right? So if you run this through any of the LLMs right now - this is GPT-4o - you will see that it doesn't know anything, right? Because I asked it to - yeah, right? It does, but like I asked it to not answer from the pre-training, right? I want more control over the answers, right? I wanted to look at my data and answer from that versus "I don't know what it is actually trained on," right?
So now what I'm going to do is I'm going to give it memory, similar to how us humans have memory, right? And I'm turning two memory agents on here. So this is memory of all of our documentation, and this is the memory that has some health data on me, right? Which should include how tall I am. This one does not include that I'm a founder, right? My doctor doesn't need to know that, and this one does not know much about my health data. So these two data sets, memories, are very different.
Now what you will see is we will be able to reason over all of this data, connect those two memories, and enable this LLM will be able to answer these questions, right? So now if I run this, it will take its sweet time - it's GPT-4o and it's already writing the features of Langbase. And there you go! Ahmad is 185 cm tall, right?
So this is pretty awesome because I did not ask how tall Ahmed is, right? And this is the agentic reasoning of it, right? So from this question it figured out I'm the founder, and from this question it automatically figured out "oh, I need to rewrite this question." And it didn't find the answer in one memory which was about my product docs - it was able to answer that question from another memory, but it didn't have any information on me being a founder, right?
Awais: So it did a lot of things that I can never predict from my users, right? But AI is so good, agents are so good - they can reason over data, they can rewrite queries, they can re-rank the answers, and then they can predict that "oh, so this is what he's trying to ask and this is where it is." And it connected those two memories which were very different and then answered that, right?
Switching LLMs On the Fly
Awais: And now what I can do is I can just go to a cheaper LLM - maybe let's try Gemini 2.0 Flash which they just launched - and nothing in this entire complicated setup changes, right? And you can see that, you know, even Gemini 2.0 Flash is able to answer these questions, right?
And all of a sudden you go down to runs and just refresh it, and you see a huge difference. So this was GPT-4o, this was Gemini Flash 2.0 - same answers, right? Run it a million times: this is $20,000 versus $500, right? And I think it's also faster. Yeah, it took about 1 second, this one took about half a second. Abdrea: Half a second! Wow.
Awais: You can go deep, deep into this. You can see, you know, what was the prompt, what were the augmented chunks, and all of that.
Awais: So the value is not actually accruing in an LLM. I think LLMs will keep updating to a point where we will not know how they did what - like, you know, what are the actual changes? It would not have any point to know. It would be like "Oh, I'm using the latest Gemini model". But the value will accrue in how you're composing this giant composable system and how you're deploying it. And we call it your cognitive architecture - your neural architecture for the lack of better terms - and how you teach it to work, right?
So you make something very intelligent and then you give it some guardrails like "don't do this, don't answer that question," maybe "remain close to this particular memory," maybe "only talk about these topics and not that." Maybe if you are asked a lot of questions from our enterprise customers, yeah, "talk to a human, get his feedback," right? That is where the value is, and that is what a lot of these companies are building.
Scaling Challenges and Solutions
Awais: They are able to build it. In the moment they have this "aha" moment, they're like "oh, I can't scale this. I have to pay like $100,000 of enterprise fee to just do like, you know, I don't know, 100 concurrent runs of this agent architecture." And that's where we come in. We want people to be able to scale these priand build their architecture easily, right? Without actually having to go to AWS or completely... like, I'm not against AWS - I use AWS - but that's not the job to be done here.
Your job to be done here is building an agent that is performing some random task that an automation could never have completed, right? Not an automation - you have to have exact guidelines of "do this and that and this and that." It cannot survive a change in your environment where it is working.
Awais: But in an augmentation, your agent can understand things, reason over that data, and figure out - can I answer this email? Can I draft this email? Should I need more information from my human to make a decision on that? That's where the value is, right?
So agents are not that important, but what you build with those agents... I think we don't yet have a name for it, and I generally just call it "agent flow," right? Like agent plus your workflow - the flow in which agents work with is different for every company, different for every person. And that's where the value is.
Andrea: Let me ask you, because I think - and I don't recall if I saw a part of this conversation on the report - about how, for example, enterprises are making the decisions of what models to use when it comes to that experimentation. Because usually that experimentation comes with a cost, right? But I guess it's not so much about the model but what you just showed us.
Sure, you're making decisions based on like whether it be the expense or the efficacy, whether it's like "stop answering questions there because you're never going to help this person - transfer to a human," for example, if you're online trying to, I don't know, return something, etc. So how do you see from your experience with the report, and particularly how people are - how enterprises - making those decisions as to what specific models to use to build their agents with? Are they taking - is the cost a factor? Is the efficacy the factor? It doesn't matter - you have hundreds of models that are available for you to experiment with.
If you're a company right now and you want to build, say, a customer care assistant, how do you go about figuring out which model would be the more... the best for you to get started, and then exactly scalable and affordable?
Model Specialization and Trust Factors
Ahmad: So this is what I see right now, right? Everybody wants to experiment a lot faster without being bogged down with one system, or that by the time they are doing this particular experiment, the model they were experimenting with has already been, you know, ousted by another model. So they want to switch really quick. So everything they want to do is very, very fast.
And the moment they hit some form of "aha" moment, they want to quickly scale it, right? Without having to, you know, like "oh, now we ship it to our machine learning engineers." The problem is they are only like, I think, a million or less than a million machine learning engineers, so there are like 50 million developers, right? Or 100 million on GitHub, right?
So developers are going to build this, and they don't know machine learning. How will they scale it, right? So that's the second problem.
Awais: And then I think then comes the trust, because the cost of intelligence is becoming so low. Nobody, at least now I talk to, is worried about the cost of intelligence - the cost of LLMs, right? People are more worried about the cost of scaling this and the cost of spending time on building it, right?
But we do - that's why we did the survey, right? We do share a lot with our customers like, you know, people are generally using OpenAI to write content, right? Marketing - you see this writing? There's nothing better than OpenAI in my experience and what we have seen. And if you are translating something or even health data, Google kind of beats everybody, right? Different models, right? If you're writing code, it's very, very hard to compete with Anthropic. But all of a sudden, you know, now there's DeepSeek, right? DeepSeek is kind of very, very cheap. But you have to decide, you know, what type of customer you have, and for this type of customer, for this type of query which is very simple to implement, you can just do it with DeepSeek.
And if they say "oh, it's not really working out really well for me," and if your thread is like three chats deep already, then maybe you can switch to Anthropic and offer "can I upgrade you to this experience?" and, you know, run it over an expensive model and give you an answer. And you can just say "okay, I'm willing to spend this much on this problem," and then it can reason over that, right?
That is the experimentation layer, you know, that we offer. And imagine - I think the value happens when you know you have this kind of architecture you built, you have this problem, and you built an agent. And now you start to, you know, do this, right? Now you need to experiment with this model, that model, this API, and then inside of each of these you have evals where you're evaluating this for I don't know, three or four, 50 different outcomes. All of a sudden it's a pain if you have to move from one API to another in this giant architecture, right? So you want to move fast.
And we have also seen that this space has become much more visual than as a developer I was okay with, right? I was like, you know, you probably need to see what we are talking about, right? So there's some visual component to this architecture that has, I think, never existed before. And maybe I see some form of that in DevOps - like for example, if I do DevOps scale diagram or something, every time I talk to a guy who's doing DevOps, they come up with like "oh, so this is how we will do this." We never talk about code - we talk about the sequence or, you know, some form of reference architecture or whatnot, right? So that is how complicated the space is, right? So having some form of data at scale, so you can decide before having to go through like "okay, I will do everything from scratch, I will start with the research framework, I will be embedded down, I won't be able to scale." You decide with primitives, you decide that no model is going to be good enough - you will have to change that model every single day if you have to. And then you come up with these insights where like "okay, I want to reduce cost, I want to reduce latency," or "I want... I don't care about latency."
Like I recently... so my father is a doctor, and I grilled him for like "how does human brain work?" and like, you know, whatever I got out of him was like, you know, we have different cognition levels. One is just gathering facts - C1, they call it cognition one, right? The second one is just trying to figure out what these facts are, understanding those facts, and even animals have that - like your dog has that, right? And then there's a problem-solving C3 level where we can solve those facts after understanding them, and even animals have that. And then there is C4 - that is the management of all of this. Like which problems do we need to solve? When? What to prioritize? How to do it? That is something that we only see in humans, right? And now we are seeing that kind of understanding, reasoning, reflection in different LLMs.
Awais: And then one thing he said that - and maybe I have heard it somewhere else as well - like we think about things in two ways. So you say something to me and I answer you back, and then there's a part of my brain that will just think about it over the night, right? Even though I've already given you the answer, we'll keep thinking about it and it will realize "oh, I should not have said that," right? "I think I was wrong," right? "I should not have sent that email, maybe I did that too quickly," or whatnot, right? That's the... I think there's some form of intelligence that requires latency, that requires an architecture that can reason over it over a length of time to figure out what the right outcome is.
Andrea: That's the difference between using an LLM like that versus using an agent, I guess, because you can ask the question, but in answering, the question is not going to stop and think "is this the correct path to what you're asking me?" So I guess the agent would just be like "slow down, think through the task," and then you're going to get a better outcome. That is fascinating, fascinating. And our brains have been doing it forever, folks - we are the original! Learning from Human Brain
Awais: Exactly! And literally, I kid you not, we literally read human anatomy because we - after reading hundreds of research papers, being at the cutting edge of this space - we run out of ideas to how to fix this problem. And there are several ideas like, you know, how would you think about it, right? But there's a bunch that we know already about our brain, right?
Awais: So just implementing those parts like... like I was talking to my father and he was like, you know, so your brain has nerve cells - let's call them agents. And then I recently figured out that our brain is extremely slow actually, right? And we process, I don't know, billions of bits of data, but ultimately I think that was like at 10 bits per second speed of our brain or whatnot, right?
So my father called it "process" - like when you touch something and ultimately you feel that "oh, it's hot or cold." The things that move across your, you know, to your brain... in I think medical science they call it "process," right? That process is, I think, the tooling that we need to figure out so we can filter it out and give our agents the right kind of information that it can work with, versus trying to have like 100,000 agents doing one thing or whatnot, right?
SDKs and Tooling
Awais: That's where the tooling of all of this - that's where I think the human aspect will come in, right? Where we will build the right SDKs and we'll teach LLMs agents to use those SDKs versus trying to do everything by yourself.
Like for example, if you want to send an email, it's much better to use Resend's SDK and write code from Resend's SDK versus, you know, like for example... I think they just did new.email recently - they're building, hopefully we are going to help them build this as well. It's like v0, but what this thing creates is an email template for you, right? And it uses their SDK instead of from absolute scratch, you know, using some random React framework or whatnot - just uses the right thing, right?
So instead of like creating GitHub, I would just rather use GitHub's API to figure out a certain data set, right? In that way, training my agent on GitHub's API is much more useful to me versus just depending on whatever the training data set it has, right?
So we're learning from our brain, we are learning from how humans work, and in this entire SDK space, this entire tooling space, I think that's where humans will come in. They will learn this new way of feeding and training this agent - new world of "this is what we want done, this is our workflow. Now train on this and help our customers," right?
Langbase SDK vs BaseAI
Andrea: You touched a bit on SDK and I want to make sure we get a question that we had earlier from Fabio. And thank you Fabio for the question - appreciate you hanging out here with us today. Okay, so given that Langbase provides an SDK, right, so I can build already - I get an SDK that you're giving me - why do I need BaseAI?
Awais: Yeah, so BaseAI is an open source, local-first, meant for web framework, right? It can do things that our SDK cannot do, right? For example, if you want to build a chatbot that learns from your documentation, right, it's actually very easy to do that. If you go in our docs, in guides, "add AI in your docs" - really simple steps: memory, agent, and a bot. Memory will have all of your documentation. Agent will reason over that, and the chatbot is how people will talk to that agent, right? The problem, however, is... let's go back to BaseAI. Oh, I keep clicking on the wrong thing. This is open source, right? And it has a lot of files in it, right?
Now I need to - if I use just SDK - how do I keep this information that is in my memory up to date the moment I update anything in our docs, right? How do I make sure that, you know, for example, if I go here in docs, in content, you know, there's "learn" and "how to configure a pipe." If I change something here... let's say temperature is not called temperature anymore, right? How would I update this memory with the right updates? How would I remember?
Awais: It's the similar problem that Github fixed for us, right? Like there are lots of changes happening in our code and teams would drift apart, and then GitHub came in and they're like "oh, you can do pull requests," right?
So what BaseAI can help you do is BaseAI can just... you can attach it with your Git and you can say "well, just keep an eye on this." And every time you deploy a new memory, it can look at the last time, last Git hash, Git commit hash that you used, figure out the diff from all of that, figure out what actually changed, and only change those parts in our memory.
We didn't want our users to build this tooling because our SDK doesn't do that, right? This is not something that everyone has this problem, right? But a lot of people do. So BaseAI kind of comes in when you need to go a little bit step further, and BaseAI also comes in when you need to just experiment with something local like Ollama, using Ollama models absolutely, you know, cost-free on your system, right? Our SDK, in fact... BaseAI actually, if I were to draw this... so this is our Langbase SDK and it is being used inside of BaseAI, right? So this is BaseAI, right? So both are meant for different things, but if you're just solving simple problems, I would just start with SDK. You will know when to go to BaseAI - you will know, you'll get to a problem like "oh, now I need to do this, this, and that," and then it'll be like "no, you can just deploy it through BaseAI," right?
Is Langbase Open Source?
Andrea: So that's very good, very good. Thank you so much for clarifying, and thank you Fabio again for the question.
Andrea: I want to - and we're almost at time - Ahmad, you and I talk too much. I want to make sure that folks have a chance to look at the report and they can noodle through this data on their own time and of course test out Langbase, give it a go. There's been a couple of questions on the chat about whether or not it's open source, and so I want you to please clarify there, the open source aspects. And then also, before we go, I do want to make sure that we touch on things that maybe any results from the research that you think are not being talked about a lot, and there are things that you feel are definitely going to be impacting the industry. Though we did talk a bit - you know, the concerns with scalability, the concerns of the C-level people, I guess, versus the folks that are writing. But I do want to make sure that we touch a bit on, based on that report, what do you think is not getting significant attention? But talk to us about open source first, please.
Awais: Okay, yeah. So our SDK is open source, BaseAI is open source, how we are able to unify all of these LLMs into one single API - that is open source, right? It's part of our SDK code as well as, you know, our BaseAI code, right?
The infrastructure, the cloud part, is not open source. Like that is the cloud - that is the infrastructure. I don't know how to open source it, you know. We have a cloud which is serverless where you deploy all of your agents, right? That is... I think the better analogy for that would be how Next.js is open source and Vercel is the platform that you deploy Next.js to. So Langbase is similar to that - our platform is the cloud, right? But our SDK, how we do things, a bunch of our research, a lot of - I think 200+ examples - are open source.
GitHub-Inspired Features
Awais: And there's one more thing to this as well. Like so our platform has over 100,000 agents now, right? Deployed. And many of them... like on GitHub - we are inspired by GitHub a lot. In fact, if you go to our About page, you will see that Tom Preston-Werner, founder of GitHub, was one of our first, you know, backers, right?
Just like GitHub, you can see on our Explore page... I lost the Explore page... there you go, right? There are a bunch of people creating agents that are public and open source, right? So for example, "YouTube video recommendation extractor" - this guy built this thing. Similar to a GitHub repository, this is public. You can see their prompts, which are open source. You can read through the reasoning, you can fork this into, you know, your own account or something. This part of collaboration is also there, right? So it's part of open source, I guess, right?
And then a bunch of our code, our infrastructure code, is open source, right? One last thing I'll mention is we also acquired a project called LangUI, which is like... if you want to build your own ChatGPT, we have 60+ open source components. Like, you know, you want to build a sidebar that works on mobile and everywhere? There you go, right? In fact, I was checking out Bolt - Bolt uses and trains on all of these components to build... yeah.
So yeah, that part is open source and more than welcome contributions from the community. I am also an open source engineer, so that really helps.
Andrea: Awesome, awesome, awesome.
Andrea: Okay, so for the folks who are watching, and I'm part of that - two questions, then. Let's talk about moving beyond experimentation and throwing these things into production. What advice do you have for companies or engineers who are watching and have the project and they're ready to do that?
Awais: I think one of the things that I really liked recently was... so while we were doing this research, one of the frontier labs - Anthropic - was also researching. So we shared our "State of AI Agents" and they shared this almost at the same time, and we both kind of came to the similar conclusions. And they kind of like... not nobody actually sees eye to eye to everything, but they said "workflows are systems that, you know, are completely predefined, and agents can make autonomous decisions." I think you can make autonomous agents and then train them on your workflows to make them reliable, right? And then they said "when or when not to use agents," and then they went on to share a bunch of frameworks that we see in the market right now that meant for research, not for actual product deployment - or maybe now they are big companies and they're kind of pivoting towards that.
The Problem with Over-Abstraction
Awais: But what they saw was they often add extra layer of abstraction, which basically hides the prompts. Prompts are how we talk to these agents, and then they don't update these prompts, and then it makes it very hard for you to debug. And then they ended up recommending that a simple LLM-augmented API is much better in most of the cases - not all, obviously these frameworks have value, right?
And they ended up creating this LLM-augmented architecture, which is what our Pipe is. So it is an LLM, it can take any your input and give some output, it has RAG memory, it has some tool calling and memory layer attached to it. And that is literally what my idea was when I started this company - like we will have PIPE: Prompt Instructions Personalization Engine. We connect an LLM to your data and to your developer workflow, right?
So this is not a framework - this is a primitive. I think what they are trying to say is primitives are much better at this stage, so you can quickly experiment, swap out one API for another, swap out one LLM model for another, and anything that, you know, any framework, even any SDK, even that promotes this architecture is really, really helpful, right?
Building Cognitive Architecture
Awais: And then they go on to say, "well, there are really a few different architectures," which you can also find in our documentation. We implemented this to showcase to people how to do all of this simply with primitives without actually using any framework, right? And it also is a really good exercise, right? Like developers want to understand the system thinking behind it. Even if they don't want to learn the machine learning algorithms or whatnot, they want to figure out like "how does this system work?" In building that system that I call your cognitive architecture, that is where the value is, right? It's not in the LLM - it's in what your agent architecture looks like and how you build it, right?
Andrea: And that's something that Langbase does very well - makes it easy for you to design and test and experiment.
Awais: Because we're not a framework, we give you the primitives to like "here's a LLM primitive, here's a chunking primitive, here's an embedding - you want a vector store? Here you go." And we also, like, you know, we do a lot, right? So you don't have to, like, you know, connect to a bunch of different services. We give you a vector store that we built, right? We... in our memory system, in our memory agents, we basically, after learning so much from 200 billion tokens, we created a re-ranker and rewriter, query rewriter model that is just there. It's working for you for free - we don't charge for it, right? But it's helping people build those architectures, right?
Andrea: Okay, well, thank you so much for sharing that.
Undervalued Trends in AI Development
Andrea: And I want to make sure that I do get this last question out. I want you to highlight - and I think actually it's probably what we were just talking about, right? The thing that is not being spoken about a lot that there isn't even a name for it, or you're calling it workflow - what is it, agent workflows? What's the from the report, the findings about how people are using agents to build anything that caught your attention as something that is not being very mainstream, not very discussed a lot, but you see it as something that could potentially be very impactful to the industry?
Awais: I think my major gripe with this world right now is a lot of people create a lot of slop, as we called it, right? So it's like a better version of spam when you say "oh, AI doesn't know this," right? So it's really... doesn't, you know, it's really dumb, right? We call it slop just because you don't know how to use something really well doesn't mean that thing is not good.
So what you're doing is it's a, you know, AI version of spam. You're basically creating a bunch of hype around media expos or whatnot, right? And what that ends up doing behind the scene is it discourages an average developer from like "I don't understand what is going on," right? And they feel like, like, maybe I'm not capable enough. And then they try this experimentation - they have to do it with probably a different language, Python, right? And then they have to say like "oh, I just talk to this piece of code and just tell it to do this thing." That feels so weird, and then they get stuck in this loop of experimentation and that, you know, instead of actually trying to solve a problem.
Making AI Accessible to All Developers
Awais: So my gripe is like not a lot of people who can build agents are building agents. And my mission is to make it so easy for you to build an agent - maybe through an API, maybe through an SDK, through a framework, or through a UI - that you feel like "oh, I can do this," right? Because that is what this space is like - you know, an average technical enough person can build an agent, right?
The rest of the things, I think there are so much happening in the industry - so many tooling startups, so many, you know, people innovating - that you will figure out a way. Maybe it's composable, maybe it's serverless, maybe it's not, maybe it's something else, maybe you want to do on-prem - we don't do on-prem right now, right? There's a solution to everything as far as, you know, that this is not a space where, you know, like we used to have the space where you would spend $100 million into an AI problem, hire a team of like, I don't know, 20 people, create a research lab, and in 5 years you will see some ROI, right?
Now the space is you don't need a PhD, you need some level of technical knowledge, yeah, and you just... you can just build agents, right? And the amount of people not doing that's why I said enterprise is experimenting more than individual developers - which is like beyond me. That was never the case. Companies are more interested in implementing solutions than an average developer. I think an average developer needs to understand that they can do this, right?
Andrea: Yeah, yeah, I appreciate that. And I will say though - and I have to - that even though you are creating the tool that is making it easier, like for people to understand, and maybe I think, yeah, like the way that developer's brain works is like you do need to have... it's like you have to know how it works to a certain degree for you to be able to like get out the concept in a way that you know... I gotta know how the sausage is made kind of thing in order to be able to feel comfortable and actually getting in the kitchen. And so there's a lot of stuff out there also to help you, even if Python - I know that not everyone is a Python developer, and that might be the barrier as well. Like a lot of people are stuck on like "do I need to learn an entire new language in order to begin experimenting with that?" And we see Langbase, we see tools like GitHub Copilot that are making it so you don't have to really... with the general understanding and being technical enough, like that should be enough for you to get in and experiment.
So more developers experimenting, more people building quality agents, no more garbage, no more slop, more enterprises continue to experiment by moving from experimentation to production. And I guess that's this year, Ahmad? Is that this year? Is that next year? Like what is your outlook for like the next couple of years in this space?
Future Predictions
Awais: Oh, that's so hard to predict, right? So like…
Andrea: We're gonna hold you to it, too!
Awais: Oh man, I always enjoyed predicting, you know, like, you know, "next 5 years we will have this, we will have that, we will have this, this, and that." But I don't know... the kind of experiments I've seen happen in the space like... like in 2020 I knew ChatGPT or something like that was coming, but it took 3 years for OpenAI to launch it to general public, right?
Like we now know like... it used to be very hard - like for example, if you want to train a robot, you have to program it like "oh, pick this thing up," and because this was not as, you know, the weight was low, not that heavy. If a robot is picking up something heavy, it would just mess it up, right?
Now we are in a space where robot programming could be runtime agent, so it can program at runtime literally how we do it. I have never picked up this chair - I'm going to pick it up and instantly my brain starts to adjust how much force I need to exert, right? And the second time I do it, I know I'm not going to pick up that chair - that's too heavy, right? So I think we will soon see a lot of agentic workflows, a lot of trust in this space, but it will not be everywhere. It will be in these particular spaces where deep research or some form of operation, a lot of grunt work, and that's what I think. Like I don't think... I think these agents will actually end up working with humans - they cannot innovate, right? They cannot replace our brain, right?
Andrea: Thank you. The human element will become really, really important, and it's now more than ever we... like learn all of this. Like you can use GitHub Copilot to train on our documentation and ask it questions and build things the way you want them built, right? But you have to do it right. If you just randomly create an agent architecture that does it, it will probably not do a fine job at it. Awais: Yeah, it turns out our brains were the original blueprint for this all along! I love these analogies about the medical and anatomy - like that's... you never mentioned that to me before, and that blew my mind. Like because it's right there - it's like there is so much wisdom in nature and in our own bodies. Andrea: Not for real! I believe in this 100%. The answers are there - we are the original blueprint for this.
Upcoming Events and Product Launch
Andrea: Ahmad, this has been such a good conversation. I appreciate you joining us. I am so sorry, but I'm also not sorry that I kept you way longer than I asked you. I was like "yeah, we'll do like 30, 45 minutes" - never, never, ever.
Andrea: I asked you - I was like, yeah, we'll do like 30, 45 minutes. Never, never ever! Um, what's next? Online-based world - I want to make sure that you're able to share, or do you have anything coming up that you want to share with folks? Like, are you going to be speaking somewhere? I know that you usually do quite a bit of public things, so...
AI Engineer Summit and Langbase Updates
Ahmad: So next week, Langbase is sponsoring AI Engineer Summit. So Langbase is literally AI primitives and AI agents runtime. So if you are becoming an AI engineer, we help you with full-stack AI engineering - you know, deploying your agents, learning every all of that thing, right? I'm talking at a meetup next week as well. Tomorrow we will announce a giveaway of a fully paid ticket to AI Engineer Summit - about, I think, $3,000 worth, right?
Ahmad: One more thing - we are launching a major product that we have been cooking for probably the last six, seven months now.
Andrea: You've been teasing this on social! Waiting for you to say what it is!
Ahmad: Oh man, we are going to make it so easy for anyone to build agents that you will be like, "Oh, why didn't I see this?" Right? So I'm pretty excited about that. You know, follow us on Twitter, and maybe if you are at AI Engineer Summit, see me there - we have a booth there and I'll be talking as well. Event Details and Giveaway Information
Andrea: Awesome! It looks like a really amazing lineup of speakers that are going to be there. So this is going to be in New York next week, folks, for all of you in New York or folks who want to make it. I'll share the link in the chat and it's on the screen now as well. You have to share with us your giveaway so people who are - you know, maybe that's the resource is a bit tough - like definitely they can enter the giveaway so they can win the ticket. And then you just share with me your Discord.
Ahmad: Why am I on the Discord? Yeah, we just started a new Discord server. Not a lot of people there - it's super new, so now is the time to get in, right? The entire team is hanging out there, and you'll be doing a lot of live streams - agent architectures or whatnot. So if you want to do it with us, you know, join us on our Discord.
Andrea: That's awesome! Okay, you have to connect me with someone from your team that we can do a stream where we actually build one from zero.
Ahmad: Yes! I think that will be super interesting. That will be super interesting.
Andrea: Not to help me spend money, but something - definitely not using agents to spend money, right sir? Ahmad: No, no! We're trying - you're trying to save us money here, so we'll stick to that.
Closing Remarks
Andrea: All right, thank you, thank you so much! I super appreciate you. It's always a pleasure to have a chat with you, and I learned so much today. I hope all of you watching did as well. Please follow them on socials because you put out a lot of good stuff, and that way you can see this announcement that's going to revolutionize the way that you're building agents - which I'm waiting, I'm waiting to see! Hopefully... Ahmad: Well, you're too kind. Thanks a lot for having me. This was super, super fun. Yeah, we should do this often. Andrea: For sure, for sure! Thank you so much. I appreciate you. Have a great rest of your day.
Ahmad: Thank you. Bye bye!
Andrea: Folks, thank you so much for joining! I appreciate all of you warriors hanging out here past the hour. I knew this conversation was not going to be quick, so I need to stop lying to guests when I have guests that are this awesome and just tell them to come for an hour and some change.
If you had any questions that were not addressed, I think I did a good job at covering the chat, but if there is something that you want to learn more about - if it's Langbase - join their Discord! Like, very few times you actually have a whole core team in the Discord hanging out ready to chat with their users, so take this opportunity and join their Discord.
If you are going to AI Engineering Summit next week in New York, make sure you say hi to Ahmad, go by that booth, watch the demo, build an agent, do some things. I'm really going to come through on that ask - it's more like, "Can you please do it?" and connect with someone so that we can actually do this on a live stream. I think that will be a really interesting thing for all of you to see and go experiment.
Thank you for being here! Don't forget - all my super serious ask - to join our little Valentine's Day giveaway. It's not quite as fun as a ticket to a conference, but it could be access to GitHub Spark so that you can go and do some experimenting yourself.
So if you had that idea that was in the back of your mind and you just didn't know how to go about it, what stack to use, how to - like, GitHub Spark does all of that for you. So join the giveaway! All you got to do is drop a ChatGPT-made or one from your heart - a little poem on this discussion. Share it with your friends too! The slots for the beta preview are super, super, super, super hard to come by, so definitely if you're interested in testing GitHub Spark, go in there, share those roses, and participate in that campaign.
Thank you again for joining Open Source and AI! We'll see you again - I'm going to be back here tomorrow. We're going to be talking to Danny about some cool things that they did. They helped Wikipedia - yeah, they did this whole translation thing using LLMs, and it's going to be super interesting. So I'll see you all tomorrow.
And again, don't forget to follow us on Twitter and please subscribe to the YouTube channel so that you don't miss any of these announcements. Appreciate you all! Thank you for being here. Now stay here so that you can see what GitHub Copilot agent and VS Code can do for you, and we'll be announcing another series of streams that we're going to be digging into agent mode a little bit deeper.
Thanks again for being here, and I'll see you tomorrow!
