How CopilotKit Builds Open-Source AI Agents: Atai Barkai with DVC's Nick & Marina Davidovs

How do we make agents work with humans?

The agent is better at some things. The user is better at some things. And together now you've kind of 10xed.

What are your metrics?

Our strategy is to be that default. It was open source before it was really a company. Took a life of its own pretty quickly. We're selling 100x increase. Imagine that you're hiring a multiple PhD professional. But they have one problem. Say what you want to do. Say what you want to learn

Welcome to our next episode of deconstructed podcast where we are trying to deconstruct businesses to different blogs and trying to understand ideas processes which are beneath the surface. Today we have an amazing guest who is building an open-source technology and a very important technology for the current world because you're hearing agents everywhere like agents agents are going to replace TV I think at some point but how do we make agents work with humans and a lot of a lot of times you might have heard this word or rather a phrase human in the loop. How do we enable people to loop in their input into what agent does? And the answer to that is CopilotKit and its founder Atai Barkai. Did I not I hope I did not butcher. Okay, good.

Pretty good. Yeah.

Yeah. Welcome at Deconstructed VC.

Great to be here.

so just maybe to start things off, how does one start an open source company? like what led you to it?

Good question. So, maybe I'll briefly say what we do. So, copilot kits, we make infrastructure as you said for co-pilots, which are userfacing, user interactive agents as opposed to fully autonomous agents. and we actually pivoted into this from what we were building before, which was an LLM based media database. and Copekit started as an internal tool that I open sourced. and kind of took a life of its own pretty quickly. and so it was open source before it was really a company is the short answer here.

That that that is a good answer and I think that applies to the majority of successful open source startups because you you just build something that you want to use yourself and then other people start using it and then like I could maybe build a business on that. Yeah. Yeah. No, that's that's kind of how we thought of it. And you know, early on we had a discussion, you know, should we keep it open source, right? Because it was open source by default, but do we want the business to also be open source? And I'm very glad that we landed on the open source side. you know, our strategy as a whole is to be the default solution for building user interactive agents. you know the world of developer infrastructure if we're going to fast forward a few years there's going to be a few solutions that everybody uses including the very large companies and so that our strategy is to be that default and from a business standpoint you know being open source you're now a pillar of the tech stack you're not just selling features and functionalities so that's our strategy and it's yeah been

how did you make this decision actually

well we realized okay how fast forward right work backwards from the end result. what is it going to look like once this field has been set up? Developers, you know, the money comes from the very large companies, right? So that's where you have to monetize. They have the budgets.

but the very large companies, they use whatever the ecosystem designates as the thing that you should use to solve a problem. They very often don't have their own weird solution that just they use, right? So if we look at the previous world, everybody uses Nex.js, right? The the indie developer just starting out and chachbt. So we wanted to become you know essentially the Nex.js of the agentic front end. and so the ecosystem kind of you know enshment is the number one priority. I think that's often true in general with developer infrastructure but especially so in this field because it's you have this vacuum that just opened up you know a couple years ago and becoming the solution that fills the vacuum is is really important. So that's kind of our number one priority. become the best solution that developers want to use that they can build on top of that they can you know grow an ecosystem around which we actually are driving even harder towards.

So we started open source but now we're also open protocol which is kind of an additional open source layer we can talk about and then work from that and you know you don't need to monetize everybody you can build the enterprise platform monetize at the top.

Mhm.

And let's maybe let's maybe dive deeper into this. So let's let's try to draw like a map of so let's say we notice this a lot like every time a larger company is looking to build agents or use agents

they're always facing this buy or build so there's a startup pitching something to them that is like end to end 360 gives them the the result but again it's a startup so unreliable

and their own engineers would always raise a hand and say hey we can actually build it using open source internally is just going to be our stuff and like open source going to last like even the startup behind it doesn't.

Mhm.

So I'm a larger company. I want to automate my customer service. I know that human in the loop is probably the answer because a lot of time like somebody wants to escalate something. We need user input. A lot of times you know we might not have enough data to to have fully autonomous agents running. So what technology stack from open source would I use? Yeah, so it's a great question. and there's actually there's two ways to look at Copekit from a business standpoint. One of them is actually this kind of bridging this tension from the all-in-one drop-in solution that just works and the low-level control.

so we're actually serving both of these markets and kind of slowly meeting them in the middle. And ideally you can pick and choose. So there's things where the drop in solution is all all you need and then you can save all this complexity and you can have your application engineers engineer it rather than the kind of top 1% AI engineers. but you really do need both. and so if you're looking at the tech stack we initially started Copa Kit with what we call kind of the all-in-one agent that is easy to use and then we built co- agents which is agent the same agent user interactivity infrastructure that we built for the all-in-one agent. we brought it to the low-level agent framework world. So we started with Langraph which is kind of the biggest agent framework. So after we built this cogence for the Lang ecosystem this really really resonated in the ecosystem and the leaders of many of the other top agent frameworks came to us and asked us to build the same for their ecosystems.

So and obviously we're very glad to do that. so we did then a partnership with curi kind of building the same infrastructure and then we learned two things one almost all the work that on our side was shared across these frameworks which we didn't know going in and two if we're going to go one by one by one to all the kind of top agent frameworks out there it was going to take us a year and a half to serve everybody we need to get to which is what led us to build this open protocol mention right AGY which stands for the agent user interaction protocol and basically brought this layer from under the hood to over the hood. kind of cleaned it up. made it an open protocol that anybody can build towards. and just less than two months since launch almost every single one of the top agent frameworks now has a first party integration with Agy and kind of pointing to Agy and Copek is the way to bring the agents into use interactive agents.

So I would say like on that you know we kind of zoom out now. What do you use in the stack? Obviously the very bottom you have the LLM APIs.

Oh yeah. and lots of great options there. Then you have the Asian orchestration layer. and there's a lot of options there. You can use no framework, you can use one of the top frameworks,

lang chain, langraph,

mastra, agno, llama index, paidantic. there's there's a I'm sure I'm missing somebody. I'm sorry. But and then at the top you have the userfacing agentic applications, right? Which is both an application and an agent. and copekit we strive to be kind of the best glue layer between the application.

But there also would be like MCP for the agent to interact with software.

Yeah.

There also would be like me zero for memory something like that. There there would be something for evaluations that companies would use. So there's a bunch of different technologies.

That's right. Yeah. Every one of these layers you can decompose into the the different building blocks that make it. so eval is extremely important layer. memory extremely important layer. MCP you mentioned is really interesting, right? So kind of the way we position AGI is kind of this

MCP to people.

Yeah. Exactly. Right. So MCP that's agents to tools and context. A2A from Google that's agents to other agents. Then AGI that's agents to users through userfacing applications. So yeah. Exactly.

Mhm. So how do people use copilot kit mostly now? So you're you're getting a lot of success, a lot of traction people.

Yeah. do do you see a lot of forks or people just take it as it is?

most people take it as it is. There are a lot of forks but kind of you know relative uhly it's it's but a lot of these forks are from extremely high quality important

contributions and then kind of merge things back into the the original workstream and especially now with AGUI we're seeing really a lot of that. Yeah. So so that the use cases so I mean by now we really we're we're the leading solution for agent user interactivity by growing margin right. So we have many tens of thousands of installs every week, over 22,000 GitHub stars, and all of these are growing at, you know, pretty high rates. and by now we're working with, you know, all the way from fast growing, you know, Silicon Valley unicorns to Fortune 100s and really kind of everything in between. And the two main use case categories that kind of dominate is one is the what we call the SAS copilot. So you have an existing SAS application and you now we want to build an agentic interface for it, right? So instead of find the right menu, the right button, just say what you want to do.

Say what you want to learn and the agent helps you do that or learn that. And the second big category is what we call the productivity cop. I'm still looking for a good name for that. can maybe call micro aents. Somebody pitched me yesterday called micro aents. And that's you have a business function that maybe in a perfect world you want to fully automate, right?

maybe you have a thousand employees doing this right now all day long. That's what that's all they do. This one one type of thing. But the technology is not there for full automation. It's clear. It's not nonsense. It's not, you know, generating. Yeah.

May maybe they're not not in the technology. Maybe they don't have enough structured data. Maybe they don't have enough, you know, maybe they don't have a single source of truth. Maybe the process is messy.

Mhm.

But yeah.

Yeah, that's an interesting point we can maybe come back to with the data because I think that's kind of our second act that we're seeing in the future there, but we can get to that. but yeah, but the end result is the if you want the agent to just do the whole thing, it's producing something that is amazing. You're looking at it is a this is amazing, right? Is this for $1 instead of for you know two hours of work. but you can't press send on it. You can't just use it as is. and so the solution is build this kind of like a cursor for anything, right? that the the user, the expert is working alongside this agent. the agent is better at some things, the user is better at some things, and together now you've kind of 10xed this person. So, so that's

like to to bring an example into this, you you take like your best customer support agent maybe who's a lead team lead or a manager

and you empower them to like do a work of 100 customer support agents where like they become manager of agents using Kapila kit.

Yeah, it can be manager of agents. It can also be I think that's an interesting UX question. What does that UX look like? we talked a little bit about this you know offline

is it single threaded not I think there's interesting discussions there but

but the end result is you see a clear lift right and we're now at the point where these companies are sharing with us hey look we put this in production it's being used hundred thousands of times the last few months and we're seeing a clear lift in productivity and I think we're just getting really started in this type of thing it's only still the early adopters the early companies but definitely the future of software and very likely you know the future of productivity unless you AGI comes next year then then we're out of business.

We still don't know what what is AGI because you know no I I I think the first big thing AGI would have to do is persuade all of the scholars and researchers that it is AGI and the definition of AGI is correct

but something is coming for sure.

Yeah. Well, you know benchmark they keep coming up with new benchmarks and then they keep breaking benchmarks because like people labs end up tuning their their networks to be good at benchmarks.

Yeah. some benchmark hacking is

yeah speaking about this usually for like choosing open open source strategy is a hard decision and what are your metrics are you keep your eyes you're looking for

yeah so the like I said the most important battle for us to win in the next two years which can't be fought in the same way in two years from now is the mind share and the in measurement in the ecosystem right and you're the default not when you say you're the default, but when everybody else is pointing at you and saying you're the you're the default, if you're using my part of the ecosystem, go use copeket to bridge bring it to user interactivity. So that's the number one thing we're going after and and the metrics there are they're indirect, right? but they're pretty good. So installs usage, right? which we don't see as open source detailed but you see that just topline number it's opt outable but a lot of people don't opt out and yeah so and you know GitHub stars it's kind of everybody's like oh it's a vanity metric sometimes you hear we've seen pretty good correlation between that and everything else I think if you just optimize for that then can be any metric becomes meaningless

GitHub stores are much better than Google review stars

because engineers are usually way more conservative and praising something.

Yeah. Yeah. but but but yeah, generally I think installs and usages are kind of topline metrics.

So now that we kind of deconstructed what like an open-source stack of technologies for an agent might look like and again I want to deconstruct this further

for our less technical audience. basically explain what the concept of open source software and how do you make a business on it? Just say any advice for people who are looking in into this direction

because yeah a lot of a lot of people are like hearing oh open source means free so it's not a business. How do you how do you do a business on something that is free to use and everybody can you know copy it.

Yeah. So it's even more hardcore than free to use right. It's you're actually giving the thing away. They can run it on their own software. They can modify it.

Yeah. They can run it in their basement.

Yes. They and most of our users do right now honestly right. which actually is a huge advantage for a young startup because the fortune 100s are not going to trust a young startup to host all the user activity data that's flowing through these systems. So I think that's a huge advantage for startups but yeah so the way you make a business right this has been done many times by now so there's kind of a playbook you offer the base technology as open source and then you offer an additional layer around it that is geared for enterprises and that's where you monetize and in principle anybody can go build that additional layer because it's open source but you are kind of the default right you're the the you have the mind share you have the funding so if somebody else is really doing a better job than you building the enterprise there.

Yeah, exactly. And you can evolve it with the technology, right? If you want to build some feature in the enterprise, you can just build whatever it needs in the open source side. so I think if you if you're not the one building the enterprise solution, you kind of screw it up. But it is something that's really important to keep in mind, right? You have to build these things at the same time, right? So you where you monetize in the end of the day large enterprises that signal is actually a little bit invisible from the signals we talked about before because they're small number of of of important users. So I think it's it is important for open source businesses to consider the enterprise and build for the enterprise early on not even just for revenue just so that you have that signal to evolve the product.

So it's not just good for the community but also for these businesses that have different needs from

and the signal also loops back because like once your ecosystem partners hear that these like five Fortune 100 companies use you they're like they might be the standard.

Yeah. No, that's true. And it definitely loops back kind of exponentially. the more the better you do, the easier it gets to some extent. Yeah, we're seeing that at this point. Again, we have a lot of the just the top companies out there, period, that are somewhere in our stack.

There's a lot of great examples of open source technologies that have become beautiful, valuable businesses. well, Red Hat is probably the most expensive one

acquired for what 34 billion I think some something crazy. But then click house that originated within Yandex as a OLAP database that they just like open sourced and then all of a sudden half of the world was using it. So they decided to build a business is now a 5 billion company and Gen X is what probably 70% of consumer websites are built on is acquired for two billion. so there's clearly a very large market for that. But who buys open source companies and why? So, I mean, who buys the companies? That's a separate question, I suppose. because it's that's not necessarily the destination you're going for, but it obviously could be. I think in our world, I guess there's there's two answers here, right? first, I do think that there's a lot of large companies who could want to own something like what we're building that have no organizational capability to build probably even either the product, but also just the whole ecosystem around it.

So, that could be a good insurance policy. it's not what we're going for but I think you know if you look at most of the cloud providers most of the you know model providers even the IBMs and the Cisco of the world I think could could find a lot of value in in that again that's kind of more of a insurance policy and and I think they the ecosystem love and inshment and mind share is one of the number one things that are very difficult for a lot of these kind of larger companies to build of course with AI that's a little bit different the model companies they have a lot on their mind anyways but yeah if you look at the IBM's the world that that's a huge advantage but I think if you look where this going right so there's you always kind of try to what's the right pattern match and everything is a little bit different but there's still some pattern matches we think of ourselves as kind of similar to the versel of the AI world to some extent so vers you know nextjs is the backend for the application front end copit is the back for the agentic front end.

And so monetization wise, you know, Versell is, I think it took him some time, but they're starting to become, you know, pretty profitable business. And I think there's parallels for everything that Versell is doing in our world. So that's kind of our pillar one of monetization is, you know, just the hosting, the CI/CD, the analytics. So we have, you know, hosting CI/CD instead of application.

Let's deconstruct that. So hosting is when people don't want to run your code in their garage. They want like this done as a service and they pay you.

Y

so you get like a cloud and you build all the security around it and you use them let them use it as a hosted solution. Y

what's CI/CD?

CI/CD is you you have a lot of developers eventually building alongside this product. Each one of them is making some tweaks and they want to see what what is working in their own development environment that runs against everything else and then when they merge it to the main branch you want that to just become the thing that serves everybody else. so that's what's called CI/CD you know continuous integration and deployment and so you can do the exact same thing for the agentic front end right so people developers they just want to commit their code push it and everything works they don't want to

glue all these pieces together

okay so and they pay pay you for that

right

okay in terms of percent revenue what do you what do you mostly bet on what do you mostly bet on like what's going to be the lion share of the revenue

yeah so our actually larger bet is on the second pillar that is not even the VSL type parallel. So that first pillar is you know kind of do the what's done for the application agentic front end with the agentic front end. the second pillar is really a data play. so you mentioned data before right you know reinforcement learning is what took GPT 3.5 to chach right and what made the coding models really good at coding and what now they know that last weekend we had the math olympiad result right so it's a really important layer that so far is really restricted to just the hyper hyperscalers the model the base model layer for the most part but there's there's a market out there for for data right so if you look at scalei right what what are they last 15 billion or what was it?

Some

so well 30 something billion valuation per company. So they paid 15 for half of the company.

Twice what I thought.

Yeah.

One of the most valuable data sets you can have is here's what the I generated here is the one sentence that I want to change into the other sentence. and the going market right now for this type of data is on the order of dollars per edit. Guess what? If you have thousands of people using co-pilots every day, all day long, you're getting this mo all the time, that's literally what is happening passively,

especially if it's expert data.

It's and it's yeah, it's expert. It's kind of doubly expert, right? It's on the one side, you have product managers in the whole organizations that are optimizing. Okay, where should we speed up our engineers? What's the most important valuable thing for non-engineers, sorry, our employees to do? and that's where the application identic application is born right and you have whole teams just working on that and then the individuals actually working that they are the experts that's their job to generate whatever content that is and they're extremely well motivated to make good content because that is their job like it's

so they're not labelers fixing AI output they're they're just doing their job and in the process you're now get this extremely valuable data that right now is just being thrown away by yeah

it is incredibly ly valuable. So we Marin and I we we were looking at tens maybe hundreds now of different unsuccessful applications of a AI agents towards different business functions

and we coined three main reasons people fail at implementing agents. The first one is they don't have a single source of truth. So that's just you know also a data problem but the different angle of that is you know they are trying to automate their sales or sales intelligence but their sales people do not really use their CRM system they have a separate booking system

it's a context problem

yeah so that's a context problem so AI doesn't know where to look for the most recent up-to-date data the data might be contradicting itself it's like this is why all of the coding agents are so good because codebase is the single source of truth it cannot contradict itself The second biggest problem is rules. So if the process does not have any rules that define it, let's say it's brand marketing. There are no rules in brand marketing. It's good to break rules.

So a lot of times the agent doesn't quite know what is expected. So it's I'd say it's a prompting problem but also somewhat a context problem. But the third largest problem that we've seen that nobody can really solve unless they've been very careful about their processes way before back in the day is having a throw like a lot of data on how are these things solved in general. what is like if this you know let's say we're making investment decisions we're in decision-m business and a lot of times we would look at a company and say that's a pass a lot of times we says like I want to investigate further a lot of times I'd say that's a that's a no-brainer we'll write a check what was in our head when we were making this decision like how did we come up to you know that conclusion when when when we talk to experts Like let's say we're looking at an open source company and we speak to one of our LPs who's who's built playright.

It was like can you look at this and say hey is it is it is it going to be good and they and they look at it and it's just like and they shrug but we have trained our natural neural networks to interpret the shrug in like three four different ways. Now, since we haven't been documenting all this,

yeah,

there's nothing to put into an agent to enable it to autonomously make the decisions or even support these decisions. So, what I think is incredible value that Capilot kit is is bringing or is going to bring is that you can decipher human interactions and human decisions and maybe in the like in the moment ask the person, can you explain this? There's a very good example of this implementation in Chinese courts. So I don't know if you've heard about it but China has introduced AI ruling into their courts about a year ago.

I did not know that.

We don't know to to to which extent but like as a

so now every judge in China has an AI co-pilot that tells them that what like decision should be. The judge can overrule it and say no I think this is what it's right. But then they are obliged to write like a page long explanation to AI for a LHF.

Yeah, I mean that's that's first it's incredible. I mean and I think as long as there is well just thinking out loud as long as the human has the overriding capacity that seems like a

a good idea

net positive I suppose although I'm sure there's biases that the result can introduce and

so I don't know maybe think about it some

if you try and asking it a bunch of questions that you know ch the party doesn't like

right but it could even bias the judges opinion right they first saw some answer

it's like maybe that's what I should do larger your point before, right? So the the the context, the prompting, I think we are actually operating on all of all three of these even before the data, right? So one of the reasons you want user interactive agents is because of the problems you mentioned, right? That there's context all over the place. It can come either at build time, right, but then there's only so much context that you can do at build time because this agent is going to run for thousands of cases, maybe more, right? So could be at runtime and then the ang has to go and look for the data but it's not really an intelligence problem like you take the smartest person in the world drop them in a new job ask them a question that everybody in the job can answer they they don't know what to do so I think this has to come this tribal knowledge has to come from the user so right now phase one is enable the agent user interactivity right allow the agent to reach out to the user allow the user to see what context they have the agent already has change it, augment it.

same with the prompts and the nudges for what to actually do. So that's where we are right now. Every time the user is doing that, that is generating the data that can be used for the second part. And right now we're just throwing it away because you can you can't boil the ocean, right? So so you facilitate the agent use interactivity that is already making these agents turning them from very cool to actually useful today in production. But then now you're kind of passively collecting data that if you wanted to buy you would be paying a lot of money for. Of course as the scale grows that it's not going to stay the same price per data label. but you know what Scali did for you know three or four or one a handful of of companies right? You can have now for 30,000 or 300,000 maybe every single application or micro agent can have its own RHF loop that is kind of improving over time.

So I think that that is an exciting it's a little bit speculative but a lot of AI is a little bit speculative. and that's something

the data markets are changing fast. So we spoke to Toka about this which is like a competitor to Scalai

and what they said is like 10 years ago they were you know paying like a dollar per maybe 20 pics for people to label sneakers on a picture and then they started paying five bucks per 20 pictures for people to tell if it's new sneakers or used sneakers. Mhm.

now they're like paying hundred bucks per like a page of text reviewed by an expert financial professional or like a lawyer.

Mhm.

And and and then they're also doing a lot of synthetic data based on that. So I think I think the more specialization AI gets like the the more advanced data has to be and the more hands-on it has to be. So you are really unlocking that. I really like how Andre Karpathy put put it in in his talk and by the way highly recommend it. Watch Andre Karpathy Hawaii combinator AI school. So he said is imagine like if you're bringing an LLM into your business imagine that you're hiring a multiple PhD professional who is incredible but they have one problem. They have their memory wiped every five minutes.

Mhm. So they need to learn how to navigate in your organization. They need to learn how like different things are solved and for that they will need capilot bit capil kit for users to be able to tell them those things when it matters and and they need some sort of a memory to store this and then has to be efficient. So there's there's a lot of

a lot of things. So I think it it makes total total sense. But let's maybe shift gears and talk a little bit less about technology and a little bit more about your company because we wanted to deconstruct your business.

Yeah.

As a like what is Capiler Kit as a company

and I am wondering about that like do you think it's different to have like is a team working on open source product is different from from other teams?

That's a it's a really good question. I think the answer is yes. and even internally right the way right now we're too small to really have three teams. So I say we have the shadow of three teams right now and people kind of overlap but generally speaking we're thinking of on the product side we have kind of the ecosystem in meshment right and that's open source. You really want to be part of the ecosystem and and do these kind of win-win games across the ecosystem. And then you have enterprise is the second kind of big pillar and you're it's a totally different mindset. you're building for large businesses. They don't care about traction and they don't care. They they care about the one problem that they have. It's oftentimes not a sexy technology problem. It's some, you know, boring permissions problem. and you need yeah different mindset.

and the the open source side, the community play is extremely important, right? You want developers who can engage with other top developers across the ecosystem as as peers as equals and and and we're lucky to have these types of folks. Yeah. In the enterprise, you want, you know, professional conduct and talk to the product managers on the other side and help them fix the problems. And yeah, these are different mindsets. So, I think both the companies and even the teams have have different mindsets.

So, what does compiler kit look like a company? How many people you got?

Yeah, I mean we're still pretty small, right? We're 1213. We have some external kind of contractors. depends how you count as well. but yeah, we're 12. We have, you know, me and my co-founder who's also my brother. I'm CEO. He's head of growth. we have a Devril community engineer. so that's help manage all the tutorials which is kind of the best converting content in our world. How do you build something? And now we went recently from having you know one partner to two partners to 10 partners in a couple months. So have to have a bit of a you know mass production motion around that. We have kind of a go to market person. So that's you know take these.

So there's partners they they they they help you develop the product or they're like design partners.

I mean they're in many cases building a big part of it on their side. Right. So when we initially did cogent integration with Langaf and then Crui, we did most of the work and they supported us. and you know that takes that time it took you know a few months per integration like that that's what we realized okay you can't repeat this at the timeline that we wanted to repeat this agi right the the protocol agent user interaction protocol factors out a piece of this cleaned it up we did some time on just that but now these partners it's the other way around they're actually doing the integration they're doing the major building and we're supporting them when they need help so so that's if we're looking at the agent framework side right there's other types of partners ers as well. right the kind of MongoDBs or there's always something you can do together.

But if we look at the agent framework side, it's really their engineers are building with us or building against the protocol and we're supporting them. we also have some things sim similar with like the larger companies at AWS and Google had yeah a few threads open with each of these and yeah it's a lot of fun. and it's you really start to see the open source kind of community standardization paying exponential dividends to some extent.

A lot of times when people hear open source community developed products, they think of you know an engineer in his mom's u basement in Bulgaria just like writing a few lines of code and you know making something better.

Yeah. But that a lot of times, most of the times it's a large company that relies on your product to sell their product.

Yes.

And they are coming to you and saying, "Hey, that would be cool like if we could build this feature for our users. Can you support this and like include this in like in in in the base, right?"

Yeah. Yeah. Exactly. So that's that's the stuff that we start to see even more of when we went from the just the Yeah. the AGI, right? because I think it was really there was a big gap there. The reason we wanted to do this quickly too is because they're all telling us, hey, we need this like now, right? And this vacuum, if it's a strong vacuum, it's not going to stay a vacuum for long. So, so that's what the reason it was important for us to move that we've seen it actually follow up, right? We we put this out and there's been a ton of of traction including for these really large organizations, right?

that they're trying to solve the same problem in their own context. Maybe they have their own internal agent framework. Maybe they're using other systems, but they have the same problem. And so now they're joining the ecosystem, not necessarily on the commercial side right now, but now they're part of the ecosystem. They're contributing to it in major ways to actually a Google ADK integration that came from one of those really large companies that at the very top technical leadership of the company. there was a really strong interest for that and and they're doing it. They're adding amazing value to the ecosystem. you know, a relationship that can grow in in interesting ways as as we grow

as a company.

From one point of view, you're like contributing to the world and it's a little bit harder to monetize than the proprietary technology, but at the same time, it opens like every door brings you immense interoperability.

Yeah.

Your technology works with everything there is because everything there is wants your technology to work with it and they build it. So in a way you don't have 12 people on the team, you have 100

or even more, right? And it's both the teams and where they sit, right? You can't buy the employee at the large company who absolutely

somebody with like a 2 million a year comp package sitting at, you know, Google or Microsoft that are very genuinely interested to make this thing work with their stuff.

Yeah. and and not only the monetary side like the fact that they're they're there, right? So they're controlling and they're kind of influencing where this large organization is spending all of the other engineers work, right? So if you now set up the foundation of thousands of engineers in your organization, that's something that's going to the there's a forcing function that grows with time. and so yeah, it's definitely delayed. You have to delay. We're thinking of, you know, be future greedy, right? It's a business to be honest. I mean, we're I'm very open source pro in general, but it's a it is a business, right? We're not a charity. So, there has to be a business case that makes that work. and I think in some ways it's kind of lucky that it does work, right? You can if you just play these kind of win-win games, it does open enough opportunities that you can monetize.

Yes. It's delayed because initially you're providing all the value for free.

Mhm. you can build additional value that is not part of the open source and and monetize there over time and yeah like you said there's a lot of very successful examples for this. So yeah.

So going back to the company, so you you got GTM people that

one

okay one person who like speaks to enterprise customers, figures out what they need like how to how to make money on all that,

right?

Who who else you got?

so the rest of it is engineers, right? So so

well what what kind?

we have I would say two personas maybe three personas. the developer infrastructure kind of hardcore we call it innovation leadership type role. So that's the architect mindset right and

like focus on infrastructure.

Yeah. Yeah. infrastructure and really can make a new protocol that is going to be used by across the whole ecosystem for something that is quick fastm moving when actually for good or for bad we're in a pretty complex part of the stack because we're the glue between

a lot of other components that are themselves doing something new and moving very quickly on their own. So it's a pretty complex glue layer to to to be in and so and you need to make clean abstractions and also be able to build them and develop them quickly. So we have that persona. a few again we're a small team but we have a couple a few of those.

we have kind of the delivery lead persona and that's okay you build something you have to make it clean and production ready and that's that's a different mindset and mentality and then you kind of have the the jack of all trades which I think is really important for a startup at our stage that can take the front end and the back end and you know the enterprise and the open source and like figure out how to make it all work and if you don't know something just you you learn it and then you build it. So that's kind of the three personas that we have.

What's your take on remote work?

well, we are remote.

Okay.

and for for good or for bad, it's definitely there's advantages and disadvantages for sure. On the advantages side, you just open the pool of talent a lot. I think that's the main advantage. It's not really about cost. So actually I I worked at Doximity which is healthc. when I joined was kind of a scrappy start on a multi-billion public company. So learned a lot there. but anyways they were remote before co or partially remote. I was in person in San Francisco office but a third of the company was remote. And one of the things I learned there is yeah it's not a cost-saving effort because you need to spend a lot of money to bring people together a few times a year both for morale for productivity. you know, you zoom too much, people start to dislike each other and then you have to bring people together to to kind of

usually you do when you're together.

We've had a few of these off sites now. We basically all get together for about a week somewhere and both do it's very intensive. It's kind of morning to evening most days. It's a lot of building together and discussing and planning the next quarter or two quarters. and also having some fun like doing some things that like you don't normally do. People every single time build a lot of camaraderie. like that's everybody leaves kind of on a high high note. and yeah, it's it's it kind of buys you the fuel to to to go back on Zoom for a few months and and make that work. And honestly, actually now that we're also growing a little bit bigger and these off sites, you know, we're still small depend everything is relative, right? But I'm also making budget available for folks who just want to get together for a week

to build something together cuz they think like there's a project that needs to happen.

so people can just do that spontaneously and you don't have to pull the whole you know team together that's kind of becoming a big bigger operation. so I think that's really important. So that's the advantage is you grow the talent pool. You have really really really excellent people. I would say that there's probably you know our engineers are kind of the best engineers in their part of the world in many cases or like the very best the disadvantage is a lot of things are default easy with in person work become default hard that you can make them work but you have to work for it process becomes a lot more important we've iterated a lot we've now finally arrived at something I think that is seems to have seems to work well enough that we're not iterating too much. But yeah, there's a lot of things that you have to they don't just by default happen the way you want them to.

You can make them work, but you have to put the attention in the time.

How much is I I don't know, Marina, maybe you want to take over. I just wanted to speak a little bit about the fundraising part. So how how how is it different to raise money for an open source company and what like is a different type of investors you have to speak to?

Yeah, it's a good question. I mean it's definitely not every investor and I think for for good actually to some extent

what you are looking for from

right exactly so I there's you want people who understand the open- source motion and what the play is because for a lot of you know it is different from traditional businesses right where you're giving away the the thing like what are you doing right like you aren't your business you're trying to make money why are you giving it away for free of course if you take a long time horizon then these things have worked out very well in the past and you can't do it the other way it just doesn't work but not every investor understands that yeah I think you want people who who understand the journey and that for us I mean you know we're lucky enough to have the ability to to choose when we raise money and and I'm very glad we have the we have signal fire as our lead they've been really fantastic for us and you guys you want people who will challenge you on what should be challenged and not on what should not be challenged.

Oh yeah.

Yeah. And that's not every person by default. You have to really understand the technology space.

How did you maybe the last question for me? How did you pick a license? Cuz you can go you know copy copy left you can go Apache MIT.

Yeah. honestly look we spent like I don't know a few days thinking about it in the beginning then we just continued on with that. so I can't tell you that it's necessarily the optimal but we have MIT license. It seemed like a good default. it's a pretty maximally open license. The reason we did that right is the war to win now is the mind share and the community love. and anything that gets in the way of that could become could come back and hurt you maybe later. But it's it's premature I think to to you can only you know control and and have attention for so many things. So that's not the you know battle we chose to to fight.

all right

and and also yeah and and I think it does make sense longterm and also kind of fits with what I would want to have in out there in the ecosystem.

That is good because we are it's it's actually a very important question for us. We are thinking about open sourcing a bunch of our stuff that we've built internally in in the fund. Not only not only the software but also our pro processes, playbooks.

Oh, playbooks. Yeah. Like

Yeah. And that was a very meaningful conversation that we had with Charles Grinton who is one of the co-founders of Mistro.

We were like never ready. We're like, we we still have to do a lot of stuff. And he's like, anything you can open source now that is immediately helpful to somebody that's something that they can spend maybe 10 minutes of their time to use it and get some value out of it. Just go and open source it now. but that's that's going to be interesting like we're going to pick a license. We're going to you know start trying to build a community. So we'll we'll definitely ask you a lot of questions.

Yeah. So that's and the whole idea of this podcast is is to learn from you from your experience and to be inspired by you. Probably you can recommend us and our audience anything books something to listen something to watch something you like personally.

Yeah, good question.

Maybe some even fiction books.

But you know maybe even before I go to that like because I think what you said before really resonates with me too. I think there's a lot of lessons I learned in the open source space over the last couple years that I think should be applied way outside of where they're currently being applied. So I think that's a really good trend to see.

Can can you share this lessons?

yeah look I think the community so I mean first what does a community look like? understanding the adoption curve really explicitly,

right? Which is something that u startups do in general, not just open source, but in open source you see that even more, right? So the early adopters, it's a totally different core you were aiming for, right? So the we didn't have docs in the beginning. We were aiming on purpose. We're aiming for the folks who will just go in the source code and figure out how to use it. And then you had bad docs and then you have slightly good better docs. And there's this adoption curve that you have to you can explicitly optimize for like you you should not be spending your time initially on good docs because it's only the most hardcore early adopters are going to care anyways in the early days. you can't really skip stages there. So so I think that mindset having the different cohorts of people right you have the users who who are sophisticated and really really large group and then you have the contributors which are even more sophisticated in a smaller group and they're just different personas and these are all parts of a community that you want to engage with on different terms.

and the fact even just the organization force of a GitHub repo I think is incredibly incredibly powerful. and we are even going to be doing even more on that side because you have all this attention that comes in and a lot of it wants to be useful somehow but it it's not you know in a non-opensource world and again it can just doesn't have to be product and technology anything like you know from I don't know parent groups on Facebook or something to academia to like I think lots of different parts of the world people can't by default be useful like what are what am I going to do? I want to I want to do something but how can I do it? If I did something how can I bring it back into the larger group. so I think that's that idea of a GitHub repo with open source type mentality and management can really extend far beyond where it's currently being used.

So so that's and I know that whatever I'm going to do in my life I'm going to kind of carry a lot of these lessons to that kind of Yeah. because I really think it has

oh yeah

application way outside of of just technology businesses but yeah in terms of books to recommend it's a good question I mean there's so many things going on that you kind of have to be to stay with the latest for the latest but I think the fundamentals maybe become more important to some extent like because you really need to be in touch with that my background before doing startups was in physics got a bachelor's and masters in physics And I still kind of like to to deep dive into that. So I don't know, maybe it's a bit of a off color recommendation, but

it's a good one.

Theories of everything podcast lately is did a really great job. If Kurt is watching this, good job. Yeah. So I'll stick to that one.

Actually, more and more talking about we start talking more about physics. Oh yeah,

we by the way if if you guys are watching and you haven't looked at or talk with

Adam Brown,

Adam Brown, he's a theoretical physi physicist at Stanford. He's a deep mind doing quantum and doing AI.

we had a like an amazing talk to him at our conference.

is it on the podcast?

Yeah, it's it's Yeah, it's on on on YouTube as well.

Yeah, I'm a big fan of podcasts. I think podcast in general that's actually where we started very initially the LM media database that was the first application so everything loops back

that's the best way to learn cuz you're speaking to people who have spent the time and effort to learn themselves and they're giving you the gest and right now like yes it's not going to allow you to become experts but I don't think we need to be experts now giving you know AI we need to be general less and we need to figure out things how they work very quickly

and learn quickly.

But yeah, I'm I'm so happy that you know we we have you on this podcast and that like I really want to persuade. I want to inspire a lot of a lot of people thinking about you know building a startup or actually building a startup right now to go for an open-source. it is it's not only a go to market strategy it's it's also a philosophy and somebody was telling me that and I do agree with this is is that if the market crash if the internet bubble happened today and a lot of people are like oh AI is a bubble AI is going to you know collapse it's going to be horrible and we're like all all going to lose money and die the biggest difference from 2001 and today. It's not, you know, revenue growth or anything or how many people use the things.

It's open source. Opensource was not Yeah. was not was not a thing 25 years ago.

It is a thing now. And the reason why open source is so important is that open-source has become the way for people to create technologies so quickly and so more powerful that we're not selling like 20% increase for a lot of money. We're selling 100x increase.

Yeah.

And that is just propelling our society towards exponent.

Y

and a lot of times Yes. People are not making as much money as they would with proprietary stuff, but because it just becomes so big, in the end, they're going to make a shitload of money.

Yeah. Yeah. It's it's kind of this virtuous lucky cycle, right? So, yeah, it's less monetization per, but the pool grows so much bigger that you're still doing really well. And now, and also you're contributing to the rest of the ecosystem. And and I think yeah, we've gotten you know in the internet I think that's a really good point right the post internet world friction is really really important input into how well something is going to do right is it exponential

growth or exponential decay and when you can nudge things to the exponential growth side it's it's it works it's good for the ecosystem and it's good for you and it's it's this win-win. Yeah, because like if we look at you know the dotcom bubble web van or like pets.com or all these companies that collapsed what happened to them because people were like oh we're overpaying that's not enough value which is going to switch to something else and and they're switching to save money and these companies just go out of business.

Mh.

If let's say all of these companies run out of funding from VCs will companies stop using it? No. Because people look at open source stack and they're like we are 100% dependent on that stack.

We're not paying like anything noticeable for that. We will keep paying and this thing will persevere. This will keep growing. So I mean like there it is just the abs absolute insurance from people changing their sentiment. It's like if if people say hey we don't want to fund robotics or we don't want to fund biotech or we don't want to fund you know something like deep tech like it's called fusion we don't want to we don't want to maybe invest like a billion dollars in into quantum everybody still needs open source stuff because actual real life businesses absolutely depend on it

and then open source saves the Yeah. No, I agree with you. And and it's it buys the insurance for these initial customers that allows the company to grow to the point maybe we don't even need to apply this insurance, but it's insurance is necessary to to take the initial leap. so yeah, it's it's definitely a very kind of virtuous cycle. and

yeah, maybe a business in I don't know, invention or discovery in its own right.

Yeah. do you want to cover something?

Yeah, I think we covered.

Yeah, I think I think that was beautiful. So, I want to probably do like maybe 10 15 minutes of explanation for like less. so I I think the way we we we I don't want to disrupt the flow because it was like high energetic but there's a lot of things that you mentioned that are very hard to understand for people who are less technical

and we want this to be able to to understand. So, I'm thinking, do you guys think it's it's going to be better for us to just put text explainers so people would just like pause and read and then pause and read or maybe

I don't know. I always just listen to podcasts, you know. So,

yeah. That Do you have any materials like

Yeah. Yeah, we have materials. I mean, we could I don't know if maybe that's kind of where you're going, right? could do like a thumbnail that instead of a could do a text only explanation too, but could be like a small clip that just talks about the thing.

Yeah.

Yeah.

But somebody who hasn't heard about, you know, open source tech before, hasn't heard about, you know, a Gent,

right?

they would need to try to dive in.

Yep. and and understand and a lot of like we do want business people to understand that too.

We don't want you know we don't want just engineers to read that as like oh yeah I know capella kit that's good right

because

we could do like yeah maybe like just a Q&A type session at the end that people can like hey deep dive into some of these terms if you don't know them.

So you you said devrils that you have developer relations why do you why do you need them? Yeah. So developer relations is the bridge between the community and the engineering team, right? So the engineering team is not out there talking to the users of the software every single day. They're not telling they're not out there evangelizing the software every day. So that the developer relations is about essentially you can think of it as marketing for developers really. That's kind of the the simplest way to understand it. but good marketing for developers speaks to developers really on their terms, right? So the content you have is tutorials that developers can go through and build whatever they're trying to build with your stack as part of it. It's not just your stack. It's the whole the whole thing together and you're a component there. it's about giving support to to developers and then taking this the input from support and bringing that in an actionable way to the team so that we can build whatever is missing, whatever people want, whatever bugs people are hitting.

So and then it's also just speaking about it in events and giving demos. It's actually one of the hardest by far roles to hire for. because you need both a technical mindset. like you need to be a good engineer actually to be a good dev row and you have to be interesting to to listen to and charismatic and kind of you want to do public speaking. These things tend to for whatever reason filter each other out for the most part, right? Most people who are technical engineers, they don't want anything to do with public speaking. or they're not interesting even if they do want to and vice versa. So like it's it's a really kind of this intersection of two different talent pools that that is can be a huge force multiplier for everything that you do. because just because you build something doesn't mean people know about it but it's hard to find but it's worth it.

Another question that I wanted to explain. So so open source has a gazillion different licenses.

Yeah. And we we kind of covered like how you chose the one that you want to do. Which ones would you find, you know, the most the most common for aentic agentic technologies?

So maybe even to take a step back before the license, right? Open source what it means is it's about open-source code is kind of I think that's where the word probably comes from. Now source means a few other things as well, but you know, you're building a product, there's source code. somebody wrote the code that makes the product. and open source means you're putting that out there publicly. So, not just the product, it's not premium. It's it's not free. It's not premium. It's different from that. It's the thing that we the company is using to make this product available. We're putting that out there. and so you can put it out there and that's, you know, GitHub probably everybody's familiar with at this point. where it lives and you can view it. but just because it's out there visible doesn't mean you're giving everybody also the right

to do whatever they want with it.

so there's different licenses that have become common over the last you know couple decades I suppose. we are as I said MIT license which means take it essentially you know I'm not a lawyer don't take it as legal advice. but do whatever you want. you can copy it, you can modify it, you can do whatever you want with it. there's more restrictive licenses licenses that let you do whatever you want with it as long as you keep the same type of license on your thing. So you cannot take it and build your own.

Yeah.

and there's other licenses that have evolved over the last few years that say everybody can use it except for really large companies. If you're over whatever a billion dollars in revenue, you have to pay you have to pay something. So there's kind of this mix of licenses or licenses and like I said I'm really and honestly not an expert on what is the most optimal strategy. We kind of just went with a sane default of just keep it simple.

Yeah. So basically anything that allows people to understand how it works, use it for their own project or their company and then if they need to change it or they want to maybe resell it, then there are different licenses that are built specifically for that use case.

But don't be too creative with licenses otherwise like it will ruin your open source strategy.

Yeah. Yeah. Can you maybe explain? so we kind of covered how like what what people need to use to build something that u enables them but what what exactly AGUI does and and and like how how do people get value out of it?

Yeah, AGUI stands for the agent user interaction protocol. and that's the layer that we factored from kind of the insides of Copekit and then we cleaned up, put it out there as a its own protocol that people can build towards. it's a lightweight open eventbased protocol. which means as the agent is running, it's emitting one of 16 standard events that are compatible with AGUI. the agent can also accept one of a few standard inputs from the user. And now you have input and output and you can facilitate rich user interactivity. So that's the protocol at a basic level. we built it to be kind of maximally interoperable by design. which means part of the protocol includes a middleware layer which makes it both. So I mentioned before the events just have to be compatible with AGI. They don't have to be exactly the same events because maybe you have your own naming schemes for events and so on.

So the middleware layer can convert events from one form to a equivalent form and it also makes the protocol transport layer agnostic which means the events can be emitted however you want. The middleware can do whatever you can write whatever code you want there. Most integrations just emit these events with SSCs which is one of the you know standard technologies for emitting serverside events. Mhm.

But it could be other technologies such as web hooks, which means the server is calling some other URL with events when things happen. and kind of everything in between. So that's the technical kind of lay of the land. Maybe it's a a bit on the technical side. Yeah, I think I think it's a it's a it's just to sum it up in more like in less technical terms, it's it's basically a standard protocol that allows agents to input like user what what users want to change in the in the process of of of of the AI making thoughts and so yeah solutions. Yeah, it's essentially you know agents, AI agents, they're just computer programs at the end of the day, right? but they are very strange computer programs compared to what we had before. they're long running. they're interruptible, they stream, they need kind of monitoring and nudging.

so so this is very different from what we had until two years ago where you have an API behind some REST endpoint. you make a call and you get an instant response and it's over. so so that's why there's a need for a different way to interact with these things versus just make a REST API call. And what it enables is the agent the the mental model it's actually really simple at the end of the day. The mental model is very similar to working alongside another type of person that or a person that is better at some things, worse at other things. It's much faster than you, that's for sure. but it can't do everything as well. you want to be able to nudge it in the right direction when it is not going the right direction. You want to easily give it context. You know, people talk about context engineering a lot these days.

which means give the agent what it needs to know to make the decision, right? Because like you said earlier forgets everything five minutes. A big part of context is almost a UI problem. How do you allow the end user to bring the right context at the right time to the agent? So yeah all of these things you know another interesting thing about these programs is they talk to the users in both structured and unstructured data at the same time right you have the structured here is the state that it's changing here's the tool calls that it it's calling the unstructured is the actual conversation right which now is mostly largely you know text based maybe soon voice maybe later video as well so so you have both of these at the same time and yeah all these different corner cases is add up to a lot of overhead that anybody building any type of user interactive agent needs to deal with.

There's just no way around it. It's kind of death by thousand cuts. and you know the V1entic applications just solved it themselves, right? That's kind of the most hardcore teams are the cursors of the world and so on. but if you think like we do that that's what all of software is going to look like very soon and maybe a big part of productivity period. then this needs to be repeated at really large scale and you don't want to resolve these problems again and again and again. You know, one thing we actually didn't even talk about before is, you know, Agui it's an instance of what people call an MN protocol. And what that means is it turns an MSN problem into an M plusN problem. Right? So you have M frameworks, let's say, right? So mentioned langu cury agno ag2 amindex and or your own framework if you're a large company or the no framework which is another kind of special case of this type of thing

you have n clients right so copekit we started with react web that's the main client we built for first but you want to expand to all the web platforms right angular is large it's for you know somewhat kind of more technology a little bit backwards businesses that are often very cash flush a little bit more older school technology

besides web you have mobile you have messaging actually we're in talks with one of the the really hypers scale is about a messaging based client there's a new text messaging protocol called RCS that Apple recently added to iOS Google has had it for a while already that let you almost have a application instead of a text message conversation so it's not just text in and out you can have buttons menus whatever so you you have all these different types of clients And naively you each one of these agent frameworks would have to build a client for each one of these environments. And with AGI now you just build okay each one of these has one integration to AGI which is maybe a couple weeks of work. and then each client you have a client for each environment with Agy and now the whole ecosystem gets this this kind of you know many to many crossover

right? So it saves a lot of time in making like agents universal

right

for like actual use cases that people need for their business.

Yeah. Yeah. Maybe you want to take a you know language agent and serve it through text messaging with all the human the loop interactions right for customer support in you know

in you know South America and Asia people do everything through messaging applications. So, yeah, that's one of the

Yeah, I think that this is a super important point, right? Well, congrats with with the launch and you know I I'd call it soft pivot,

right? because you're like your product's already very successful but now it seems like you are very close to becoming the actual standard for like universal technology and I hope that since the market is growing at an incredible speed 3 four years from now it's just just going to be billions of users using maybe not even realizing that they're using your technology but everybody's using it and no one can really you know do stuff without without the the tech that you're working on.

Yeah, I mean that's that's what we hope to do and I think that that agent human interface layer I think is is really important really kind of thick layer that we want to grow into more and more and more. and yeah that's that's our goal. So thanks so much for having me.

Super cool. Thank you very much. good. So I think we'll we'll have to figure out how to like reorganize all of the questions but I think we can we can do it.

Do you have any time frame when you want us to publish this?

what what's the relevant time scales here? Right. Is it it's

so from from couple of weeks to we can we can maybe hold it for like two months and then release it like depending on like when you want people to.

So we don't have you know too much audience but we can get usually like 20 30,000 people to to to to listen to it. A lot of so I'd say half are engineers and a lot of people would just like oh I want to workers founders

and founders like either I want to use it or I want to work for this company but also a lot of people who are just just generally trying to learn new things.

Yeah. I think September makes sense to me. I think that's

that's good.

Yeah.

Perfect. Yeah. And also like it's always good to to use this stuff as like something to to make fundraising better when you need like this what what's it called like I forgot the word but it's basically oh the echo echo chamber for you know

yeah cuz when

usually when when Marin and I we were fundraising for our startup so we tried to concentrate everything into like same two weeks. Yeah.

So all of the sudden somebody who starts like looking at our company all of a sudden they notice oh like everybody speaking about us

yeah I've seen us about

yeah some somebody published a review like from the customers somebody like somebody published like a media article that was like and like all of a sudden people realize that they've been missing something.

Yeah.

And they're like oh I want to write a check cuz you know you want people to feel thumb. 100%. I mean, we we're likely going to go raise before the end of the year.

Okay.

so September sounds like a good

It sounds like tentatively good. Also, look, it it's this type of old ages, you know, quickly, right? So, if we wait two months, then there could be a lot of other things that are

Oh, yeah, 100%.

so, yeah, so I think that's a good kind of middle ground. and yeah, we whenever we will go fundra, we're also going to do a yeah, San Francisco Blitz. Just try to be everywhere like you said. So,

yeah.

that's good.

Yeah, I think that's good. Good. You know, close enough to potentially be relevant for that. and if not, then I I wouldn't want to wait too long anyways because then could be a lot of other things going on that were

And again, we always want, you know, more people to to hear about Yeah. U the the company, more people to use it, more people to try it.

Yeah. Yeah. 100%.