JoyStream.ai Builds an Operating System for AI Agents

Hey everyone. Welcome back here to Techstrong TV. I don't know if it's because I've been around this long or not, but I enjoy seeing my friends start new ventures.

Right? And I have a lot of friends like me. I don't know if it's a question of if we all have ADD, or we're just serial entrepreneurs, we're born to be serial entrepreneurs, and we like starting new things.

ai, and I couldn't wait to hear all about it, so I reached out to Harpreet, and here we are. Happy to have him back on. ai.

Harpreet, welcome back to Techstrong TV. It's great to see you. I can't wait to hear all about this.

Oh, thank you for having me, Alan. It feels so good to see old friends, so thank you. Absolutely.

But wait, there's more. Harpreet brought someone with him. And also, there's a bit of a Cloudbees connection here.

I want to introduce you to Vivek Pandey. ai. But first of all, Vivek, welcome to Techstrong TV.

It's great to have you here. Yeah, thanks for having me here. Yeah.

I don't want to steal your thunder. Why don't you tell people a little bit about your background? I'm co-founder and CTO at JoyStream, and before this, I was head of engineering at Mirantis, where my team built and shipped cloud-native and virtualization products, things like Kubernetes, Docker Enterprise, OpenStack.

And I got there because Mirantis acquired my previous startup, the startup before JoyStream, Shipa, which I co-founded and was head of engineering there as well. We built a layer on top of Kubernetes to make Kubernetes invisible. Very challenging product, but was awesome doing it.

And before Shipa, I was co-founder at Cloudbees with Sacha and a few others, Mick and Ryan. And I started out at Sun Microsystems, where I worked with Harpreet. And at Sun, I did everything from Solaris OS, that's where I joined Sun at.

So all the way from Solaris OS and X servers, people who are old enough would remember that, features over there to Java middleware. So basically, I have worked across pretty much every layer of the stack. And now putting everything I've got to work at JoyStream.

Thank you, and thanks for joining us. I hope this won't be the last time you're on Techstrong with us. So thank you and welcome.

So many people I know came out of Sun. Mm-hmm. What a collection of brainpower they had over there, huh?

It was an amazing company, but it was an amazing company, I think, was that why it attracted amazing people, or did the amazing people, chicken and the egg, or did the amazing people make it an amazing company? I'm not sure. Probably a little of both.

I personally, my very first company that I started was built on Sun technology. Right? We were a hosting company back then.

Before we called it hosting, I thought I was a digital landlord. I would just put more hard drives into my SPARC servers, right? Never ran out of real estate.

And I self-taught Solaris because I had to do the graveyard shift for tech support. Right? And so the first thing I learned was kill.

Right? How to kill a process that was... Anyway, but we've come a long way since then, right?

And here we are in this age of AI, and almost, when Sun was out there, we had this feeling of just anything was possible, right? There was such great stuff coming out of... And back then, it wasn't just Sun.

It was the golden age of, anyways, Oracle, and SG machines. And Microsoft, you can say what you want, but Microsoft was innovative. Yeah.

Right? Apple too. Right?

And there was NeXT there for a while as well. " Mm-hmm. We can build anything today.

Mm-hmm. Tell me, Harpreet and Vivek, what was the ignition point for JoyStream? What made you say, "Hey, we should do something here"?

Yeah. So since you touched upon Sun, I think we'll start from somewhere there. Okay.

Let's start there. Because a lot of the lessons that we learned there are actually making it into JoyStream. So before I get into this, I think very quickly, if you are to remember us, we are calling ourselves the operating system for organizational agents.

Right? And you could build agents on us, or you can bring agents from elsewhere. You can collaborate and operate there.

Right? And so a very quick side note about me is when I came to the US, I came to do something called AI and a master's in AI, and somehow my path led to a thesis in AI agents. Right?

Which is nothing like what it is today. So in spirit it was AI agents, and it felt like the world would be there in the next couple of years. It wasn't.

So just to give us a little context, I'm not saying you're old, Harpreet. I'm not saying you're old. But what years were that?

My thesis was like in '99. '99, 2000. Yeah.

And their dot-com bust had not happened. It was still the boom. But I got out, and I couldn't see anybody doing AI or building agents, and I joined Sun in the Java EE division, where we built out the spec and the app server in Jenkins.

And so when Launchable got acquired, I spent a year at Cloudbees, and I decided to take some time off, and it felt like, well, this is not the best time to take time off. This is last year. Right?

It's like all exciting things are happening here. What the heck am I doing sitting outside trying to travel the world? And so I started building out a few agents to see, like, all right, let's see what these agents are.

And I was bouncing these ideas with Vivek. And very quickly, it became apparent to us that what's happening is people say they're building agents or will say that they'll build agents, but they're all kind of building in their laptops. Right?

Today, if you ask anybody, the definition of agent is very wide, but most people are putting stuff in Cloud, and they're writing a skill, and they're calling it an agent. In our Java EE world, we looked at that, and we said, okay, on one side you have prompts. That is your entire logic of your agent.

And the other side is software. The prompts can change, and the software can change. And so how do you say, like, this work needs to happen repeatedly and reliably?

You just can't in a non-deterministic model, right? And this is where it connects to Sun, and we both were talking about this, and we're like, what's needed is a layer in between. That is a specification.

That's a definition of work. And that definition of work has to be frozen. And to put it in real terms, it's like most things that we do today are standard operating procedures.

Right? " And so that's been thrown out, and we're like, that cannot happen. You just can't have an agent willy-nilly do everything that it decides it's going to do.

And my favorite example of this is if an agent was driving a car from, say, San Francisco, where I live, to LA, one trip it might go directly to LA. Another trip, it might go to Tahoe. The third trip, it might just go to Seattle and then decide to come back, and I have no control.

But I really want this guardrailed path. So Vivek and me were talking about this, and we were like, okay, the work definition really needs to be upfront and central The second thing as we were talking about this, we're like, okay, if work is turning into software and we just call it AI agents today, they can't be sitting in your personal laptop. Like, geez, we moved to the cloud years ago, so why aren't these agents running in the cloud?

How long are we going to just sit on our desktop and do that? And the third thing we were like, okay, if agents are software, then why the hell did we throw out all the learnings that we have done on delivering reliable, productionizable software? And we came to this place, well, whoa, when I ask somebody what productionized agent means, nobody knows that.

Right? And so all of that coalesced, and this journey coalesced between Jenkins, between DevOps, between CloudBees, Sun, and we were like, looks like we were working our entire careers to come to this point where, in my opinion, what's going to happen is domain experts will be writing their agents that are automating their standard operating procedures, and we wanted to give them tools to just talk to that, maybe in Cloud, it's maybe on our agent builder, and just take it to production. Right?

And when you think about it like that, a lot of interesting things start dropping out of it. Right? So one is it gets to the cloud.

The other is this notion of agents themselves become versionable. You can collaborate on them just like GitHub. You can version them.

You can push them to production. And a whole bunch of features just naturally fell out from there. So that maybe a five minute on how we landed up here.

So I'm really, really excited about this because it brings our collective histories and all that we've learned over the years, and we are saying this is the operating layer that organizations should be running their agents on. Excellent. Now, both of you have long experience with open source.

Mm-hmm. Is this the kind of thing where you build it open source and build sort of maybe an open core model or something like that? Or you think this particular thing, this particular functionality, it's not for open source.

This is going to be something that needs to be commercial software, day one. So I think there are parts and parcels to this answer. So it's not the entire story that this entire solution should be open source.

So think about, Vivek just keeps educating this to me. When we talk about agents today, we are just talking about the chat, right? But it's the entire life cycle.

Like what does it mean to take it to production? What does it mean to monitor it? He wrote a blog yesterday.

Can you have a kill switch for your agents within your organization? Well, I think we're going to hear about that in the next Congress, right? This is going to be something they're looking for.

Exactly, right. So we think that there's this whole operating layer that needs to be a platform that runs. So that, in my mind, does not become open source.

But there are components of this that need to be open source. So for example, when you define an agent, and this kind of goes to our Java ethos and Java EE ethos, that agent should be portable. Like if Alan, you wrote something on Perplexity, and now you're locked into Perplexity, but tomorrow if you wanted to come over to us, you should be able to just take, this is my work.

That should be portable across. So we think agent plugin specification that has come up is the way to go. So we are agent plugin specification compliant.

So that's open source. And then the next piece that we think, I really think about this for JoyStream, is just as when GitHub came in, what made GitHub really, really successful is this had this notion of open source repositories. Right?

And I think JoyStream is this place where if your work is portable and your agents are written to a particular specification, you can have an entire marketplace or a set of repositories, we call this Agent Universe, that you can just plonk it up here, and it's open source or closed, and it could run anywhere. So it's a mix and match answer. Right.

So just to add on that, right? There's a piece of the platform, right, which is a bunch of systems that basically manage your agents, and they're the agents themselves, right? And then it's the agents that you are building, right?

And I think the industry at large is still a little bit fragmented. md. There is this specification there.

No one say that I align with that, but there's a general agreement across frontier model that this is how all our niches are going to consume a skill or define a skill. Then there's an agent plugin. There's a separate standard that basically it has come up with OpenAI and a bunch of other folks, but you don't have Anthropic there.

So then the industry is kind of, but then Anthropic plugins are more or less aligned there. So there is this whole industry has to come together to standardize how the agents are defined, how their life cycles are defined, and that's where this whole idea of, Harpreet mentioned software, right? Agents is just another software.

It's a new application, but you have to look at it from the, what people are calling ADLC, which is instead of software development life cycle, you have agent development life cycle, and how it gets represented in a open and portable format. So there's a lot to talk about in terms of open source, which basically has to start from open standards and then work backwards. I think any software can be cloned and white coded, so open source or not, it doesn't really matter at the end of the day.

No, I think that's a really important thing, Vivek. We were talking about this on the gang today. Mike Vizard, our chief content officer, used the word reverse engineering.

You don't have to reverse engineer anymore. You just have the AI duplicate it, right? Mm-hmm.

Or make its own version. And I think the whole open source model has changed in that we have open weights, right? We have open weight models.

They're not open source, but the weights are open. We have open standards. We have interoperability, right?

These are things that we identify with open source, but they're not necessarily open source, right, as Jenkins was open source or Java or something like that. But here's the other thing. You're calling it an operating system for AI agents.

There are a lot of big companies, Salesforce, ServiceNow, right, Amazon, AWS, who want to be your agent manager. Yeah. Right?

For lack of a better word. Park your agent in our garage, and we'll maintain it, manage it, service it, what have you. It could be our agents, it could be agents you wrote, it could be other people's agents.

Why JoyStream versus Salesforce, let's say, or one of these companies that I might already be using? Yeah, great question, actually. For that, I think what you need to do is look at the agent space on a spectrum.

Okay? So we primarily focus on Alan writing his agents and maybe uploading them to JoyStream or Perplexity or Cloud. On the other side, you have people building, what I would call vendors building their agents.

So somebody just tells you, "Hey, we have an AI SDR, and you know what? " We kind of have some disagreements with that. That's why we're going with this philosophy.

The third piece is this notion of vertical. It's in the same space you have behemoths like ServiceNow, Salesforce, offering their agents to operate in that world, right? And what we are seeing coming across, and we've talked to a number of people, is our people are saying like, "I actually need an operating system," or maybe you can call this as a platform Right?

A platform where if I'm on the Salesforce world, I'm locked into the Salesforce world. It is not talking to the rest of the organization. So we are talking to people who are thinking in terms of, "I have marketing, and my set of operating procedures," they're calling it as their marketing operating system.

But guess what? When I get my marketing operating system running, HR needs its own operating system because they have their own standard procedures. So you could end up in this world where everything that runs your operations is fragmented across multiple tool vendors.

Or you may decide that no, that's not the way I want to do it. I actually want to have a platform which all my agents are running across my organization. And you could talk to Salesforce from here, you could talk to somebody else from here, but these are my stuff.

So that's how we think about it. Now, the question is why us versus, say, Amazon, versus, say, Microsoft. In my opinion, that's very interesting.

What we are seeing is some of these vendors are really talking in terms of operations platforms that are really targeted to engineers. And I think there will be a place where enterprises-- Having sold to engineers all my life and being an engineer, I think they're going to come in and say, "We'll build a platform for my organization," but they're taking components from Amazon and building a layer above. And we are that layer above.

It's almost like a platform as a service versus the underlying thing. And the lens that we are taking is that we are actually addressing business users, and we are focused really on the organizational management of things. So we are not coming in terms of the underlying sandbox that Amazon might sell us to build this platform, or the underlying identity system, but we have mapped these two in our world to the components that are first class for running agents.

So I think that's where we probably are. Vivek, you want to add anything to this? Yeah, I think I would just add to what you were saying, Harpreet, which is, look, most companies and enterprises are using frontier models, but they are also getting concerned about not having enough visibility into what is really going on in the model itself.

And that's where these open weight models are becoming ever more popular in terms of having that kind of control. It doesn't just stop there. Even having an organization, mid to large enterprises thinking in terms of taking these open weight models and making it their own by fine-tuning them.

So we are going to live in the world where these hybrid models are the thing. And if you pick any kind of these vertical solutions, or even if you like working with Codex or whatever other frontier hardnesses, in essence, you are locking yourself in, but you also have these kind of models that are running on inside your on-prem or however you're renting it. And that's where this governance that works across multiple models come in, and those are the values we bring in that you have JoyStream, which is a platform that cuts across any kind of model that you can work with and govern your agents on top of them, have these deep insights into the audit of it and which agent is doing what, what it is accessing, and things like that.

So that basically is how I see it. Got it. Guys, we're going to have to continue this conversation because it's supposed to be a 15-minute video.

I think we're coming on 25 minutes. Right. ai?

That's right. Is there a trial, something they can download to get their hands on? Yeah.

ai, right up on the navbar, we have something called Agent Universe, where we built about 30-ish agents that you can just double-click Start, and just go with it. So you don't have to do anything. So I would recommend people to just go up and sign up and try it out.

All right. We will continue the conversation. I want to hear how we came up with the name JoyStream.

Ah. Well, not today. Not today, okay.

We're out of time. All right. I'm just teeing- Teeing it up ...

teeing it up for next time. Sounds good. " Best of luck with JoyStream, and we'll continue the conversation in the weeks to come.

Thank you so much, Alan. Thank you. Thanks, Alan.

Bye. " We're going to take a break. We'll be back.