Apps Are Obsolete: How Startups Can Leverage the New Era of AI Agents (Spotlight)

I get the feeling I'm in for a fight. Let's do it though. You're listening to The Startup Podcast. This is a spotlight episode.

Hey, I'm Chris Saad.

I'm Yaniv Bernstein.

Hey, I'm Jeremiah Au-Yang.

On today's show, we're gonna talk about the future of AI and the future of agents with one of the best people in the world on the topic. Jeremiah runs a fund, he runs the top event, amazing keynotes on the topic, and he has compiled some amazing insights into the future of agentic AI. He's going to share it with us today. We're going to discuss it, unpack it, and tell you what you need to know to survive and thrive in this fast-paced emerging space that is changing the way everyone thinks, works, and lives. All that, much more after the break. Stay tuned.

I know a lot of you use dev agencies to help build your product. Problem is, a lot of those agencies suck. I've been looking for one I can really endorse, and I finally found one. Warp Development are the powerhouse behind countless successful startups and enterprises for over 22 years. They have 100+ experts who don't just code, they solve business problems. From custom software to AI consulting, Warp delivers results that move your business forward. And they're in your time zone. They have offices in Australia, the United States, and South Africa. So you get local support with global expertise. Are you ready to build something that matters? Visit warpdevelopment.com, that's W-A-R-P-development.com, and mention the Startup Podcast for a free consultation.

Guys, if you're listening on the audio-only version of this on Spotify or on Apple Podcasts, make sure you go find it on YouTube or watch the video bit on Spotify because there's slides to go with the conversation and it'd be super useful to be able to see those slides to understand what the heck we're talking about. So make sure you go find the video version of this episode. All right, with that, let's get into it. On today's show, we're going to discuss AI and agents with one of the people who know all about it. We got Jeremiah Ouyang on the show, of course, and he is one of the best investors in Silicon Valley. He's the general partner at Blitzscaling Ventures, the AI fund backed by Reid Hoffman. He's the host of one of the best AI events in the Valley, Llama Lounge, and he is regularly one of the best thought leaders and presenters on the topic of AI and agents. And he happens to be one of my very best friends out here. We thought we'd bring him on to the show because we caught one of his talks about AI and agents recently., that he presented over here. And I thought, hey, it'd be cool to get him on the show to get him walk through the presentation and then we can interrupt and debate and discuss and add color to, each of the thoughts and, and have a general conversation about it on the show. Jeremiah, thank you so much for joining us, man.

Hey guys, it's great to be back. I think this is my third time back.

Friend of the pod. You're basically a co-host at this point.

I could only be so lucky. How's it going? Things are great here in Silicon Valley, by the way.

It is super warm and nice. Let's jump into it. You know, you have this as a keynote. It's something you've given at a few conferences. Usually the audience is nice and respectful and quiet and listens to your profundity without interruption, but we thought we'd get you to—

Oh, we don't get that today? You don't get that today. Usually I can throw things at them, but not here, huh?

Not here. Uh oh. We thought maybe let's go through, you know, maybe a slide or two or three at a time and then Yaniv and I can jump in, add color, agree, disagree, push back. Build on your points and just sort of give the audience an update on, on what's happening in the space. It's all moving so fast.

I get the feeling I'm in for a fight. Let's do it though.

Let's do it.

Let's talk about the agentic strategy and forecast from my perspective. I've been in Silicon Valley for 27 years. I've been investing in agentic startups, talking to the founders, lead events for them. So there's a couple of terms. So shout out, Chris Yoniff, what does that stand for?

Oh gosh. We talk about this on the show all the time, of course. B2C, business to Consumer.

Yeah.

Then this one, B2B, business to business.

So what's this? Don't know.

Well, I'm cheating, but B2A, business to agent.

That's right. It's not business to Asian. That's a very different thing. That would be every business who serves you, but that's different. Okay.

That's right.

It's business to agent, not real estate agent. We're talking software here. And then consumers are going to talk to their own agents. So there's a term called meta agent. That'll be the one that you interface with the most. There's also another term called an execution agent that actually completes the task. You may or may not interact with that agent. Then this, agent to agent. Then that leads to this, business to agent to agent to agent to agent, ad nauseam, forever. So I'm Jeremiah. I've worked in big tech at traditional companies. I was at a market research and forecaster at Forrester. I was a co-founder at Altimeter. We had an exit. And now I'm a GP at Blitzscaling Ventures. And my, my AI deals are actually reviewed by Mr. Reid Hoffman himself. He's one of our major LPs and an advisor, and we also get to invest in his projects. And then Chris Yeh is the co-founder who wrote the book with Reid, and I've been friends with Chris for 20 years since Web2. So that's kind of where we see things. And I do, I want to be with the community. So I started this event 2 years ago in San Francisco, and it was for founders, and they started to bring their laptops and show it off. Show off what they're working on. So I created a submission form and it turned into a demo event. So that's 2 years ago, Llama Lounge number 1. And one of our good friends, you can see him on the lower left, just the top of his head, Ben Metcalfe was right there. Hey, Ben. Now here's the last one I did last month. This is Llama Lounge 17. Here it is. So it's become the largest ongoing demo event in the AI community in Silicon Valley and therefore the world.. It is the largest AI startup event, demo event in the world. And we usually have 10 founders. There's usually 50 to 75 VCs and usually 25 to 50 corporate AI buyers. The next event, the mayor of San Jose, the executive team from Zoom, Zoom's hosting it, and the CEO of Product Hunt, Rajiv, my friend, will be the MC. So it's really become this big thing. And then today in the newspaper, in The Atlantic, Llama Lounge was covered. The Atlantic. Just couldn't believe it. So it's really grown to become the event for the founders in Silicon Valley, and people fly in. It's growing really fast, and so we invested in Crew, and the forecast is 10x every 5 years of growth in this space. But let's talk about how we define it, and maybe we can pause and have a discussion after I share this. Now, I interviewed over 20 agentic founders. I didn't just make this up, and I didn't just ask an LLM. I actually talked to the founders. So what the heck is an AI agent? It's autonomous software that can perceive the environment, that can actually look around or read content or have visual capability. It can make decisions. This is why it's different than RPA. It can take actions to complete those goals, and then it can learn over time. And so it can access memory, determine that plan of action, and that plan of action is typically coming from a large language model or multiple. We'll see verification and permissions. It'll complete actions, and then it learns. Now, the advanced cases are, we'll hire or recruit or even build or launch other AI agents into a swarm or orchestrate that. So a couple examples for home, maybe you want to have an AI agent plan your meals, make sure you have the right food in the fridge and the pantry, or to book those restaurants. For your email, I don't know about you, but I haven't been at inbox zero more than one time in the last 2 years, you know, Could it actually triage and respond to basic emails? And then 3, at work, maybe it can collect all the data and reports, maybe like for your show, and so you can get all that information tallied for your sponsors and whatnot. And then physically, you are in San Francisco, you see the Waymos around, those are actually AI agents per the definition, and those are level 4 autonomous, so they actually are autonomous in those ways. So love to hear your thoughts on like how I see the, the big view, and then I'll talk about the forecast.

And what I think is going to happen. Yeah, Jeremiah, this is really good. And, you know, people talk about agents a lot, so it's great to have a definition. And for me, the sort of the key words here are achieve specific goals, because I think sometimes, you know, you hear sort of people on LinkedIn saying, oh, agents is just a fancy name on top of software, right? But if you think about how software is typically created or has been in the past, it has been the how, not the what, right? You give it instructions, do this, then do that, follow this instruction. If this, then do this. If that, then do that, and so on, right? And of course they can go out into the world and do those things and maybe even perceive their environment in some sense. But to me, the big difference is when you prompt an agent, you're not telling it what to do. You're telling it what to achieve, right? You're saying, this is the goal, figure out how to do it yourself. And that I think is the new part.

So to add on to that, you raise a great point, Yaniv. So the agents are non-deterministic. Let me say it a different way. It'll complete the task differently each time. In many cases, it'll do it differently each time. And so it's actually kind of hard to forecast what it's going to do. So that's kind of important. That's different than like the traditional way people have been using AI when it's if this, then or that, where it's pre-scripted. These are not pre-scripted in most cases.

There's a lot of people these days calling basic generative AI that you, you know, do give a one-shot prompt and it gives you a response. They're calling them agents. And I think that is a terrible and distracting conflation of terms. And I think this definition that you've presented here, Jeremiah, is very, very useful. And I think also the example of Waymo is very, very useful. You know, people forget that Waymos are essentially robots and their agents. And I think that's a really useful thing to hold in your mind. And I think that matters the most to me about it is that it's, as you said, Yanev, you give it a destination, a goal, and it figures out a multi-step adaptive process to achieve that. And it does it in a long-lived asynchronous background way, meaning it's not you give it a prompt and it responds with one piece of information. You give it a goal or a destination and it takes multiple steps over a long-lived process, a process that is adaptive to changing circumstances, and then it comes back to you many minutes, hours, or days later once it's succeeded or it needs some input. So yeah, I think it's really important to separate these terms of, you know, a one-shop assistant or LLM and a long-lived agent.

But Jeremiah, like, I do think there's a gray area. And like, I do think this definition thing is important. We can, we can move on in a moment. But if you think about something like ChatGPT deep research, right? Is that an agent or isn't it, right? You say, hey, give me information on this task. It goes on, makes a bunch of decisions. It's a reasoning model. It searches the web, it gathers information, it generates a report. You could say in a sense it's an agent, but I feel it doesn't necessarily capture that, that sense of agency. It's in the gray area. And what I'd put forward for your consideration is when we talk about an agent in the, in this sort of context is really, it's not read-only. It's actually taking action in the world and changing the state of the world, right? Like deep research doesn't do that. In the end, everything it's doing is read-only. All the actions it's taking is gathering information, but it's not changing anything.

That's exactly right. So we invested in, uh, Please AI, which is an AI agent that will help you with planning. Oh, Chris, you know the founders. They used to be called Multion. They did a pivot. And the new app, and you can see on their website, will help you with planning and also to get trips together or plan your meals like I talked about in bullet 1, but also execute, do those things or reserve that restaurant or go book that gym class, right? So the execution part is going to be critical. And that's how it is different than a traditional large language model, which is basically a chat or a discussion or content being generated.

You're right. Yeah, I would say if you want to put these things on a spectrum, deep research is probably amongst the lightest weight version of a common agent. You give it a basic prompt and it does an awful lot of reasoning about how it's going to go fulfill that research task, but it doesn't affect the world. And then I think Waymo is probably one of the richest and most sophisticated agents that, especially if you're in San Francisco, you might interact with on a day-to-day basis in today's world., and, and it is affecting the real world. And so those are sort of maybe the two ends of the spectrum at the moment with, you know, sales agents and email processing agents and booking agents and so on living somewhere along that line at this point.

Let's talk about our thesis. And I worked with my partners, of course, at Blitzscaling Ventures. So right now, gentlemen, and those that are listening, you had to go find this podcast and subscribe, or you have to go get the news, or you have to go to a website and book your flights. Or you have to go to Google Maps and find that restaurant and then reserve that. You know, there's all these steps that you have to go do to research and then complete the task. You are your own agent. And we're traversing different apps and websites. And in my opinion, it's not a great experience. Every user experience is different. Our data is not respected. Even the most popular search engine in the world, Google, they know the answers to the questions, but they don't tell you. They put 10 blue links up because that's how they monetize. Right? They're having to change now with Gemini, of course, because they see this trend coming. So that's the way that it was or is. Here's where we see things that are going to go. Tomorrow, the AI agent will be the interface that will go to the internet and the apps, whether consumer or within your workplace enterprise, gather that information and bring it to you and where you want it and also complete those tasks out there and they click through it. So if you're not familiar with agents and We invested in a couple. They actually simulate the click-through paths. They actually read the text and they look at the images. They fill in the fields in the forms. Some of them are going to get access to private data and they will complete financial transactions. We back Skyfire, which has agent permissions and also an identity and the ability to do the transactions.

Go back into our back catalog for an interview with Skyfire.

Oh, that's great.

Awesome.

So that's how it's going to work. When the agents are completing the task for us. So let me ask you this question. Actually, let me show you the next thesis, part 2. What we believe is that in the future, that information will be reassembled in the way that you want, Chris. Chris, you're very visual. I know you. We're close friends. You love beautiful things. Maybe you'll have visual representations of information, even if the content they gathered was only text or audio. It can convert it and generate it. Ayanov, I don't know you well, but do you like a lot of information or a little bit of information?

I like a lot of information.

Well, then you're going to get that even if there was only a little bit of information that that agent saw on that website or app. The way that you want the information and when you want it, how much and in what format is going to be up to you. So let me ask you guys this. Let's have a discussion. In this vision, whether you agree with it or not, and we can talk about that, how many apps would you actually need and how many apps would we actually— websites would we actually need to see? I think this changes everything.

Yeah, you and I have talked about this, Jeremiah. In fact, I think even on the show we've had a little bit of a friendly debate about Hey, do I need a swarm of tens or hundreds or thousands of agents, or do I need one personal agent that then recruits others behind the scenes? And I think this thesis is more aligned with my intuition, which is you will have a personal agent that learns who you are, where your data is, what you care about, and how you consume, how you learn, how you like to be fed information. And it will, I love this idea, you know, compose that information for you in voice, video, and so on. And it also, this taps into my second hypothesis about this whole thing, which is I don't know how many AI companies or B2B SaaS companies or B2C SaaS companies survive the great agent apocalypse when the most sophisticated foundation models can generate not just text, images, and audio, but also applications and user experiences on the fly for you and the small groups of people. And, and pe— those who are longtime listeners of the show will have heard me say this many times, um, unless it is physically manipulating the real world or it is a massive multiplayer game where the rules have to be consistent for everyone, it is unclear to me why a lot of software won't be generated on the fly by your favorite fund— foundational model and personal agent.

I'll give a few exam— so my, my thinking has evolved on this too. And, um, I've listened to you and I've, I've changed some of my points of view. So I think we will have some apps and websites where they will be designed by the designers. I think it'll be in the case of media or gaming or luxury or, or social networks where there's a certain experience that needs to be crafted that's consistent, or the brand or a famous artist wants to craft that experience, and that's the experience that you want. So I'll just leave that. But going to fill out expense reports or doing banking apps, they all suck. They're horrible. I want my agent to collect that information, bring it to me, and so I don't have to visit some of those horrible things.

Let me not quite take the other side of this, but I think what we're talking about here is levels of abstraction. And if we think about the evolution of software and the evolution of human technology, really, it's about increasing the leverage that I as an individual have over the world, right? So I can get more, more change, more of the things I want for less effort, right? And you could imagine that you could, in the broad sense, you could say that even a, you know, a motor is a type of agent, right? It's an agent. If I push on the accelerator, I get propulsion, right? And so this is nearly the next level. And where I'm going with this is that, you know, a lot of the friction that exists in software is something that an agent can take away from you. But I would also caution against this idea that it's going to take away all of SaaS or something like that, because I think that this sort of layer, the agency layer, isn't the full stack of a piece of software. And if, if you think about enterprise SaaS or B2B SaaS, there is so much more to it than the interface. One of the reasons the interfaces suck is that it's often, that's not where the value really is in this system, right? And so yes, I think interfaces change, right? But it doesn't mean that that is all of software.

That's correct. In fact, you can see that Aaron Levy from Box is pushing hard into Agentic because he knows that his core value is the data stored in all of his clouds, right? Yeah. So the interface of Box doesn't matter in the future if Agentic can reassemble all of those docs and PDFs and Excel sheets in the way that you want. So he understands that. And I think the big trend is this, the data or the content layer is decoupling from the presentation layer in the Syngentic space.

Exactly.

By the way, advertisements, you can read the second line, advertisements could be ruined, right? If it's a premium agent. So this also brings up the topic of, um, business models too. Like if, are you paying for this agent, then you probably don't have to see ads unless you want to. But if you're not paying for it, like the Rufus one in Amazon, then it's going to serve up products only from Amazon, right? So that's a whole nother topic. But essentially the SaaS companies, getting back to that originally, the way that I'm imagining this is that they will provide an API to agents for fast, rapid transactions. 5 minutes ago, I just told you that the agents are actually reading the screen and doing really crude click-throughs and reading the English on the, on the page and inputting English or other languages. And that's a very slow way to do it. So we're, I'm expecting to see a new type of standard. By the way, Anthropic has MCP, Google has A2A, MIT has NANDA, Stanford has their version of a protocol working with Consumer Reports. And all these teams are at my events called the Agent Congress. So I'm right in the front seat of these conversations. Like there's these standards that don't really exist. That's actually why we invested in a company. Oh, actually, I can't talk about it yet. So we invest in a company that glues all the APIs and LLMs together in one place. A very, very hot, fast-growing company because I could see this problem was going to happen. And Chris, you and I worked together during data portability during Web2, and it's just really hard to get the whole industry to standardize. So that's why I bet on company to do that because I learned.

I think this taps into another theme I've spoken about a few times, which is, you know, you know, have you mentioned that the history of technology is toward increasing personalization. I've described this as unbundling. You know, we all know the term unbundling in the context of unbundling features from into separate apps. But my thesis is that unbundling is occurring everywhere to everything. Unbundling music from albums and unbundling work from companies to freelancers, unbundling ownership from access like Uber and Airbnb, you know, unbundling intelligence from biological brains now with LLMs. And, and the end result is that you get a personalized experience at, at runtime. And now we're talking about unbundling functionality and utility from apps and, and content. And so you're, you're, we're, as you said, we're separating the, the fulfillment layer from the UX. We're, we're unbundling, we're, we're disintegrating these two things. Yeah. But reassembling it. Yes. The, the, this continuous disintegration, this continuous unbundling allows you to compile or to combine at runtime. For maximum flexibility and personalization. And it's, it's just, just incredible. Like you just get exactly, exactly what you want and no more at any given time.

I will say that in, so we invested in Crew AI as well. And so there are templates that developers upload of how agentic orchestration works. And then of course you're gonna find these patterns. So not everything is going to be, um, reassembled in real time, 'cause once we find some best practices, then people might gravitate towards those. But those patterns really aren't clear. I was looking at their data of the use cases being adopted in enterprise agentic, and the pie chart was just slivers. Like, there wasn't even, like, a real pie piece yet because people are just trying everything right now.

I just wanted to double down because this is at the core of it anyway, right? The point I wanted to hammer on here is I think while the consequences of this may be revolutionary, I actually see AI agents as very much an evolution in the sense of they are building on what's come before, right? And, you know, you guys talked about data portability, about Web 2.0 APIs. Web services used to be the term of art back then, right? The, the idea of exposing everything via an API so you can compose it, that has existed for a while now. And what the agents do is they give you an ability to leverage that with way less effort than you could before, right? It used to be like, okay, there are all these APIs, but it's a lot of work to stitch them together. You need to be a developer, you need to have a lot of logic. And now it is the agents that can stitch these APIs together. So they're not starting from an empty foundation. They're building on this rich ecosystem that's already out there.

The problem with data portability and REST APIs or APIs of any kind was, you know, we dreamed of this world of open standards where APIs could talk to APIs without having to code defensively. And the reality was data portability was misnamed. It should have been called data interoperability. And I take full blame for that. I'm the one that, that, that popularized the term. So the, it's, it, we never achieved interoperability and each of these APIs were different and complicated and had their own object models, their own authentication models and so on. The dream of agents and LLMs is number one, we seem to be coalescing about around standard protocols like, uh, agent to agent and MCP and the data model or the communication protocol is English. And so they're able to negotiate and discuss with each other in a way that is more fluid and dynamic and doesn't require pre-made integration. And so yes, you know, this is an evolution of an old idea, but hopefully with, you know, if we squint real hard and pray real hard, hopefully it kind of snaps into place and this kind of composition at runtime becomes real finally and once and for all.

Looking for a tech partner to scale your business? Here's one I can personally endorse: Warp Development. With 22 years of experience and 100+ experts, they deliver custom software and AI solutions that drive results. Visit warpdevelopment.com and mention the Startup Podcast for a free consultation.

Let me show you a few examples. So it's not just big tech and startups. The enterprises are coming too. Like Salesforce pivoted hard into agents. In fact, they renamed Dreamforce Agentforce and they rolled out their agents for every single large corporation in the world. There's 3 use cases: marketing, sales, and customer care. And then also we're seeing HubSpot launch this marketplace of agents. I communicated with Dharmesh, the CEO and founder of HubSpot, and basically these will be integrated into the CRM. You can hire or find agents to complete tasks like doing research or to generate images or make a meme or do lots of things.

And there's hundreds of agents in there. Yeah, but going back to that slide, it looks like all of them are being made by Dharmesh.

Yeah, this was an early screenshot, but now there's many developers in there. But you guys should go to agent.ai and see what's in there, right? And you can like glue things together. Basically, it's like gluing together a bunch of Fiverr or freelancers. It's a pretty big deal. And by the way, you saw like the freelancer marketplaces, CEOs are like, we're shifting to Agentic. Like it's happening. Skyfire, if for those that haven't listened to the show here in the interview, you should. They have two, actually three major features. Have you heard of the term KYC, know your customer? Well, KYA, KYA for those that are in the know and they watch the show here, know your agent. So every agent will have an identity and that identity will be tied to the human, right? So it'd be verified identity. Every agent can also complete financial transactions on behalf of the human with guardrails, of course. I know that's the first thing that comes into your mind. And then there'll be a wallet system to manage all of these things for you. So imagine when in the future when AI agents are completing most of the financial transactions on behalf of humans, you know, the simple stuff, right? The, the, the easy things. Obviously if you're gonna buy a house, you're probably gonna be very involved. And then Crew, we talked about Crew. They are growing at a crazy rate. Okay, this is a crazy stat. They are in 60% of the Fortune 500. The company is only 18 months old. Now here's the crazy thing. It's a small team, but they have 150 agents working at the company, AI agents. So that's how they're scaling. It's agentic first. And so that company's just screaming and just, it's really fast growth. Let's talk about Google. They have, of course, they're not gonna wait. Big tech is coming through. They're very concerned. So we have a Chrome-based It takes over your browser and can complete tasks for you. And OpenAI, they're not going to sit behind and wait. In fact, they were on my panel at TED AI. And so they launched Operator, which would do something very similar. By the way, lots of these companies are trying to launch browsers. And that's probably a losing battle. Getting into a browser war? No, the winners are pretty clear. I think that's really going to be hard to unseat whoever's dominant in the existing browser space. So there's just a few examples of what we're seeing in the market.

You know, when you talk about the crazy growth of something like CrewAI, I guess I'm curious from the adoption side what you're seeing, like what sorts of businesses are able to take advantage of agents and what are the use cases that they're finding sort of giving them the most bang for their buck? Where's a good place to start? Because I'm thinking about our listeners going, okay, you know, I may not be running an agent startup, but I sure as hell would love to have an agent workforce helping me.

So where do they start? So the most common thing what agents are doing is doing the repetitive tasks on behalf of humans. And so they're training agents to do customer care things or doing reporting. Anything that's repetitive that humans are doing on a digital interface, that's where agents are stepping in. So it's a wide range of use cases. Earlier I mentioned I saw the data on the use cases a couple of months ago. And it was just slivers of just different use cases. There wasn't really one clear one. Crew is open source, so therefore it spreads rapidly through an enterprise. And so it doesn't necessarily start with the IT department. It doesn't necessarily start in marketing or sales. It's bubbling from many different places, which is a unique play into the market. So it's not clear exactly what the main use case will be. Of course, there's many other agentic startups out there. Some of them are saying we're agentic for marketing, we're agentic for sales, agenting for HR or finance, and of course they're going to tell you what they're seeing from their perspective. So let's talk about this market map. So I interviewed the leaders in the markets, the actual CEOs of agenting startups, and at a high level, there's 4 levels of the agent stack. So they're accessing data, and 80% of the world's data is privatized data— private cloud or corporate cloud or government cloud or military cloud. And of course they're grabbing open data, and there's this bubbling conversation, Chris, you should be there about unified APIs or protocols or standards, whether it's MCP, ADA, or whatever's coming out of Stanford and MIT. The next layer is the management layer. There's the permissions. Every agent will have its own identity and security. This is a critical thing for consumers and for businesses. And then there's the management layer. There's a lot of startups in this category. There's agent orchestration where you can manage all of these different agents that are doing things and making sure they're doing the right thing. There's arbitration, which agent gets priority. There's payments, which we talked about. And then of course, the— we haven't talked about this a lot, but the agents are learning, they're improving. And you need to— they have self-evaluation and reflection. There's even simulations to forecast what would non-deterministic agents do. There's a startup at YC called Covalve. My friend Brooke is running that. And so there's just lots of startups with these individual use cases that are trying to, and they're all being glued together. Then there's the application layer. There's many different consumer apps and business apps in the agent space. And then there's platforms on the right-hand side, which tend to be the horizontals. So that's where, um, LangChain and Crew fit into there. And then at the top are marketplaces. And I showed you Agent.ai. And we should assume that every app store in the world is going to have agents in there that you can eventually download, use, or hire. And this is a new thing to think about, that employee base might come from an agent store. So that's where we see this agent ecosystem right now. This map is constantly changing. It's already out of date, but I'll update it as quickly as I can.

Yeah, I love this diagram and I can see my name down the bottom there as one of the contributors, which is awesome. I've never quite understood why you have LLMs in the middle there at the application layer versus maybe just above the data layer as its own kind of intelligence layer. That felt like the more natural place for that to fit, but—

Probably nuance. It could fit in multiple places probably, but the apps are calling an LLM in order to have natural language processing to read a website or an app and then to put input text into the app. That's the main use case, but it could also go into management layer because they're sourcing LLMs to actually determine tasks in which to complete. So I guess it can go in multiple places.

Yeah, just in, in a technology stack diagram, you typically have kind of intelligence as its own layer in the abstraction stack. But in any case, the thing that I've noticed is missing from this, having advised a startup on this, is because this envi— this environment is so heterogeneous, there's going to need to be increasingly a, some kind of multi-platform publishing tool. Where if there are multiple personal agents or multiple agent marketplaces or multiple agent frameworks, and they're not all going to be the same, you know, there'll be the ChatGPT ecosystem, there'll be the Siri ecosystem, there'll be the Meta ecosystem. It seems to me that there's also room somewhere in here for some kind of Shopify for agents or MuleSoft for agents where you can describe your agent once. Oh, interesting. And then manage its publishing to multiple ecosystems and endpoints. Uh, and it's not something I've heard anybody talk about. Uh, and I've, I've seen a couple of people attempt to start and then give up halfway through. Uh, and it feels like a really big opportunity.

That's interesting. So the Interop Connections, we've done an investment there. They were actually at the Llama Lounge event you attended at AWS as one of the demo companies. So you can probably figure that out, Chris., but having a place to tell it how to go and operate against all these apps, no, that I haven't seen that feature.

I haven't seen that developed yet. Yeah, I think it's a big opportunity.

I do think it's a big opportunity.

If you find it, you gotta bring that to me. So Jeremiah, if I'm a founder looking at this and I'm saying there's an ecosystem and I think, you know, this is the next big thing as you believe, and I want to look for opportunities to do my own agentic startup, what, how would you think about it, right? Like how would you look at this map and try to map it to specific niches, specific use cases where there are startup opportunities?

So I think there's deep vertical, so application layer agent apps. So a deep vertical agent that goes start to finish, the whole stack and the whole thing, um, that is an opportunity. And I am very sure based upon my friends at OpenAI, they don't want to do any of that. They want to leave that area alone for other people to do. They want to go horizontal. So that would be a place to do, but you have to still make sure that you build a, a business model that scales too. It has network effects and viral effects. Those are two things that we look for investment thesis. So I think the application is the untouched. The marketplaces, I think the big tech companies are moving into that category. It's kind of hard to build there now. There's still a lot of work to do in the management layer. We're going to see consolidation. So each of those dashes of those and each of those words separated by a comma, there's like 10 companies in each category. It's a lot. So we'll see consolidation and suites emerge out of that. It's just a natural way of tech.

There's many ways to play this game. That's right. So I think the application layer is where I was thinking, right? And as you say, it's the largest sector. So if you were actually to draw this ecosystem to scale, that would be far bigger, right? 90% of the amount. That's right. And if you think about the diversity of different employees, right, the different job titles that exist out there, each of those, you should be thinking, you know, what is one or more categories of agent that with the correct expertise could do some of the tasks, some of those more repetitive tasks that a person is doing in that job. So I don't know if that resonates with you, Jeremiah, but I think as a framework, in a sense, like you said, the agent marketplace is like, here are all the different employees you can hire, right? That's what like Agent AI is nearly like Fiverr for agents. You know, what's, what's the equivalent? Like what can you create with perhaps some of your domain expertise for a specific vertical or a sub-vertical or a role within a sub-vertical where there is a big enough market and you can provide enormous value?

Yeah, I mean, you could essentially map out all the industries and for each industry, map out all the roles and for each role, map out all the tasks. All right. And you could start on a task-by-task basis cannibalizing the tasks of a given role, and then on a role-by-role basis cannibalizing the roles in an industry, and eventually cannibalize a player in that industry and build a multi-million, multi-billion dollar player in an industry. And when I say industry, I mean healthcare, material science, you know, education, insurance.

Yes, all that. So if you look at the latest YC applications,, what they're looking for. They're like, please don't build, productivity apps for law firms. Build an agentic law firm. It literally says that. I'm like, wow.

Okay. Let's go there.

Yeah.

We discussed this on the last Reacts episode, Jeremiah, with Amir. And, that was the first item on their call for startups. And we discussed it a lot because it's, it's my favorite thing. Uh, you know, the longtime listeners of the show have heard me say a million times, stop propping up incumbents and disrupt them., and that means building full-stack startups that rethink the way an industry works. Are fully agentic. That's it. Well, originally it was a full software stack, right? So Uber didn't prop up taxi, they destroyed it. Um, you know, Netflix didn't prop up, uh, Blockbuster, they destroyed it. Uh, you know, your agentic startup possibly, probably, likely shouldn't be propping up a law firm. It should be destroying one. That is where the most value creation is going to occur. It's not for everyone. It's not, you know, it's not for everyone's taste or capital availability, but it's definitely my favorite thing to do. The other thing you also mentioned that we've discussed recently a lot is the idea of, you mentioned this a few minutes ago, the idea that some things will be ads-based, but more importantly, a lot of things will be paid and you'll turn off the ads. And this era is going to be very much the era of paid applications or paid functionality. And you also touched on this idea that OpenAI doesn't want to do a lot of this stuff and there's a lot of space for you to play. We also discussed that on a recent Reacts episode and the platform risk of the hyperscalers coming in and eating your lunch. You know, they just bought Windsurf. They are getting into, uh, agentic, you know, IDEs. And it's only a matter of time, I think, until they decide to justify their valuation, to maintain their momentum, to, to eat more of the value chain. They're going to get into more and more applications. If not them, then Google, Microsoft, Apple, and so on. And so yeah, you have to play a game of, you know, avoiding the giants and also avoiding platform risk from these LLM API providers who are maybe going to one day cannibalize you from the bottom.

Yes. And they might launch, they might play just at the marketplace. So they might have a generic agent that's horizontal that can complete tasks if you give it the instructions. But they may just have a marketplace where there's vertical specific apps like they try to do with GPTs. So it's, it's really not clear. Um, but either way, the founder is going to, should expect incredible competition in this market. It is, there's 35,000 AI projects now and it started at 3,000 in the database. There's an AI for that 2 years ago, 10 times growth. So this was made, this forecast by Gartner last year that search engine volume will drop by 25%. And I'm sure you've been talking to startups and I've been talking to companies and marketers. They're seeing their referral traffic from Google plummet. I'm hearing from multiple sources now. No surprise. And then Bill Gates also said, hey, whoever wins the personal agent, that's the big thing because you'll never go to a search site again. You'll never go to a productivity site again and you'll never go to Amazon again. So shots fired there, right? But we've already talked about how we see that. Just to close things off, this is how I close off the keynotes. AI agents are going to bring information to human and they're going to convert it into the way that you want, when you want, and how much you want. We're gonna see changes in media and business models, right? This is significant when we see that unbundling, as you said. The traditional companies are launching their own agents to talk to my agents, and then there will be APIs and there'll be new business models that emerge. And then we're seeing agents throughout the enterprise. They're like assistants, and then will be colleagues, and then sometimes managers. By the way, every Uber driver reports to an AI for over 10 years. It's like, that already exist. Uh, your partners and customers and even your competitors could become an agent. And that leads to this final, very science fiction thought, and I get a lot of pushback from people on this one, but I'm going to say it anyways. I think we're going to see a new type of organization emerge in a few years, and the best term that I have right now, guys, is called an autonomous organization, where there's no human propping it up. So right now Silicon Valley is abuzz about the a billion-dollar unicorn run by one employee. Well, I actually think we're gonna see an autonomous organization run by no employees, and it'll offer a service, probably a digital service that people hire or developers hire or humans hire, and it'll generate revenue and it will improve and grow on its own. So that's what I see.

This is my forecast for the agentic space. We often remind people on this show that Ben Parr is the one that coined this idea of a unicorn, a billion-dollar company powered by 3 people., and then you took that idea and posted it and shared it on LinkedIn, uh, and, and, you know, mentioned the show and, uh, that went really viral, right?

That was over a million views on LinkedIn, which is very rare, extremely rare because it upset a lot of people.

People were really upset and I actually got a lot of hate mail for that. And so that, this idea, you know, I'm sure it was independently, uh, you know, invented in a lot of corners, but I love to take credit for it on this show. We popularized this idea together. You know, one little bit of pushback on this, uh, I'm not convinced that the concept of a traditional API, right? A REST API, you know, call and response survives this. In my worldview, an agent replaces an API. And the difference is, is the agent is an English-speaking protocol to access or manipulate a system or data or the world.

And so I'm not sure why you—

That's basically what MCP is.

It doesn't have to be English though. No, but MCP is basically a way of creating an API without all the hard work of defining a really precise technical interface.

Well, MCP, as I understand it, is a sort of a way of generalizing data access or functionality access APIs for agents. But what I'm saying is that every piece of functionality or data will have an agent sitting on top of it, even a rudimentary one. And so the protocol will be English.

Or conversation and reasoning. I think that's an assumption. It may not be English. They may, an agent, they might start to develop a new language on their own, like the way modems have.

Okay, let me generalize what I'm saying. Except for the most rudimentary and primitive interactions, I suspect there will be an agent of some kind sitting on top of functionality and data, and that agent will be conversational, and therefore the protocol will be conversational. English or some made-up jabber talk language. And so defining a structured REST API with JSON objects and whatever is antiquated when you can talk to every object feature or function or dataset. The APIs become less interesting or important. One might argue in, to counter, as a counterfactual is like, that's a very expensive way to negotiate and transport information because you have to add reasoning on top of everything and burn a whole bunch of tokens. But I think you could argue the same thing about, you know, it's expensive use of memory to render a GUI on top of a computer system until memory became almost free and everything is GUI. I think when tokens become free, all communication or most communication will be conversational or agentic or generative.

I'm inclined to agree with you, Chris. Like, the only thing I'd say is this agent that you're talking about, that is the API. So it's still an API, just to your point. It's not a REST API. That it's nearly like you're walking into a store, right? It's got a bunch of data or services sitting behind it and you say, excuse me, this is the goal I'd like to achieve. How can I achieve that? And the agent is effectively that expert shopkeeper or that, you know, expert consultant who helps you get the best value out of their services.

Yeah, broadly speaking, an API is an interface between two systems, right? That's the most broadest definition. But when people generally talk about an API, what they're really saying is that there is a an opinionated structured endpoint with an opinionated data model, with an opinionated method of exchange, with an opinionated call and response. And it tends to be fairly static in its design. And what I'm saying is, yes, an agent is a form of API, but it is, it has high degrees of flexibility and it is not pre-prescribed in many ways. And so it is a next generation API, a conversational adaptive API. And so therefore we don't need APIs as we have traditionally imagined them, structured with, you know, formal objects and so on.

Guys, thanks. This has been amazing to get your feedback and think about how this could evolve and how APIs will evolve.

Thank you for having me. It's been amazing, Jeremiah. I think this is, I think hopefully we have created the canonical artifact for your talk here. You could just hand this out and basically skip your speaker's fees, right?

I should just give all my money to you and the agent.

Feel free. Jeremiah, it is always fun to have you on the show, always too infrequent between visits. How can people pitch you their startup? How can people hear you talk? How can people follow your work on the internet?

Where can people find you? Sure. I'm on LinkedIn. That's where I'm most available. And we're looking for startups that have viral effects and network effects. That's the prerequisite. And if you know how Reid Hoffman is invested in his history, you know, that's what we're kind of looking for.

And he does review our AI deals. As I said, the massive multiplayer games are very defensible in the generative AI world. And so yes, if you have a network effect, I think that's a really smart place to be. Yes.

Yanev, how can people find you on the internet? I'm mostly on LinkedIn trying to make it less boring. I love getting connection requests from TSP listeners. I just got one this morning. So if you'd like to connect with me, just let me know you're from TSP and I will accept the connection. Now, Chris, I know you like to help startups going through this complex world of change and come up with a really sharp strategy that meets the moment. So if people want to work with you, how can they do that?

Yeah, you can find me over at chrissaad.com and chrissaad.com/newsletter, and you can find out about my advisory work at chrissaad.com/advisory. Jeremiah, an absolute pleasure as always. Thank you for joining us, and thank you for listening.

We'll catch you in the next one. Bye-bye. Bye-bye. This episode of The Startup Podcast was brought to you by Warp Development. I know a lot of you rely on dev agencies to help build your product. Problem is, a lot of these agencies suck. Ask me how I know. I've been looking for one that I can really endorse, and I have finally found one. Warp Development are the powerhouse behind countless successful startups and enterprises for over 22 years. They have over 100 experts who don't just code, they solve business problems along side you. From custom software to AI consulting, Warp Development delivers results that move your business forward, and they're in your time zone. They have offices in Australia, the United States, and South Africa, so you get local support with global expertise. Are you ready to build something that matters? Visit warpdevelopment.com— that's w-a-r-p-development.com— and mention the Startup Podcast for a free consultation.