Making $$ with AI Agents

Howie Liu is an absolute legend. I mean, this guy started Airtable, half a billion in revenue, $1 billion in the bank, growing quarter after quarter. So he's one of those people that when I want to know where is the world going, I call Howie. This episode is structured into 2 parts. First, where is the opportunity when it comes to AI agents? I think that there's $1 trillion up for grabs in AI agents. Does he think there's more? Does he think there's less? Spoiler alert, he thinks there's way more and we get into it. The second part of the episode is where he reveals HyperAgent.com. Now HyperAgent is an AI agent builder that allows you to build digital employees, allows you to build apps on different ideas, and I don't know why more people aren't talking about it. So I had him just give us the tips and tricks for how to use HyperAgent so that you can outperform 99.9% of people. I got good news. Howie is going to give you $1,000 of HyperAgent credits, no strings attached. You just log into the account. There's going to be $1,000 right there to go and build the business of your dreams. The catch is first 1,000 people do it, get the $1,000. He's committing $1 million. How crazy is that? Just writing a million-dollar check of tokens to you, to the Startup Ideas Podcast community, to play with HyperAgent, to automate some stuff, to do some research. To build their business. So thanks, Howie. You know, all I ask is you like and comment on this video. Show some love for Howie for doing such a cool thing. We need more entrepreneurs, more builders, and it's—I'm stoked to see him support you all. Thank you to Airtable for sponsoring this episode. You guys are legends. Enjoy the episode and have a creative day. Startup by this podcast.

It's sipping time, baby.

Feeling really lucky right now because we've got Howie. He's the co-founder and CEO of Airtable, and today we're going to talk about agents. He's going to do a little show and tell of his new product that I've been using for the last few weeks. But first, Howie, I have been— I haven't been sleeping very much, to be honest.

Is it agent psychosis?

Yeah, exactly. And I just need your reaction to just some things I've been thinking about.

Yeah.

So this chart over here, it's by Sequoia. In what domains are AI agents deployed? You can see software engineering is at almost 50%, back office at 9%, marketing and copywriting 4%, sales and CRM 4.3%, and down. When you see this, like, what's your reaction?

I mean, I think 2 things. One is I think it absolutely reflects the underpenetration of AI in industries that clearly could already be disrupted or benefit with even today's AI capabilities. If you took frontier agents today and deployed them into every one of these categories, you should get to 100%. And then 2, I think even the higher numbers like software engineering is actually kind of an overestimate. Meaning, as I think frontier developers and companies applying frontier agentic development practices are finding, The new model of software development is not even just every engineer using AI autocomplete, like tab autocomplete, which we all figured out 3 years ago with even GitHub Copilot. But it's now like you don't even need the IDE. The way I develop on HyperAgent is I have 30 different Claude Code instances running in parallel and each one is coupled up to a browser, fully autonomous. It can go and get other agents to comment on any PRs it creates. And so this modality shift of no AI to what I would call Gen 1 AI, which is basically AI augmentation for still very human-driven development workflows. Andrej Karpathy talked about in October, November is when he completely inverted from mostly still human-written code with AI augmentation to completely the opposite. And that's what we've seen the frontier companies leap into. I think even the 50% is an underestimate because the number of companies and even people who have switched into that new frontier mode is actually definitely less than 50% of software engineering today. So I think what we're actually seeing is the frontier is advancing so quickly and many companies and many industries and many functions are barely catching up to the 3-year-ago state of the art, let alone disrupting themselves and their industry with the new state of the art.

Right. I mean, another way to think about it is like there's copilot territory. These charts are from Sequoia, right? There's copilot territory. There's autopilot territory. Like, how do you see— you look at this, right? This is what Sequoia says. There's a trillion dollars up for grabs within agents.

Yeah.

But they're very different. What's your reaction to this?

I mean, look, I think to me it's like these agents really reached a breakthrough, really call it like 4 or 5 months ago.

Yeah.

I think developers felt this with Opus. Opus 4.5 just kind of set a new high watermark of like, whoa, this thing for the first time really feels like a true software engineer that's able to work on a task that would've taken a real human engineer. maybe many hours, if not days. It can go do it completely autonomously and it ships me a perfect clean PR that I can just review like a reviewer would. And I think that that experience is going to be unlocked and already is unlockable across every single other domain. Because we've kind of just reached this point where the models are more than smart enough. You talk to these models even in a more synchronous chat interaction, not like an autonomous agent interaction, And you can ask it the most advanced things, give it really complicated subject matter content, like management consulting. You give it some really hard meaty problems in the context thereof, and it gives you really smart answers that truly are expert level. And so it's clear that the model intelligence is there. The models are smart enough also to coherently execute across multiple terms with lots of tools and context. I think it's more of just a matter of how and how quickly we can deploy agents into every role and industry before we can truly just almost do anything that humans could do in each of these functions with agents. The TAM for that is not even a trillion. It's probably the whole GDP of all white-collar labor, which is obviously many tens of trillions in even the Western Hemisphere alone.

Right, which is sort of like, I don't understand how you're not— how people aren't motivated to create startups right now in that sense. Like, the person listening to this is like, yes, yes, Howie, you know? Um, but it just feels like, you know, I can't think of a better time to be creating a startup than now.

Totally.

Right?

I think like, I mean, yeah, I think the weird thing is like, it's almost like using is believing, right? Like, it's really hard to fully grok the power here if you haven't actually gone and hands-on spent at least a full weekend playing with agents. And that means more than just the superficial, you did some naive one-shot thing like, hey, who's going to win the next presidential election? Kind of question that you could have asked a chatbot. I think people are not actually coming in and when they're doing light experimentation, they're not actually putting an ambitious enough prompt or task in front of the frontier agents. And they're still using it like they use Gen 1 chatbots. And like until you actually experience the full power and autonomy of these frontier agents, you know, I think it's hard to fully extrapolate like what types of companies can be built now that weren't possible before. Structurally, how could you build like a multibillion revenue business with one human and like hundreds of agents, right? Like you have to use it to get it.

Also, you know, this is another chart I can't stop thinking about, which is the unit economics just absolutely crushed when you look at a human Human person versus an AI agent and what it costs, you can create some serious gross margin businesses on top of this.

100%. This is the funny one because I've seen a lot of people complain about the cost per token of the frontier models. Opus 4.6, now 7, clearly the most expensive model. Then GPT-5.4, very good, still kind of expensive. Even open source, it's cheaper, but it's not free. And I think some people are struggling, I've seen, to adapt to this mental model of, in the old days of software, a lot of stuff was free. Even ChatGPT has a free version that you could use however much you want. You get a cheap, dumb model, but you're not expending that many tokens because it's not actually doing autonomous multi-turn work and expending a billion tokens every few days. It's much more token cheap or token lean. And I think that we have to get over this hump of anchoring our price expectations for AI on traditional subscription software where it's like, oh my God, I have to pay $20 for Netflix per month now instead of whatever it was, $12.99 before. And instead think of this as, yeah, to your point, how much would it have cost a human to do the thing, right? Like, you know, if I wanted to go and like create an entire marketing campaign, or actually in my, you know, CEO role, like, it's funny, like one of our recent board memos that I wrote and sent out to our entire board and kind of major investor list, like, you know, a lot of it was researched and crafted by HyperAgent, right? Obviously with like my, you know, kind of instincts and context and whatever imbued into the agent. And of course, I oversee it at the end. But like, I got feedback that that was the best memo from some of our best investors that I had ever written. And I'm like, yeah, like, you know, because an agent did it. And by the way, I got to do it in like 10 times less time. And so like, even if it cost me, let's call it like $150 of tokens to generate that output, like, think about the opportunity cost of my time. And so I think that is a real reframe moment that's needed is Let's think of this as like, what is the human equivalent time cost versus, wow, $150, that sounds really expensive versus like a $10 per month sub.

100%. Yeah, I think the way I always think about it is like, I anchor it around value, right? What's the value I'm getting out of that? I mean, the truth is with your, you know, your board deck or whatever, like, it probably was the best, you know, it probably was the best because you had so much research support.

Yeah, totally.

2 more quick graphs and then I want to get into HyperAgent. Percent of enterprise apps with embedded AI agents. This is the fastest adoption curve in enterprise history, right? So when you see this, how do you react?

I am not surprised and I think even this reflects the pace at which incumbents can even integrate AI into their products. And I think even that is like, stimmed by just incumbency and kind of how seriously did enterprises, enterprise apps or enterprise app makers or internal app teams take this. I think the real show of how profound this growth curve is, is if you take the aggregate revenue created from zero of all the leading AI companies or companies doing AI things. Take OpenAI and Anthropic alone. Let's just say they have a combined revenue probably of $80 billion plus right now, up from basically zero a few years ago. In the history of software, has there ever been an industry where any company, let alone, or even in aggregate, across all the companies, you got a category that went from $0 to $80 billion plus. And that's not even including all of the other AI providers, inference providers, and tooling, etc., out there. The revenue of the AI category is an even sharper curve. And I think that really reflects just how profound this lightning in a bottle is.

Totally. And just from an opportunity perspective, it's like, Selling to these enterprises and helping them figure it out and just helping them transform is just a huge opportunity.

Totally. I think it's probably one of the bigger cash grabs in business history. There's kind of 2 angles, I think, to create a very valuable business right now with AI as a wedge. One is PLG, and obviously we see a lot of these PLG products. I kind of put OpenClaw itself in this category because even though it's not actually a monetized business, it is getting this massive amount of adoption. And just the raw token consumption through OpenClaw is, I'm sure, in the many hundreds of millions, if not billions already. And likewise, other products in the PLG genre. So that's one way. It's just like, let people use the AI thing that actually works and you're going to get profound growth. But the other is to come in top-down Palantir style. This is why OpenAI and Anthropic and The big guys are also doing it. There's new companies as well going after this opportunity, which is go pitch to every enterprise board and CEO. We will fix your AI problem. Pay us a massive check. Give us a $100 million plus check and we will purportedly solve your problems for you. That is an existential risk mitigation that every large company incumbent should be willing to pay. Because frankly, the CEO's choice is either I pay it and I risk wasting $100 million and maybe getting fired over it, or like, I don't do anything with AI and I'm definitely getting fired over it. So on a game theory level, it's like everybody's gonna pay it, right? Now, whether that actually results in like long-term substantial structural, like, you know, kind of transformation to the business that probably could be run now with like 5 people maybe instead of like 50,000, right? In some cases, that's a bigger question.

Yeah, and this sort of speaks to my last point too, which is if you can help a company run a fleet of 20 agents doing customer intel, content production, competitive research, lead enrichment, all these different things, this is the future of work in one image, right? An agent command center, right?

Yeah.

So when you see this, your reaction?

I mean, look, that literally— is a view in HyperAgent. I feel like I'm looking at a HyperAgent and I think this is the future. We are building towards a world where it may not be that every company is literally one person and we have a lot of one-person companies, but I do think every company will have a fleet of agents. What's interesting to me is actually that agents are converging on these purposeful— they almost map to job roles that humans were playing. Maybe it's a little bit like, why are robots, hardware robots, converging on a humanoid form factor? Part of it is a lot of the infrastructure of everything we have in our homes, in construction sites, in factories are built for human ergonomics. So for the robot to effectively insert themselves seamlessly with the current infrastructure, they have to have human-scale scale capabilities. I think there's a very similar phenomenon happening with agents. I guess 5 years ago when people talked about superintelligence, I always imagined there's going to be this single omnipotent AI that just figures everything out and looks at everything all at once. Everything everywhere all at once. I think now I'm more and more of the belief that There are going to be fundamental and always present limitations on context windows, for instance. I just don't think we're ever going to get to a point to where an AI model can have infinite context window. And I think there's a physics to that. You can just literally only have so much attention on so much context at once. And I think what that means is that for the same reason why we partition humans into different roles and org structures so that not everyone in a company has to know everything and work on everything all at once. I think the same is true for agents. Hence, you get this overview of agents that actually maps to intuitive human-played roles really well. That's the really interesting emergent phenomenon for me. I just recently spent some time playing around with Paperclip, Which is kind of fun because it literally creates the org chart metaphor. But I think this is really exciting. In a way, it's both familiar because we're not just completely upending everything we knew about job functions and roles in the old world to the AI world. And yet there is a rethink and reapplication of, okay, how do I play that content production role with an agent?

Right. Well, I think we should get into HyperAgent. I think now is the time, right? So, you know, for the listener, like, what is HyperAgent? Why are you building it? And this is a show-and-tell podcast. So, you know, by the end of this part of, you know, by the end of this episode, like, you know, can you commit to, you know, giving all the sauce around how to use HyperAgent to sort of build a business?

Sure. Yeah, let's, let's go for it. So this is HyperAgent. I'm currently in a thread. I'll zoom out in a second and kind of I'll show you what the entry point looks like. But think of HyperAgent as if all of these other agent products out there like OpenCLAW, etc., are more like Linux. HyperAgent is our take on the Mac version of it. We want it to just work, to be secure. It's cloud-native. You don't have to run a Mac Mini. And perhaps most importantly, HyperAgent is applying a lot of the same design philosophy and obsession with great UX that we applied to the no-code app category 10 years ago, but now to agents. Meaning, apps are kind of complicated. If you're a developer, even at that time, you could build a Rails app. You had a data layer, a logic layer, a view layer, but it was kind of technical, right? Or very technical. And the whole idea of Airtable was to distill that into a really intuitive experience. In fact, we were very inspired by The Macintosh, the GUI, taking terminal-based command-line computing and making it into something that people could just grok immediately. Hyperagent is really intended to be a very intuitive and visual way of using agents. This is actually a task thread that I ran a little bit earlier. This is actually one of your startup ideas, Greg, that we had Hyperagent work on. And basically the pitch was hyperlocal market reports for real estate agents generated from public data. Right. And, and so basically this agent went around and did research on the landscape of the market. It ran a bunch of like analysis. It's got full coding capability, it's got a full sandbox environment. So it is running a full computer. It's just one in the cloud, not like, you know, kind of your own computer. And you can connect it to all your accounts if you want. Like it can access your Slack and Granola and email and send stuff if you want it to on your behalf, or just pre-draft emails. You know, it's got already preconfigured ability to do things like pull from Twitter, use advanced tools like generate imagery or use Google Maps, etc. But basically what happened was it went around, it did all of this, it researched the opportunity, right? And then created this research brief. And let me just show you what this one looks like. this is kind of the business case for the idea you pitched, right? I kind of love it because I actually think these, what I would call medium-sized markets, it's not like a $100 billion market, which is going to be super competitive and there's going to be massive incumbents going after it. But I really love this idea of the kind of like, maybe it's not micro, it's more like mini or medium market, like a couple billion TAM large, which is to say you can build a very lucrative business even capturing like a double-digit percent chunk of this. Like, you can make a few hundred million per year, and yet, like, it's small enough to where really big guys are not coming after it, right? So, you know, this agent created kind of a business case for it. It found some really cool, like, user validation of the problem. So it saw, like, you know, looked up Reddit, like, you know, and found, like, some real real estate people who are actually saying, like, I need this product, right? So it's kind of validating the market need. Here's actually the current problem. I didn't even know about this, but apparently I guess there was some legal thing that changed the dynamic of the market. People don't want more software, like another tool with an interface. And did some competitive analysis. Here's who else is out there. And then just put together the case for this. But then better yet, you don't just have to stop there. You can go and actually tell it to go and just build a V1 of the product. In this case, because HyperAgent has full coding capability, it just went ahead and created a V1 of this product. I think this will actually work. Where do you farm? Here's my report style.

It also looks really clean.

What's that? Honestly, a lot of this is just like if you have a good frontier agent running a frontier model, i.e., Opus 4.7 or GPT-5.4, It just does a lot of this really well out of the box. So any frontier agent powered by frontier models should be able to create an app of this quality. What's unique about HyperAgent is that it can do that perfectly well, but then kind of do that in the, in the workflow of like, it's not just an app builder. App building is just a feature now. It's a commoditized feature. And what it can actually do is like go and research the end-to-end of like, here's actually the business context of what I'm trying to do. and then build the app informed by it, right? So it's more like HyperAgent is the founder in this case. It's not just the developer, it's the founder. One other cool thing I like about HyperAgent is like it just comes out of the box with like really powerful tools. So it has like, you know, Google Maps as a tool and it can actually go and like, let's say, I think I already did this, but like I wanted it to go and actually find like real Street View imagery of billboard locations. So it knows how to use Street View to find actual points of interest and then to take that image and use that as a reference seed image for AI image generation or video generation. Another cool thing you can do with HyperAgent is you could tell it, take this house and I want you to redesign the house using interior photos from Zillow or the exterior shots. And it will do that like really, really well. Right. So that's HyperAgent in a nutshell. I can walk through some of the other stuff here. You know, once you actually build like a lot of agents, then you get like this, this ability to start looking at like, well, what if I wanted to see, you know, not just my one agent— sorry, but, but an overview of all of my agents. Right. So this is not like a very built-out account. This would be like your first week of HyperAgent use. But literally that command center view that we talked about, we want you to be able to create many different agents that each play a role. Here's the content marketer, here's the market researcher, here's the customer email responder, and just manage and oversee an entire fleet of agents, constantly improve them because we actually have this ability to go and curate memory and skill improvements from every run that you do. And then finally, to be able to deploy them into a team setting as well. So if you wanted to take any of these agents and actually give it the ability to talk in Slack, right? So I can actually say like, let me put this into Slack. Let me have it always on, always listening, in fact. And, you know, just sit there in my channels listening to everything I'm talking about, my team's talking about. And when I have something relevant to add, to automatically chime in and then people can interact with me truly like I'm a virtual coworker. I think that's part of the OpenCLAW experience I've seen some of the power users achieve that's really quite magical. Your Slack coworkers are now agents in addition to humans and they're really smart and they have their own expertise and context. You get that with a single click out of any agent that you build in HyperAgent.

You mentioned skills. How do skills work on HyperAgent and how should people think about it?

Skills are, I think, the most important concept or primitive in the frontier agents world. Meaning the models are generally intelligent enough. It's like, find Albert Einstein, who's obviously super smart in a general sense, and he may not know how to solve problems in real estate. But if you gave him just the right briefing on, here's a playbook, here's a manual to learn everything you need to know to do this job in real estate, Like he's going to go and like figure it out pretty well. Right. And so what's really powerful about skills is like they're a really, really composable concept. Like you can interactively create skills. So let's say I'm actually going to create like a new thread here, just keep this super clean, but like help me create a skill that posts Greg Isenberg-like AI content. Okay. And so what's really powerful about this is like—

No, don't create this. Don't create this. No, no, no.

Okay. But worse enough that, you know, we don't take Greg's business. Exactly. But what's really cool about this is like, it's not gonna just like go and like say, okay, like, you know, I'm just gonna have a prompt that, you know, pretends to be Greg Isenberg. It could actually go and like, you know, research how you actually, Do content. So it's coming up with a plan. The plan is like, I'm gonna first go and like research your style, figure out like what platform I care about, like look at some of your actual posts, and then distill all that into a skill that I can then pin to an agent or like just use on demand at any point, right? So let's say just for fun, like what platforms do you wanna post to? Let's just say X for now. We're gonna have the skill only generate drafts, so it's not gonna auto-post for you. Is there any kind of content you want your agent Eisenberg to be focused on?

Yeah, let's do contrarian AI takes.

Okay, cool. And then any topics beyond that?

Like solopreneur, bootstrap, life challenges.

And then how do you want to use this agent if you end up using this agent? Do you want to start with an idea? Do you want it to just come up with ideas for you? I don't want to do it anymore. We'll go full autonomous. Someday we're going to have to see if real Greg is actually just sitting at the pool all day. It's just created the Greg avatar version of you and is doing everything on its own. But okay, so now it's going to go and do some research about you and figure out how to distill the perfect skill for Greg into How should people think about Hyperagent versus Perplexity Computer versus Manus versus OpenClaw itself?

Codex? How do you see it?

Against Codex, it's quite simple. Hyperagent is a more general-purpose agent platform. Against OpenClaw, this is much more turnkey, ready to go, safe and secure by default, cloud-native. And I think just much more focus on great UX. OpenClaw, when you actually have to go into configuration or you're trying to edit memories or do any kind of curation or configuration, it's quite raw. It's a very raw product. It feels like it's more for very technical people who've become experts at it. I think Perplexity Computer and Manus are the closest comps for Hyperagent. The key difference is Hyperagent has more powerful tools out of the box and also has more focus on UX out of the box. I've spent some time playing with both of those products. I think they're great products and at their time, or at least when Manus first came out, it was truly groundbreaking. It was the first kind of real holy crap YOLO agent. Look at everything it did. Kind of like before even OpenCLAW, long before OpenCLAW. I think they were really pioneers in this space. With HyperAgent, we've just taken a very UX-focused approach. For people who like seeing visually and being able to interact with the outputs and see more visually what the agent is doing and have a more visual way of defining skills, deploying skills, creating agents, etc., HyperAgent is just much more of the Macintosh experience versus the Linux. I think secondarily, We've also done a lot more to make Hyperagent immediately ready to run not just one agent. I think the nominal experience for Manus and Perplexity Computer is still like you use those products and you have this agent that's pretty awesome and you use it directly. You can do that with Hyperagent. That's exactly what we're doing here. But it's also designed from day one with much more of the scalability and deployability story in mind. Like once I have an agent that kind of works for me, I can now deploy it one click into my Slack channel. And now everyone in my company can benefit from this agent just always on, like kind of chiming into conversations. You know, they can ask it questions, it will respond. You have the command center, that fleet view where it's not just one agent, you can oversee your entire fleet of multiple agents. And we even have things like, you know, the ability to oversee and curate like the learnings that keep making each agent better. So like they kind of have this automatic self-improvement loop where over time they're accumulating not just new memories, but also like suggesting to you, hey, maybe you should add this additional skill or update or tweak the skill, or even like maybe you should go and actually try changing my agent system prompt or give me access to different tools so I can do this type of job better. And best yet, like we actually have this concept of what we call rubrics, which is exactly what it sounds like. It's like an eval rubric. And what you can do with rubrics that's really powerful is actually like define what does good look like for a certain type of task, right?

Mm-hmm.

So I could create one here that's like, what is a rubric for great Greg Isenberg content? And what it basically does is I can then have a full eval loop where every time my agent runs, like once the Greg Isenberg skill is ready, I could say like, I'm creating the virtual Greg agent and I'm gonna pin a rubric to that agent that then says every time Greg creates a piece of content, I wanna score that content along the dimensions that you care about. Using a separate LLM as judge that fires off. And then I can literally oversee like how well is my agent doing over time, right? And if I want to double-click in and inspect any one task run to see like how did it get scored, I can do so. So we basically, you know, have this complete full loop of it's not just like you get a day one agent or thread experience that works really well out of the box. And it's not just like you can curate agents and deploy them and like improve them over time. But it's that you have this complete observability layer and kind of this orchestration story where you can actually just like look at all of your agents running all the time and see how they're doing. And so if I pinned the eval rubric to any one of these agents, I would see like the trend line of how it's scoring. I could then automatically like suggest, hey, maybe I can reduce the model quality. So I drop from Opus to Sonnet, get a 5 times reduction in cost, And the score didn't go much down. Once people actually start running agents at scale, these secondary capabilities become really critical because it's not just about, can I get one agent to do one thing, but how do I oversee and run an entire business with many different agents and ensure consistent quality?

Which is a big deal because, for example, if you're using Manus, who is the judge around the output? The judge is you. A human being, right? It's not Opus 4.6. It's a human being. So if you're trying to actually create what we were talking about before, which is like an agent-first business, you know, managing a ton of agents, you're realistically, you're not going to have the bandwidth to be looking at every single output at all stages, right?

Yeah. It's kind of like Management 101, right? But like applied to agents now where it's like, as you scale up, if you're the CEO of a business, like you just literally don't have time To go and look at every single thing that every single person in the company has done. And so you need to create better automated checks and balances to oversee what the agents are doing and inspect quality of work. This would be like if you actually had a giant army of human content creators, you would want some way of, in a scalable way, to detect if they're posting good or bad content or not. And then know, okay, we got to tweak the guidelines for each of these people. Okay, so now we have the Greg Isenberg Contrarian Draft skill, and I'm going to go ahead and save this skill and I'm going to try seeing like, okay, let's do a dry run. It's going to scan today's AI and news and trends and then create some contrarian drafts. Right. And the whole idea here is like, look, like it's probably going to do an okay job on like the first effort here. Like it did some research about you. It kind of like, you know, it has a lot of like context about how you work, right? And if I wanted to see more about this skill, I could actually open it up. Here's when it should be used for. Here's the actual kind of skill contents. Greg's voice is a smart friend at dinner saying the quiet part out loud. Not a corporate communicator. I would agree with that. You know, you've been inside all these companies, blah, blah, blah. Like, doesn't mean be a jerk. I think it's very astute. Like, you're loud, but like not annoying or like You know, kind of rude. And then, actually, I'm curious if you agree with some of these stylistic things, right? Like, you got a hook in the first 7 words.

Yeah.

You know, you don't want, like, long blocks of text, which I'm guilty of. So I should take some of this Greg skill and apply it to myself. You love ordered lists. Never end with, what do you think? Which is super generic. So let's just say, like, this is a pretty good V1. Like, maybe it's, like, 50% of the way there. But the idea is that these skills should be evergreen. It's not like you do one and done. The whole point is every time I use a skill, either automatically using the LLM generating learnings and suggestions to improve itself, or because I am looking at the content and saying, oh, that's not quite right, here's why you got that wrong, you can interactively tweak and improve the skills and performance of the agent over time. I think this is the challenge that a lot of people face. They one-shot something, it's not quite as profound as what they hoped for, and they kind of give up. I think my strong guiding and urgence to folks, and I think this is very aligned to how you've thought about it, is don't give up after the first shot. Because it's very, very clear that the agents are powerful enough to do almost anything you want it to do. And the issue is not whether it's capable of and whether you should like give up on it. It's whether you are able to invest the kind of time and coaching and like curation to get it there. And I think that like it is well worth it, right? Like if you get it there, it's obviously going to be so much leverage for you that like what's the value of like having an always-on now employee that just like does the things that you care about like behind the scenes at all times? And like, you know, runs for trivial costs relative to like the cost of hiring a new employee.

Well, it's like real life too, which is like, you know, when I first started playing tennis, I was bad at playing tennis. And when I, you know, would go to play tennis, I almost didn't want to go because I was like, I'm bad at this. But you sort of, you go through the messy middle and you get better and better. And over time, then you end up Wow, this is a lot of fun. So I think that once you get to the point where it's a lot of fun and it does feel like the outputs are really good, the truth is 99% of people are not putting in the work to get the great outputs, right? So this is the arbitrage. It's for people to actually invest in spending time to optimize and get it to a place where it's high quality.

Absolutely. It's funny. One of the Benchmark Partners sent out this memo about— it was basically a wake-up call to all of the portfolio companies to get with the program and really radically rethink how you operate your business immediately with AI. The assumption is you think you're doing some things for AI. You have an AI center of excellence. You have this AI thing. Feature, but it's not enough, right? And the kind of parable that they ended with was like, imagine there's 2 friends back in, call it, 2003, and they're both going door to door selling knives, right? Or some other kind of in-person, offline product. And one of them decides, Every night and weekend, I'm gonna spend like 30 minutes like trying this new Google like AdWords thing and trying to like get some extra leads for my business as supplementary. And you know, like one month, like they grow a little bit of revenue like from from the SEO or the SEM thing. Next month, they get a little bit more. And the other person is like, this thing is awesome. Like SEM is awesome, and it's early, but I need to figure it out. And so they stop going door to door. and selling knives at all. And they just spend like the next few months like just focused on like, how do I get this entirely internet business to work, right? In the early days of it. And like, you know, 2 months, like they have zero revenue. They're like living off like their savings, but they slowly start to get this thing to start get humming, right? And they get like really versed in the best of SEO and SEM techniques and how do I create an e-commerce, you know, kind of, you know, website that like allows people to transact directly there versus like, just giving them a number to call me. And, you know, the end of the story is like, okay, like project forward like 5 years. Where do you think each of those people is? Right? And like the obvious answer is the second person has probably built like one of the early multi-billion dollar e-commerce businesses and just like carved off like the next Amazon. Right? And the other person's like probably still selling door-to-door, which is getting harder and harder. And like, you know, kind of that market's shrinking. And so I think it is one of those things where it's like you kind of have to like hit a reset moment, and what feels like, you know, maybe experimentation and not actually bringing home the bacon actually is the most profound thing you can do to create like real business leverage in the, like, not even like 2-year time frame, but like maybe even like the 6-month time frame. And I'm curious, like, in your experience, or when you see like, uh, solopreneurs doing this, like, where do you see, or like how often, like what is the average like break-even point? Literally, either in terms of like you get to the point where you can like self-sustain a full-time, you know, kind of like business, right? Like, and that becomes your paycheck, or just even where it like even feels like it's starting to pan out.

I think that there's like multiple milestones that people hit where they, you know, it's a game of confidence, you know? Yeah. When you make your first internet dollar, no matter what it is, it rewires your brain.

Yeah.

So if you can take an idea and make $1 a stranger, just $1, it's gonna rewire your brain. Then I think once you get to like $10K a month, just something about that number, you know, for the most part, once you hit that, you're probably quitting your job, you're probably going all in, you're probably like, okay, there's something here and there's a path to something bigger. I think that with respect to like, agent products and products like this, you know, the mistake I think a lot of people make is they try it too sporadically. So what I encourage people to do is to actually try the product, you know, every single day for a certain amount of time.

Yeah.

So commit to 30 days, 60 days, 90 days, some amount of time so that every single day It's like in your calendar, like literally I have in my calendar like 30 minutes here, 30 minutes there, right? Yeah. And that's what gets you to be a top 1% agent builder, right? Because you make it a part of your workflow. And then you end up seeing like, you know, outsized returns because it compounds.

That makes sense. I mean, it's kind of like, you know, I'm not a writer, but like, I've heard from writer friends, like, the most important thing is not to like wait for like the one weekend where you're gonna have like the spurt of brilliance and write the whole screenplay or the whole book all in one go. go, but it's like you have to force yourself to write like some pages every single day, like no stopping. Like some of them are gonna be crappy pages, but like the forced habit like just gets you better and better and better, and then it becomes like natural. And so I could see that being very, um, applicable and kind of like analogous here for the world of like getting agent savvy.

So do we have some tweets?

Uh, so okay, let's look at this. Um, Let's see. The consensus narratives are— oh, this is not loading for some reason. The consensus narratives are getting louder. Every Medium post reads like the last one. Okay, so here's one. The $10K a month AI solopreneur boom is mostly content farm fiction. They cite 82% of US businesses have zero employees. What do you think about this one?

I mean, what I like about it is when I do tweets, because I'm a human being, largely there's no data. It's just like, I have a hot take. So what's cool about this is there's research. And the truth is, people obviously want data associated with their tweets.

Yeah, maybe with a team of hyperagents doing all the research for you and coming up with content ideas, now you have time. Oh, this is kind of cool. Is this true that MedVee is actually not Yes. A legitimate business. I'd followed the first arc of that story, which is, oh my God, this thing is so massive. But I mean, it's a little letdown for the billion-dollar startup story. But maybe there's a take on it that says, no, but it's still possible for real. This guy just gave us all a bad reputation. Your AI agents didn't replace your VA, blah, blah, blah. It's kind of interesting.

These are all what I would call kernels for really good, great tweets.

Yeah, BYOP. And the cool thing is I could give it feedback. As an example, let's say I want to give you feedback on your skill. What's one thing that you want to give it some feedback on?

I would say, you know, the tweets that tend to do well are— sound very friend to friend.

And is there like, do these all just feel like a little too like, like they're not like colloquial? They feel like— Exactly. Yeah. These feel a little too formal or like stiff or something. Exactly.

Yeah, and that's something like I would notice that, right?

Yeah.

And so what we can do, like we would put this in the eval, right?

Yeah, you could do both. So one is like you could immediately go and turn this like, or update the skill based on this feedback. You could also have it immediately just like turn around like a new draft of these tweets, right, to sound more colloquial. And then finally, to your point, I could go and create a rubric that actually says like, okay, like Here's the 5 dimensions I care about, and then auto-evaluate every future output. So you kind of have a number of different options depending on how far you want to go right now. If you just want to get your job done right now, you don't want to bother with rubric, you don't have to. But eventually you get to the point where you want to set up a scalable system for this to just constantly work and get better and better. And that's the point at which you would do a rubric, which is not that hard actually. either go in through the UI and build one, or you can actually in this chat say, help me build a rubric to score great Greg-style content, which I'll queue up for after it updates the skill. And then it will go and help me create that rubric, save it, pin it to this agent or to this skill, and then automatically run every future time I create content.

Is it possible to For example, get an email every single day at 8:00 AM with some ideas?

Yeah, you absolutely can. The way to do that would be, in fact, you could just tell it in the thread, can you turn this into a recurring daily email at 8:00 AM? Then what it's going to do is say, I want to now save this thread into an agent and the agent is going to be given a run schedule of every day, 8:00 AM, go and do this thing. We're actually about to ship something that we're calling Alive Mode, which is kind of inspired by the OpenClaw heartbeat behavior, where you could already have configured an agent to do this just by saying, I want it to poll every 30 minutes. But we're making it much more of a first-class thing within HyperAgent, where you can literally just click a button, turn any agent or any thread alive. And then the feeling is going to be that, wow, this thing is just constantly on and looking at all of the new tweets out there, coming up with new ideas, and then pushing them to me. either via Telegram or over email or in Slack whenever it comes up with new stuff. So like the UX or the mental model is meant to be like, wow, this just becomes like an always-on, like 24/7 agent that pushes ideas to me or even like can go and like preemptively draft and post content. Like if you wanted it to go full YOLO, you could actually have it just go and like tweet the content itself, right?

Good old full YOLO mode.

Yeah.

Yeah, I don't recommend full YOLO mode just because, I mean, there's no need for something like this, right? Like, in order, for X specifically, in order to win, if you can get one good tweet out every single day, that's all it is. No one, you know, and That just means that you could, and you can batch these, you can schedule it out, but just look at it, make sure that it's high quality, meets your bar. I think it's definitely worth it for this specifically.

Yeah, that's fair. I mean, I think that content is a very hits-driven business. And so like fewer high-quality hits is what matters. But there are tons of use cases where Like maybe for my own emails, right? Like there are a subset of emails that like are low stakes that I just, you know, want HyperAgent to just automatically not only draft a reply, but like if it feels confident, it's like not a sensitive, you know, kind of situation, like, you know, then just go ahead and like respond to, to it, right? Like, you know, it could be simple, like inbound inquiries from like internal folks saying like, hey, when you have time to meet, it could just preemptively go ahead and like. suggest a time, right? Or even like pre-book it on my calendar. Or customer emails that are like innocuous or like asking for like, or trying to give input on a feature. It could just compile all that feedback for me as a report, but then respond like with a smart personalized acknowledgement to the user or even ask for like clarification.

And I think you all have like a ton of connectors built into HyperAgent, right?

Yeah. So what's really cool actually is that not only do we have a ton of connectors that just work out of the box. You click a button, OAuth in, in the thread, right? So maybe starting a new one, I could say like, what's a tool that you want to use with Hyperagent? Could be like Runola.

Notion, maybe.

Okay. Yeah. Connect. Can I connect to Notion and pull in all my notes? And so it will just in the thread like say, hey, here's an OAuth link, like connect to your Notion. But arguably one of the most powerful parts is like even for things that we don't have a connector to, like let's say there's some like very obscure API that you're trying to work with, right? You could basically have HyperAgent go and learn that API. So I actually I'll say like actually never mind on this. Can you instead help me build an API integration to what's some like Fairly new tool that you know of that has an API.

I'm assuming— well, do you have Linear built in here?

We do have a connection to Linear, but actually maybe Twilio could be a good example, right? Like where I don't know if you can OAuth into Twilio, so it has to be an API skill. And we may have a prebuilt connector, but I'm gonna have it like build a custom skill regardless. So can you still help me build a custom skill to integrate with Twilio via API, right? And so now what's going to happen is it can go and research the Twilio API docs, create a skill for itself to use the API, and then actually ask me to enter my credentials in a safe way and then be able to use the Twilio API fully. I think the powerful thing now is a frontier agent should be able to literally do anything.

Right.

But it's just a matter of, like, you have to give it access to the right context and you have to, like, you know, tell it like, hey, like, yeah, you should build a skill for this so then it can do it every single future time effortlessly. Okay, let's say what we want to do, SMS, voice for now, maybe phone numbers. We'll do an API key auth and any specific workflows. Think like Maybe actually I want to build a voice and SMS service that can call restaurants for reservations or something, right?

If you're listening to this and you're not fired up about building a business right now, like, the fact that you can do this is crazy.

Yeah.

If someone has heard about, you know, this is the first time they're hearing about HyperAgent, they want to, they want to get started and they What's a plan for them? What should they do? How do they get started? How do they get the most out of HyperAgent?

I think often the hardest thing to get over is not how to use the product. I think our users have said, wow, this product is super intuitive. I can usually just ask the agent to figure something out and it goes and does it. So it's not like I have to learn a ton of new configuration or UI or anything. I think the hardest part is actually picking the right problem or the right business opportunity you want to try to attack with HyperAgent, which HyperAgent actually can help you brainstorm that. In fact, we just shipped a new, better onboarding flow where instead of just landing you into a generic empty canvas where you have to just pick a new thread and we have some templates and so on, now when you first land in, it's going to suggest like, hey, do you want to connect me to all of your contacts? So connect me to your Gmail, to your Slack, and to your Notion and Granola. And what I'll offer to do is actually go and research you in your context. So I want to read through a bunch of your past week's emails and Slacks and look at your past Granola meetings. And of course, all that context is private to you. But now HyperAgent is going to be able to suggest to you Hey, based on everything I've learned about you, like, here's some use cases that might be relevant to you. So it seems like you're a VC, maybe you're like doing a lot of deal flow. I could create an agent to just go and automatically, like, you know, kind of summarize and do research on every investment pitch that you get, right? So, um, like, you can turn me on all the time, like, I'll just run in the background and then, like, ping you, uh, every single time you get an inbound pitch, or you can even have it learn the behavior. to thread a private reply to any email that you get inbound from a founder. So you get an inbound pitch, HyperAgent on behalf of you sends you and only you a just threaded reply within that email chain saying, hey, I researched this company. I also summarized all their materials. Here's what you should know about them. But the whole idea is that HyperAgent itself can help you identify use cases, or you could come in just with a really broad prompt like, Kind of interested in building a solopreneur business. I don't know, I'm kind of interested in like real estate. I want to pick one of Greg's, you know, kind of ideas that are open source. Like, help me plan this out, right? And it will do a very good job of like going and running with you on that. So I think the main thing is like, don't get stuck in the blank slate starting point problem. Like, just come in and like, you know, figure out some place to start. Maybe it's your personal, like, you know, contacts. Maybe it's like you come in with an idea. But like, once you start getting into it, like, it's, it just sucks you in even more because like you realize all of what you can do. And it's just so powerful. Like, you won't help but to get better and better at it.

Last question before, before we head out. You know, I was just talking to someone on the, on another pod— on actually this podcast talking about Hermes Agent. And one of the things we were talking about is when you're picking one of these platforms, be it OpenClaw, HyperAgent, Codex, whatever, you're sort of like investing in an ecosystem. My question for you, Howie, is why should someone invest in the HyperAgent ecosystem? Like, where do you see HyperAgent going over the next few years?

Yeah, so We have a lot of experience building great PLG products. I mean, obviously Airtable itself is a PLG product that also scaled up into real serious businesses. There are companies that still run their major operations, whether it's really, really large Walmart-scale companies, like the OpenAIs of the world, but also we have like really innovative, fast-moving SMBs, some of the like fastest-growing companies like Mercor run a lot of like stuff on Airtable. And, you know, I think like the experience that we have of building a product that's both extremely low floor and intuitive, but then also has a very high ceiling and scales up even as you need to scale up the number of agents you have, how you deploy them, how you oversee them. Like that's our commitment is that we are going to be the best at giving you both a low floor and a high ceiling, especially as you want to actually run a serious business or operations with HyperAgent, right? So I think that's, that's going to be kind of unique where I see the landscape fragmenting into like there's going to be really easy, fun kind of prototyping tools and products that are kind of like easy to get started with, but then ultimately don't scale with you as you want to become like a real serious enterprise built around these agents. And then conversely, there's going to be more heavy agent builder products with configuration and controls and all that stuff that are going to be better from a control plane standpoint, from being able to oversee a fleet of agents standpoint, but make the initial experience and the graduation path a lot more clunky or just a really sharp wall to overcome. So I think our commitment is this product is going to be the best combination of low floor and high ceiling. And we're always going to have this obsession with great UX. Like, that's our DNA. That's like what I obsess over. And the only kind of company that I want to build is one that wins in a product category where the value of the software or the technology is very, very high, but the accessibility is really kind of the key differentiator that we win on. Right. So Agents are going to be powerful. We're not going to be the only powerful agent product out there. I think frontier agents are all going to get better and smarter and faster and so on. But what we can do is use really great product design, just like Apple did with computing, to make the powerful experience also really accessible.

Yeah, it really is the most— HyperAgent is the most visual agent builder I've ever seen. It reminds me of a desk. Like, when I'm looking at my desk, it's a wood desk right now, and I've got— I have a paper over here and some scribbles over here and my iPad over there. To me, that's what a hyperagent kind of feels and looks like.

Yeah.

It feels like a desk that I'm visualizing. Um, so I think for people who, like, you know, connect like that, and I'm certainly one of those people, I think a lot of people are just gonna be like, Sign me up.

Totally. Yeah, I mean, look, like, you know, for people who don't like UI and want to just like use their computer through the terminal like all day every day, like, well, some people—

yeah, Howie, some people are like, they're, they're, you know, um, they're obs— what they love doing is like obsessing over tuning every single detail and stuff like that. And those people, though, you know, That an OpenCLAW might be for them.

But I believe that you don't have to sacrifice the tunability or the power. One of our strong design philosophies here is that Hyperagent still does give you a lot of control. You can go and tweak agent configuration if you want to. If you want to choose the exact model and system prompt and tools and give it a lot of refinement, You can go quite far in terms of curating memories. We actually just shipped yesterday a defrag tool for your memory so that as you accumulate more and more memories across all these different agents, you have this really elegant way of defragging them where we can auto-suggest here are related memories clustered by both keyword as well as embeddings similarity so that we're actually understanding the content of the memories.

And you can consolidate them.

But we want to really serve both people who are power users who want control over how the agent is set up so they can get maximum bleeding-edge performance. But then also, you shouldn't have to do all that to get value out of the product. So it really is about the range. I think it's more just that if you are truly happy just doing it all yourself through a very low-level command line interface kind of experience and you're okay not having the control plane, the deployability, the ability to oversee many agents and deploy them at scale and manage across a team, then maybe those people aren't going to appreciate HyperAgent as much.

Totally. Well, I'm stoked to see how it evolves. Thanks for doing a little show and tell. You got me fired up. Howie, I'll include links where to follow you, but also where to sign up to HyperAgent in the description, in the show notes. And we're going to do a really generous credits giveaway for your listeners.

I mean, one of the benefits of launching HyperAgent within Airtable, which is a half-billion revenue business, we're going to generate $100 million of free cash flow this year. We have over $1 billion on our balance sheet. that's not to like, you know, just be pretentious about it, but is that like, you know, we've built a good and growing and like, you know, kind of profitable business with Airtable that allows us to be even more generous and liberal with like, we just want to get people to really adopt HyperAgent, get value out of it, and we want it to become the standard, right? Like, we want it to become like the iPhone. And so, you know, we're willing to be very, very generous. Like, we're not trying to make money and nickel and dime people. on pricing. In fact, we're giving away multipliers to your audience and early adopters for both just straight-up cash that gets applied towards real model costs, including Opus, which now, as a lot of the OpenCLAW community has gotten kind of sad about, you can't get subsidized credit for use in OpenCLAW. But You can use Opus, you can get the frontier models, and you can get it much more cheaply because we're willing to subsidize it through HyperAgent.

Well, this is a group of people who are listening to this who appreciate that because this is a group of people who actually listen and actually go and build stuff. So thanks for the love, Howie.

I love this, the solopreneur and small early-stage startup and small business owner audience. I think it is where more AI innovation is going to happen far faster than frankly within many large incumbent companies. You just have the agility and the only thing keeping you from going and deploying agents everywhere is just your willingness and putting in a little bit of time. But we're already seeing in our early adoption base with HyperAgent, you know, some of these like small shops have become super sophisticated really, really fast and are running their operations in a kind of game-changing way that frankly, like a 50,000-person company would not be able to do for a much, much, much longer time. Right. And just has all kinds of like, you know, kind of reasons why they wouldn't be able to go and pivot on a dime. So I think this is a really, really awesome audience and, you know, I kind of lived to see You know, entrepreneurs like do awesome stuff, right? Like, so super exciting to be plugged into the community. And like, I want to see, you know, your listener base generate like, you know, $100 billion, you know, kind of legit companies with like less than 5 employees.

From your lips to God's ears, baby. Thanks a lot, Howie. I'll see you next time.

Awesome. See you.