Why Howie Liu exited Airtable to go all in on Hyperagent

Every software engineer should be using agents every single day, every minute.

I think the best practice now for harness engineering is to keep the harness as minimal as possible. Kind of let the model take charge and be as open-ended as possible. Where do we go from here? What's next for Hyper Agent? Will it be running the company at some point? Like how far can this thing go in really being powerful?

Yeah, so I have this like kind of macro thesis which is, you know, while if you look too closely at like model progression on a week-by-week basis you kind of lose the forest for the trees, right? And it's like, you know, is GPT like, you know, 5.5 like that much better, you know, or not? And I think the point is, you know, if you zoom out like there are these kind of big step function changes that, you know, if you really kind of step back and think about what you can do with the models, you know, is quite different, right? Like there's a step function inflection. You know, my thesis is that like the product form factor needs to structurally and quite dramatically change with each of these sub-functions. So, you know, tangibly, like we started with basically LLM, you know, kind of response completions, right? So you give it some like string of text, it completes, you know, kind of the rest of the string. And before InstructGPT, like it was only like kind of modestly useful, right? And then InstructGPT came out where like it became more and more of a natural interaction style, a chat, you know, bot. And so chatbots come out, that's like the first killer form factor of AI. But of course, like as we're now talking about, like, you know, chatbots are no longer the right form factor, right? As the models became more and more capable, like, you know, it enabled an even more autonomous form of agent, right? And so, you know, going from like the original, let's say like precursor, like GitHub Copilot autocomplete experience, right? Like that was more of like completions model. And then we went to like Cursor 1.0 for development where you could go and like talk to this agent composer, like their original composer agent, not like the model agent, but like, you know, you could go and basically talk to composer and have it do, you know, much more complex changes, right? But it was still very, you know, mostly a synchronous interaction, like you would tell it to do something, it would do some stuff, it'd be like a minute or two max.

But now we're at this point where you have fully autonomous agents, right? And so I think for me the big like kind of light bulb moment was, you know, really with this latest generation of models. And, you know, I think Opus 4.5, as many people discovered, was kind of the first aha moment model that kind of like broke the dam for this new era, like was so capable that you could just set it off to do like a really really open-ended and long-running task and get pretty good results, right? So it was the first model that to me felt like it was, you know, more or less like AGI capable, right? Like it's capable of doing work that truly would otherwise feel like you'd have to find a human to do it. And so I just was very enamored by this model. Like it was exciting to play with. I think like we all kind of had this like model psychosis, you know, or agent psychosis moment where you would just like, you know, set off your agents to work and like be mesmerized by what they could do and how much they could do. You know, I wanted to go as meta as possible in a way, like, you know, mirroring what I did with Airtable, like, you know, if I thought like building, you know, database-powered apps was really cool, like what's the best version of that? Well, it's an app to build database-powered apps, right? Like kind of the meta app platform. And I think like with agents, you know, I was really excited working with these frontier agents. And the most meta thing I could think of was like what if I could build like a meta agent platform where anyone can go and, you know, kind of configure and build and deploy their own agents. So that was kind of the genesis of the Hyper Agent product within Airtable. But really just kind of was born out of like a personal, you know, kind of passion and like, you know, perhaps product, you know, kind of curiosity around what was possible in the space.

Yeah, so I mean the founding vision for Airtable was always to democratize app creation, right? And so just, you know, we had this strong and still, you know, present belief that like building application software is just like a fundamental need, right? Especially in a like increasingly digital economy. And so instead of having to like accept that every app you use, like CRM, inventory management, etc., is going to be off-the-shelf by somebody else who, you know, tries to build a vertical solution for your needs, everybody, every customer should be able to go and create their own custom applications, right? So if you want a CRM, you want it to be lightweight and fully configurable, like you should use Airtable to do that, right? Like if you want to build like a content production pipeline, you should be able to do that. And so that's been the vision of Airtable. I think it's still constant today, right? And if anything, like I think the fact that now humans alongside agents are there to both create the applications, the building step, but also to use the applications and arguably need even more than ever a source of truth. Like for the same reason that like humans need to centralize, you know, kind of a source of truth for whatever work we're doing, right? If you're a developer, you need like a source of truth for your literal code, right? You need like something like git source control, but you also need a way to like manage your issues or your product roadmap. You know, I think agents also really benefit, especially as you scale them up and they start becoming more and more autonomous, like they all need to share a centralized source of truth. So that's kind of really the continuation or the evolution of the Airtable vision. It's not a radical pivot, but it's really just kind of like progressing that same, you know, constancy into now an era where agents and humans need to coincide around a system of record.

For sure. Well, you know, it's funny, like when we went out to pitch Airtable in the early days of the company, we were a pre-seed company, you know, kind of relatively nobodies. You know, the number one piece of feedback we got from every investor was, look, like you seem like a bright team. You know, you've built like kind of a cool product MVP. It's clearly very technically sophisticated, but it is impossible to go and like get product market fit for a horizontal platform, right? Like you're saying you can be this app platform for anyone, right? Every use case instead of centralizing on one.

Exactly. So that was like the number one consistent piece of feedback we got and we kind of defied the traditional wisdom. Like I think, you know, I think you do need to like have some kind of, you know, secret in Peter Thiel's terms or like just like a non-conventional way about thinking about the problem to win. So that was kind of our unique take on the space was like, look, like I don't think we have to go and revert into being a vertical solution company, like build like dental CRM software, which personally wouldn't have excited me nearly as much as building a horizontal platform. But I think instead, like I looked at a different dimension for Airtable, which was, you know, while we're not going to be able to like gain the benefit of just like vertically marketing and like building the functionality for this product, what if we just like focus on making the barrier to entry? Like the reason why horizontal products typically have a struggle is because like the user has such a hard time grokking like what they can build on it and then how to build the thing they want to. And then even if they know how and what to build, they might still run into limitations if your product doesn't have the right building blocks to do so, right? So there's kind of three extra hurdles for a horizontal platform versus if you've built the vertical solution, like dental CRM, you know, it works for that use case and the customer, right? And yet, you know, like to me maybe like I just love like solving hard problems and, you know, wanted the challenge, but like those three problems actually can be distilled into like a product UX and design problem, fundamentally, right?

Like, you know, you think about like the advent of personal computing and it was a similarly, you know, maybe even like more ambitious horizontal problem to solve, like how do you create the right GUI or like the right operating system metaphors so that anyone can figure out any personal computing interaction, right? All the different types of apps and interactions you might want to do. And so, you know, we were very inspired by like the origins of like personal computing, Apple Macintosh, like Xerox PARC, etc., and applied a lot of that same design philosophy into distilling every single part of app building into the most, in our opinion, like intuitive form. And I think that's really what it came down to. And it turned out like it was just intuitive enough to get a large class of builders who were not technical, not necessarily your typical, you know, kind of Power Apps admins or like, you know, people who would have built with like an existing app platform or a Salesforce admin, but instead just like, you know, people in their business, right? It could be a small business owner, it could be, you know, an operator within, you know, a function like marketing, but they were able to just go in and build it because it was intuitive enough. And I think solving for that untapped demand or that underserved demand with a product that for the first time enables the people who know what they want to do but are prevented from doing it because they haven't had the product that's like intuitive or delightful or capable enough to do it is a really fun kind of product design challenge. So in many ways, like Hyper Agent is applying that same design philosophy. Obviously the actual building blocks are different from a relational database with, you know, business logic on top of it and interface customizations. For agents it might be more like, okay, what kind of system prompt do you want? What kind of model do you want? What kind of behavioral guardrails do you want to impose? How do you want this agent to be invoked, right? Do you want it to be always on listening in Slack? Do you want it to be explicitly invoked? Which tools should it have access to, what integrations and so on. But like there's kind of a very very parallel kind of design problem to thinking about these different building blocks for agents as we applied to app building. So it's really fun and kind of a same adventure but like applied to a new playing field now.

Yeah, absolutely. I think, you know, to me OpenClaw was a great demonstration of what frontier models were capable of, right? Like the kind of default and like, you know, kind of magical experience of OpenClaw is that you just give it access to all your stuff on your computer, right? And people were doing it on their actual like real machines, right? Not like a separate like kind of sandbox machine, but like they would just install it on their local machine or locally on their machine, access to all their credentials. They would have like full unadulterated ability to do pretty much anything you could, right? And it's that very fact that like the models were now capable of performing these really long horizon tasks. People would like tweet about how, you know, they had their agent overnight like, you know, try to find them a really hard to get reservation at some restaurant and overnight the agent to do so not only figured out how to like, you know, navigate the website of the restaurant, figure out like what days were available and get the number, but then like would write a script to maybe execute against the Twilio API and perform an actual voice synthesis call, like synthesize a voice to have like a voice call to like a real human and negotiate with a human to get a reservation that otherwise wasn't available. It was these kind of like sensational stories of like how far the agent could go when it was given a very clear goal but then like a lot of like open-ended capability, like very few guardrails and lots of like basically token bandwidth to go and achieve its goal, right? And so I think like, you know, that kind of experience could not have worked with prior model generations, right? Like if you tried it a year ago on older models like it just would not have worked, right? Like we actually saw what happened when you tried to do this with like AutoGPT and BabyAGI, like there were these kind of like experimental early efforts to do something very open-ended but the models just weren't there. So to me, like OpenClaw was a really great almost like spec implementation of how can you like most take advantage of the long horizon autonomy of these frontier models, and yet, like to me, it was mostly just that, right? It was a cool demo but it's obviously really really kind of limited in terms of like it's missing the necessary kind of behavioral and policy guardrails. You know, it's kind of a hard experience to set up, right? Like it kind of feels like, you know, back in the day I was a nerd, you know, I would install like Linux onto my computer or even like install multiple different variations of Linux and like partition my hard drive still.

But, you know, like it felt like it was very very hard and clunky to do that, and that's kind of what OpenClaw feels like today, right? Like to do it, like period, let alone safely, requires a lot of like technical knowhow and kind of like being a hacker to make it work. If you can do it, it's awesome and it can do really awesome things, but like it's very, you know, kind of fraught with like a lot of complexity and error and potential for like errors and security holes. So, you know, in many ways like Hyper Agent is to OpenClaw like what maybe like a Macintosh is to like a Linux box, right? Like we wanted to just build like a cloud-native, secure-by-default and also just like a beautiful, delightful UI around all of the above. And, you know, because we own kind of the entire verticalized experience, right? Like, you know, we get to, you know, define how memories and skills work within our system. We get to define, you know, how tool calling works and which tools we get to use, including Browserbase. But like we get to control that entire experience in a way that makes it really really effortless and allows us to add new enhancements. It's like one of the big things we're working on now are the kind of policy and behavioral guardrails that allow you to set free like an agent to be increasingly autonomous and just run on its own like all the time and do, you know, pretty open-ended things while also having the right level of guardrail so it doesn't just do something crazy or get prompt injected to leak data or secrets.

Yeah. So, yes, you know, I built the MVP of the product like over the holidays. You know, I think a lot of people, a lot of especially like AI-excited operators and founders, like, you know, just spent all of the holidays into New Year's like building with especially Opus. And, you know, I think the reason I did that was, A, I just believe that like the frontier is moving so fast. The only way to stay at the edge is to go and like get directly hands-on with it. Like you can't read just like tweets about it or like get updates from your team. Like it would be like, you know, back in like 1997 not actually using the internet yourself and like learning about the internet and like trying to develop an understanding about the internet and how it could profoundly disrupt your own business by listening to other people, right? Like you just kind of have to use it, right? And so, you know, I think one, like you just have to be fluent with it yourself. But then two, I think like, you know, there's this very meta-ness to again like Hyper Agent in that, you know, it's not only a product that I built originally agentically. Now our team, on, you know, building Hyper Agent, which is still a very small leveraged team within Airtable, is fully agentically leveraged in building the product, but obviously it's an agent product itself. And so I think like the very act of building the product agentically and developing like best practices and really good intuition for what you can do with agents in the development of the product also lends to giving you like better product intuition for how to make the agent product itself behave really well, right? So the quirks and things we noticed with, like, you know, let's say using GPT 5.5 versus like Opus 4.8 versus like Fable 5, like we can now learn how to kind of productize that into Hyper Agent for our own customers. So, you know, I think it just kind of becomes like this very virtuous cycle where if you get hands-on, you get to be part of the flywheel. And ultimately, like I think, you know, if you really love thinking about products, like this is a new medium, like in the same way that like mobile versus desktop require just like an intuition that could only be developed through like actual hands-on, being a hands-on practitioner. You know, if you didn't use an iPhone yourself, it'd be really hard to design for iPhone form factor, right? Or like internet vs. pre-internet. And so I think here, you know, you're able to just express more product vision directly. And it's really exciting because like the speed between coming up with an idea, coming up with that vision, and then being able to actually express it in a real functional form has just, you know, compressed by like quite literally 10 to 100x versus before.