Hiring Employable Agents: Howie Liu on the Next Form Factor of AI

Please join me in welcoming the Chief Executive Officer at HyperAgent, Howie Liu. All right. Hello. Who's excited about the end of this event? This has been an amazing few days, amazing talks, amazing Gary talk right there. And by the way, I've never been introduced with so much bass. I am very, very pumped. This moment. Um, so I consider myself a product thinker at heart, meaning, you know, as much as I like to spend time on the technical side of building and, you know, I, I love to think about architecture and algorithms and so on, you know, really I'm a product form factor thinker at heart. And, you know, what I want to end today off with is a brief and kind of high-level overview of how I see the world going with agents as it relates to product form factors. So I'm calling this Hiring Employable Agents, and it's really a reflection on as these models have gotten better and better, and we've heard from some really, really great speakers about, you know, the latest frontier models, whether GLM 5.2 or others, and how to coax the most performance out of them. And really, this is not about the technical deep dive, but how do we think about these from a product building standpoint? So who am I? I'm the founder and CEO of Airtable, and also now HyperAgent, which is a product within Airtable. And as many of you already know, Airtable is basically an almost laughably horizontal platform. When we, in fact, set out to build the company almost 12 years ago, we got basically this negative feedback from virtually every investor we talked to early on, which was, you can't possibly go out with a horizontal platform approach and try to build something that is everything for everyone. Like, any type of app that people want to build, you can't solve all of those use cases. I think we ended up doing a pretty good job of solving for many of them at least and solving for a large builder community. We're basically applying a very, very similar philosophy to agents now. I'm going to break down how we're doing it. When I zoom out and think about the progression of models and therefore the product form factors they enable, I really see it as this spectrum. We went from basically completions, which is you give a model, one of these LLMs, string of tokens, and it just continues emitting more tokens until it gets to a stop token. More or less, it would just complete your thought or complete the pattern. Until InstructGPT happened and enabled a chat form factor, it was arguably powerful but more narrowly applicable. Once we got to a higher level of model intelligence and then also the innovation of InstructGPT and tuning these models to actually have more of a conversational back-and-forth behavior, with users. Then we enabled the chatbot, right? And of course, we all know ChatGPT was an awesome breakout product. We're now all very, very intimately familiar with that form factor of AI. The next level I call an agent, right? And, you know, I think agent has been very overloaded. Um, we've probably talked about that a ton over the past few days, but my definition of agent really is the same as the Anthropic definition, which is distinct from workflows. And an agent is really a model that recurses upon itself, right? And has, you know, the open-ended set of tools that it could call, of reasoning steps it can make, of decisions it can make. So it's not prescriptive to a linear workflow, but it can kind of make an open-ended set of choices. Still, you know, I like to think of agents as being a little bit distinct from the next level of agents, which I'm using the term CLAW here because OpenCLAW obviously has popularized this concept. You know, there's a market shift that I see from agents that were still mostly reactive, um, you know, to user input. So you would have an agent go and perform a task, it might spend 20 minutes, maybe up to an hour to perform the task, and then it completes. CLAWs, um, as I call them, are really, you know, about agents that go and perform work for much longer, right? They can even have a mechanism like the heartbeat mechanism within OpenCLAW, um, and now in other products like HyperAgent. That enable them to wake up on their own and almost exhibit this always-on behavior. Underneath the hood, what's happening is each wake-up basically results in many different turns. The agent loops upon itself. It's performing tool calls. Each tool call response then invokes the LLM again until it reaches a stopping point. Then that stopping point gets re-invoked or restarted with a heartbeat mechanism where you can wake up the agent again and again. And so this idea of this always-on agent that's actually constantly moving towards its goal and kind of figure out creatively how to unblock itself where maybe prior agents would get stuck, I think that's a really, really important kind of next form factor leap that we're seeing. And then finally, as you think about models that are now capable of, you know, really, really intelligent orchestration of complex tasks, right? So I've personally been using the workflows capability in Claude Code quite a bit, where it can fan out a lot of work to a massively parallel set of agents. But this concept of, you know, this agent that can go and orchestrate with other agents, right? An agent-to-agent interaction to perform even more advanced tasks, right? Longer running jobs is I think the next leap yet. All of this is enabled because the models are getting smarter and smarter to the point where they can achieve coherence even on longer and more open-ended scoped problems, right? Um, if you recall back to the days of, um, uh, Baby AGI and AutoGPT, those were really kind of fun, cool science experiments. But ultimately, you know, as much as they were really interesting to watch and, you know, see them work for hours and hours on a problem, they would all inevitably, you know, kind of end up drifting off to something not super useful, right? It would go off into a rabbit hole and kind of get stuck. And I think now the models are actually smart enough to enable not only longer running jobs, but actually enable us to start applying these kind of organizational principles to break down problems into differently scoped tasks that different agents can hand off to each other, all of which is about managing each agent's task, uh, or context window and ability to coordinate work with each other. So all this is kind of abstract. I wanna make it very, very tangible and show you how we've applied some of that design thinking of building a horizontal product and distilling a lot of the complexity of agents like we did for building a database or a database-powered app with Airtable for agents in HyperAgent. So this is a very realistic use case based on some of our actual customer usage. One of the interesting things we found from launching HyperAgent is, you know, the prototypical users for HyperAgent are not who you would think they are, right? I mean, we definitely get our fair share of like, the super AI-forward, you know, kind of pure software companies, like AI companies themselves. But we also get really interesting other companies that are, you know, sometimes like traditionally offline businesses. And in this case, we're depicting a very realistic scenario, which is a customer that is in a landscaping business and actually uses agents to run virtually end-to-end every part of their business, which is finding clients to propose landscaping projects to, like physically going and redoing their garden or their backyard, and end-to-end managing everything from the prospecting to building a quote and a pitch to managing the work itself. I think this is a huge, huge interesting insight that I've just stumbled on from working on HyperAgent, which is ingenuity of agent builders is not contained to just AI companies or software companies. I was really inspired by Gary's talk just a second ago. I think the motive behind that applies to every industry, every sector. There are really, really creative people, many of whom are in this room, who can go out and disrupt your own industry, whatever it be. It doesn't have to be a pure software business. I think all it takes is the ability to imagine what work could look like when you do go and become fully, not just an AI-native company, but I like to use the term fully fleet-of-agents-native company. Yeah. You're actually going above and beyond and not just thinking about small, kind of narrow steps to automate with AI, but actually how to employ a whole fleet of agents that have governed behavior and have the right context and the right tools to be able to do really useful and very end-to-end autonomous work. Let's walk through what this looks like. This is a realistic example of, in this business, this landscaper wants to go and pitch clients. They get all these inbound submissions of clients who are interested. Somebody submits this inquiry saying, hey, here's my current backyard. It's a little messy. Here's my info. Can you give me a quote and show me a proposal of what you could do? This basically gets routed into one agent that is the triager agent. This agent is basically taking in this intake and saying, hey, what is the scope of the project? Is this something we want to take on? And is this a quality lead? Is this a legit person? And so it does some research, but ultimately it hands off then the work to another agent, a surveyor agent, to actually go and then look at the, the physical landscape of this person's backyard and come up with a proposal of what a re-landscaped version of that could look like. So here the surveyor agent is going and looking at the original imagery or even video. that the, um, submitter gave them. Um, HyperAgent, uh, in our case, uh, works with a fully, um, capable sandbox VM. So every agent, in fact every thread with an agent, is able to fully write code. It can manipulate files. It can actually use things like FFmpeg to like, you know, clip out individual screenshots from a video file and then analyze those. Um, and actually go through and like find all the relevant information to create a fairly high-quality, um, and high-touch, uh, proposal that would have been unfathomable for a small business like this before, right? If you think about like a really high-end interior designer or maybe like a very high-end landscaper, maybe they could do a very bespoke proposal like this for like a $5 million plus, uh, client, you know, if they're working for a big hotel or something. But it would have been unheard of to do something this high-touch and actually mock up real imagery and a real client pitch deck for just a small backyard landscaping proposal. But this person has done it because they've figured out how to use agents in an extremely capable and autonomous way to do the job that otherwise would have been undoable by humans. What happens next is the surveyor agent, it went and created this proposal and pitch. Now, the intervention has to go back to the human. There is a step where humans still need to unblock the agents. Sarah Guo put out a great post a few weeks ago about, you know, just describing like where the value is going to accrue in this AI landscape as things evolve and models get smarter. And I firmly agree with her take, which was, you know, the ultimate blocker to still being able to deploy useful agents, there's still a human element to that. You know, as much as models are getting smarter, the ability for agents to get unblocked by humans that still own policies and decisions, and really importantly, just kind of gate, you know, really critical, high-impact actions, like in this case, actually making a pitch to a customer that has a binding quote, um, or perhaps like agreeing to do the work, that becomes something that the human then intervenes at, just like a great manager who's getting upwards, uh, reports or, you know, upwards work updates from their team to then unblock. So here we have Sage who has come back to, uh, the real human, uh, owner of this business and saying, hey, here's a lead. I already surveyed the, the project. Here's a beautiful pitch deck and a, um, proposal. And have come up with a realistic quote based on all the factors I've already surveyed in this, uh, particular project to give to, uh, to you to now approve sending to the client and see if they accept it, right? And you can see like even the, um, the proposal itself is this really beautiful kind of interactive, uh, webpage that again would've been unfathomable for a business like this to be able to send to every single client, even for small tasks, uh, or jobs like a $10,000 landscaping proposal. Right. You know, better yet, as you interact with these agents, you know, as Gary talked about with GBrain and just kind of like creating this very sticky context layer that learns with every interaction, I completely agree. I think that is a, you know, kind of a universal principle and kind of a form factor for useful agents going forward is agents should not be static. They should not just be, you know, either static LLM calls or static even agents that just have a prompt and, you know, some static skills. They really need to learn with every interaction, right? And you can't wait for the next, you know, kind of fine-tuning run of your own models where you have, you know, a bunch of collected data or rubrics to, to retrain the model. It should learn in real time, right? It should accumulate the memories and skill updates and just take user feedback and actually learn from real-world performance to constantly adjust and get better, just like humans do. And so in this case, you can see, you know, within Slack or, you know, in our case, within either Slack or email, or other invocation methods, you can basically go back and forth with the agent just like a real human and say, hey, like, you know, remember to take this into account next time. You kind of, you know, here's something differently formatted or a different way to think about the proposal, or you missed kind of this additional cost lever in this particular project. So, you know, remember that for next time. And I think it's critical that in this next generation of agents, agents need to remember these things, right? They should feel fluid and, and, evolving just like humans are in order for them to fully realize the potential that the latest frontier model capabilities enable them to, to actually go and do in terms of autonomy. Um, and then finally, I think what ends up happening at the end of all of this is with all of these agents that are each individually capable and even interacting with each other, I think the job of the human becomes increasingly about unblocking agents. To be able to do work effectively. And so, you know, I, I compare this to the development workflows that, uh, have evolved from, you know, originally before any, um, you know, kind of AI-enabled coding. Obviously, we just sat in front of a computer. We as humans are pretty single-threaded, right? You'd stare at one code file, you think about a problem, you'd kind of bang out some code, maybe take a break, come back to it. To then we had, um, augmented coding with the first generation of GitHub Copilot and similar tools. Where, I mean, I think of that as really just a completions form factor on coding. You type some code, it can autocomplete a few extra lines. To then we had things like Cursor Composer 1.0, like the original, uh, experience, not the, the composer model, um, but the chat experience where you could talk to this agent, it would do more complex changes, and that was even more leverage, right? But we're now moving to this point where, you know, the best frontier agentic developers I know, and many of you are sitting in this room, are, are really just kind of overseeing a fleet of agents, right? Like, in fact, going to sleep without setting off your agents to perform useful work overnight feels like you're, you're taking this huge loss. Like your, your team is not working because you're not unblocking them. And so the UX of how you oversee these agents, whether it's in development workflows or general purpose work like this landscaping example, needs to shift as well. And so the way we've thought about that is you need to see this kind of orchestration control plane where you can see at a glance what all of your agents are working on, what they're blocked on, who's handing off work to each other. And your job becomes almost like, you know, zooming out to see the entire SimCity landscape and orchestrating across everything going on at a macro level between your different agents, rather than going and having to play micro and zoom into each individual project or each individual task. Um, and so I think this is a really, really exciting moment as we have shifted from, you know, not just completions and kind of the slight augmentation of existing human, uh, work, but really a radical now leap into humans as orchestrators of agents that also can orchestrate themselves. And finding the right way for us to interact with those agents as they become more and more capable is gonna become not just the opportunity to boil the ocean, but in my opinion, table stakes. for survival for every company in every industry. This is going to be the way to get ahead and thrive in your industry. If you don't, as Jensen put it, it's not going to be AI taking your job. It's going to be somebody using AI who takes your job or takes your business. Really, really excited both for the opportunity and for everyone in this room to be part of this transformation and this really bold leap into the next form factor of AI. All that being said, we'd love to see you try HyperAgent. Hyperagent.com/AIE and you get $1,000 of inference credit. Um, you know, that can be spent on all of our different model offerings, including, uh, you know, the latest, uh, Opus 4.8, Fable 5. You can use it on GLM 5.2, uh, on GPT-5.5, et cetera. And we'd love to see what you do with, uh, with agents on HyperAgent or elsewhere. Thank you all for being here.