Agents, Metaprompting and Tech-powered Parenting with HubSpot CTO Dharmesh Shah

AI still cannot write even a decent original dad joke despite my best prompt engineering attempts. You know, I love AI as much as anyone. I have a huge amount of respect for the technology, what it's going to be capable of. My money is still on the humans. I'm sorry.

Thanks for having me, Cameron.

Wow. The multitasking is going to be hard, but I'll try it. So, I'll tell you the story. When I was in my teens, my parents visited Singapore for the first time. I didn't go and they brought back this electronic device which ended up like changing my life. Funny thing is when I tell people this story, which I haven't told that many times, it's like they automatically assume it was a computer and it was not a computer because I didn't get a computer until I was in my 20s. Much later. The device I did get though that they brought back, which I'll try to sketch out here.

Looks like this. So it's a tiny little Casio keyboard. This is probably like almost drawn to size. I had not shown any inclination for music. I'm not exactly sure why they got that for me as a gift, but I sort of became obsessed with it even though I didn't know how to play it at all. And I taught myself through brute force, which is randomly pressing keys and seeing if the notes sounded right or not, and then pressing one of the other keys until something sounded right. It would take me like weeks to learn a single song. And I've been doing that for about 30 years now. You know, ever since.

I've had this thesis that in the future we're going to have agents and humans working together in what we would think of like modern hybrid teams. I'm like, okay, well, if that's true, if we're going to be putting kind of AI agents on teams along with humans, then we're going to need all the things that we have in the kind of analog human world. It's like, how will people find these agents? What will their resumes look like? How do we know if they've had experience in all these things? And so the original spark was there should be a professional network for agents because that's going to need to exist someday. So, that was the original kind of genesis of the agent.ai idea. We launched in beta about a year ago at about 50,000 users back then. It's now up to 2 million. So, there's been some resonance with the idea. What's happened now is people are sort of building their own agents, kind of creating them on the platform and we have about 2,000 kind of publicly shared agents now on the platform. So, it's been a lot of fun.

One of my favorite agents, one of the most widely used ones is a company research agent where you can type in the company name or the website domain and it will do this kind of deep research from all across the web, hitting a bunch of proprietary APIs. But what's cool about it is just like if you had a researcher on your behalf, you would say, 'Okay, well here's the parts that I'm actually interested in. Here's the business that I'm in. Here are the questions that I have. So, I want you to answer these five questions, focus on these particular areas, and give me a customized report every time I ask you for a research report on a company.' That's one of them. But I've probably built a dozen or so agents from relatively sophisticated ones to really simple ones. And I'll tell you the one that I've been trying to build, which is like a dad joke agent to be able to generate a dad joke, which remains one of the last remaining frontiers. Despite all the reasoning models and all the advancements and I've played with all of them, AI still cannot write even a decent original dad joke. That's what I've discovered. Despite my best prompt engineering attempts.

Yeah, there you go. Yeah.

There's a spectrum. When we started the platform, we had a low-code agent building platform which primarily used deterministic flows that says here are the steps I want to take in a multi-step flow. Here are the inputs, here's a minimalist UI, and then it would call the LLM or call APIs and we have a bunch of data connectors to get social media data and pull all that in. And now we also have the ability to have agents that are built off platform. Pick your agentic or vibe coding tool of choice and build an agent that way and just wire it into the platform, which I've been spending a lot of time on. Building these much more sophisticated UI agents that still kind of plug into agent.ai.

Yeah, for some definition of creativity, right? I use AI for what I think would be classified as relatively creative use cases. They're not just data research. It's not just kind of writing or wordsmithing something. One of these tricks I sort of learned from my son, who's 14, is around kind of creating what I'll call a simulation for lack of a better term. So what he does is he's an aspiring fantasy author. So he'll do world building inside of ChatGPT. So, he's got like a 2,000-word prompt to say, 'Oh, here's the world. Here are the characters. Here's the power structure. Here are the things that people can do.' And then he'll pressure test the ideas like, 'What happens if I do this?' And it'll come back and it's this very iterative process. I do something similar with kind of business ideas like, 'Oh, here's what I'm going to tell you.' And I've got this kind of very long knowledge base that I can pull information in from and say what would happen if HubSpot launched a product X or we did this change to it or whatever. Or if I were to ask five people across different disciplines to give me feedback on this thing from different regards, like what would that be? What would that look like? In a way, creativity is sort of combining primitives that already exist in novel and unique ways. Not like we're kind of building things always from whole cloth. I don't use it for writing full articles or help with talks and things like that, but it's really good at connecting dots that might be hard to connect and doing like transitions or like, oh, I'm trying to wire this up and I can't quite find a way to get to this thing that I want to get to from where I am right now. Can you give me some options? It's a great critic. It's a great brainstorming tool. It's a great way to kind of just spark some ideas and add some dots to this kind of mix of dots you're trying to wire up.

One of the neat things you can do, it does it really well, is you can — so imagine you had a spreadsheet or just a two-axis thing, whatever, and you have your kind of products or product features down one side and the same product features along the top, right? So it's the exact same thing along x and y dimension. And then what you do is you run the simulations like, I want you to combine this product with this product or this feature with this product in novel ways and come up with a description of like what would that look like. What might be like, oh, we happen to have this and we also have that. Is there some unique blend of things? And some of them or a lot of them are not going to make any sense. But sometimes it comes up with things that are like actually useful. Like, oh yeah, I hadn't thought about how we could take this feature that's in our kind of sales product and connect it with this customer service product and that would be kind of cool actually.

Yes. Yeah. It's weird. Back on the dad joke thing. Not only is it not able to write dad jokes, well, it actually struggles with even when you tell — So, I have a database, which is what I call it. It's a database of curated jokes that I've collected over the years. And so, it can actually explain to you what makes a joke a joke and why it's funny. So, it actually does a very good job at that. But then if you have it rate the jokes on a scale of 0 to 10, this is — I'm going to give you 100 jokes, you tell me which ones and give me a quantitative score, you know, based on the joke, it does actually not that good at that. And I'm not exactly sure why. Even though it can tell you what makes the joke funny, like the underlying mechanics of it, it's not good at actually quantifiably kind of measuring the quality of something, which is — I don't know if that's just a function of the way they work or something like that. Anyway, I haven't been able to figure that out yet. But

That's a good question. You know, maybe I'm asking the wrong question. Maybe the question I should ask is: I want you to simulate an audience of a thousand people and estimate how many people would laugh at this joke.

I have, but I actually use real-life humans and human audiences to kind of test that prior to actually delivering it on stage. But that's — it's some interesting ideas like maybe just run a mock simulation of an audience with a certain makeup and say, 'Okay, well, I'm going to test this joke or this humor and see if it gets a reaction or not.'

I'll be honest, the answer is not yet. I think we're very early in this notion of agents being more than just tools. So they're sophisticated tools now, as AI is, right? Because they can actually carry on conversations. We have a customer support agent, we have a prospecting agent. The thing we haven't quite figured out yet, and I think we're right on the brink, is I think the really interesting things start to happen not when you have an agent that's purpose-built for a specific thing, which is what our agents currently do, is when agents have knowledge of each other and can actually bring in — it's like, okay, I'm trying to resolve this customer support issue, I'm going to bring in the prospecting agent to try and resolve this issue for whatever reason, like bring in some other agent that's disconnected from the core mission of the agent that's currently running. And I think part of this is going to get resolved. We have things like the agent-to-agent protocol from Google and MCP where the large language models can be exposed to kind of tools and other agents and we have some sort of discoverability. But that kind of dynamic of unexpected things — now it will do unexpected things and be able to answer questions we didn't think it would be able to answer, but it doesn't really draw outside the lines, right? It's like, okay, well, sort of within the system instructions and the RAG-based knowledge that we kind of provided it will do.

Mostly stay within the lines. And we see this with humans as well. It's like very, very interesting things happen when the number of humans exceeds one, right? It's like, okay, there's a certain amount of creativity I can draw in a brainstorming session with myself, but then when I kind of put other people in there is when the truly interesting things start to happen. And the other thing, I think this is more on the kind of technology side, and AI is getting better and better and that's great and protocols will get better, but I think we need a better way to kind of weave agents into all the same tools that we use, be it email or Slack or whatever it is, right? Like they need to be sort of part of the same systems that we use for doing the regular work that we do. I think that'll probably happen probably next year sometime.

Yeah.

Yeah, I think so. And this is not just the optimist in me. It's, you know, from a just a raw IQ perspective, you know, models already have the intelligence to kind of do a lot of the things that we do that we would ascribe and say those are human-oriented tasks. There's a couple of things where we can continue to do work. One is around EQ, around understanding — despite its lack of human experiences, can it actually develop true, at least a proxy for empathy? And it's starting to get better at that because it sort of fakes it pretty well. And there have been some studies that are done. But the thing I think will sort of really unlock it is not the technology's problem. It's our problem as organizations. The analogy I use for agents is like — so let's say you were hiring what I think of as a very adept intern. That's the way to think about an AI agent. And let's say this adept intern was really adept, had a PhD in everything, was very, very knowledgeable, super smart. But when they show up for their first day of work, they don't know about your business. They don't know about your policy. They don't know about their teammates. They don't know anything other than all the external training that they've had. Or we're like, 'Ah, we hired this really smart agent to do X and we sort of unleash it on the world, unleash it within our organization and expect it to really perform.' And I think that's an unreasonable expectation. So once we start truly treating them as potential teammates, well, they're going to need onboarding. They're going to need some training. They're going to need some feedback. They're going to need some objective, measurable way to know whether they're making progress. They're going to get some things right. They're going to get some things wrong. And then over time, I think we'll have a much more feels-like-a-teammate kind of person that just so happens that they're digital, but they're sitting in my Slack. They're responding to emails. They're kind of watching what's going on and they don't have to always be prompted to do something. And the distance between how much you have to sort of look over their shoulder, you know, it sort of grows over time just like we have with a regular person. The trust builds, they get more awareness and more context about the organization. I'm hopeful we'll get there.

I still think of it as a tool, right? So, I don't like personifying tools. It's like — I love AI as much as anyone. I have a huge amount of respect for the technology, what it's going to be capable of. I don't think it actually replaces a human in the literal sense. I think it replaces individual tasks that humans may or may not should have been kind of investing their calories in because they're capable of so much more. And the reason we did those things, and we have the same thing with the industrial age, is like we sort of as we invented machines and automated things, like, okay, now we can sort of work at a higher level as humans, work at a higher level of abstraction. I think these will still continue to be tools for a while, but that doesn't mean that they can't be accountable and responsible and overtake things with a high degree of trust. But I think ultimately like the tastemaker, the thing that's going to hold it all together, the glue is going to still be carbon-based life forms. I think there's still something particularly special about humans. I may be biased being a human, but

I understand the arguments that argue that AI kind of squashes or suppresses creativity because it allows you — it sort of replaces the need for that kind of creativity, and I think the exact opposite. Creativity is like, okay, you have some concept, some idea in your brain. And the challenge that many of us have, and I'll talk — you know, like my son has, so he's, you know, this aspiring fantasy writer someday, but you know, when he was nine and 10 and even like now, his skill set in terms of being able to put words together into sentences and actually write and do it was lacking, right? That it's going to take him years to kind of develop that to even a rudimentary level to where he can kind of craft a story for his target audience. But now with AI, it's like, okay, well, he still has the ideas. He's always had the ideas in his head, right? It's like — and I, you know, from a design perspective, since we're here on the Canva podcast, it's like, okay, well, Canva did this for me, you know, back when you folks first came out and now AI does this for me. It's like I've never thought of myself as — because I'm not like a designer, but it's not because I don't have visual ideas, right? It's like I have things that are in my head that I want to accompany a blog post or something like that. And now with AI, it sort of reduces — kind of lowers the bar in terms of the kind of entry level for people that do a particular thing. And so I think it amplifies creativity because now it takes the ideas that you have in your head, the creative ideas, and helps you manifest them, express them. Be it music, be it visual design, be it writing, whatever it happens to be as your kind of skills catch up. On the whole, I think people are going to be more creative. We're going to see more output, more creative things from people because the bar is just lower. More people can do it.

Yeah, a couple of things. So, I just launched a very simple kind of agent. You know, so we've heard of like prompt engineering, which is a craft of writing a prompt in order for you to get good results from AI back. But there's this kind of step up from that and it's called meta-prompting. And the idea behind meta-prompting is using AI to actually improve your prompt. And the agent that I have is metaprompt.com. It's completely free. And so what you do is you type in the prompt that you have been using on a recurring basis. And what it will do is say, okay, here's what you're trying to do. It'll come back and ask you questions. It's like, oh, you're doing this, but you haven't told me this. You haven't told me this. You haven't told me this. Answer these questions. And when you do, it comes back with a well-structured prompt that's actually written to — and it's like a one-time investment to take your prompt that you've been using all the time into something that will demonstrably prove to give you better results, right? And I use that all the time. And this is one of the nice things about AI now in agentic or vibe coding is that these are the kinds of things I would build for myself back in the day. And it wasn't the calories necessary to be able to kind of launch something and put it out there for the world to use. I could just never make that kind of trade-off, but now it's like, okay, not that hard to build, not that hard to share, and I can just put it out there.

Totally.

I tend not to use it for writing, per se. What I use it for is editing. I use it for visuals. I use it for what I think of as less about writing and more about copywriting. So, I'll use it for headline ideas or subject line ideas. And I'll use it for coming up with like punchy ways to kind of capture something. It's like, okay, I've got this thing. I'm like right around here on this particular concept. I need a pithy way that maybe uses alliteration or a figure of speech that makes it more memorable, more remarkable, and sort of improve and take it up a notch. And I've made multiple attempts at this. And I think it's pretty good, but it still can't capture my voice, and it's not good at that yet. And I say yet, maybe it will be someday, but

It can replicate my voice. What it is likely unable to do — and maybe this will change — the thing that it doesn't have though is it doesn't actually have human experience, right? So, it's not had the day that I just had. Kind of going back to comedy, which I'm not a comedian either, but you know, I've heard comedians say it's like their best material just comes from observation of like, I had this day and it was nothing particularly remarkable, but that was kind of quirky. I had this or whatever. And then they kind of pull on that thread and that turns into, you know, a skit or a joke or something like that. And I think it's similar for writing. It's like, okay, I was solving this particular problem or I had this thing where I had this debate on a call and it's like, oh, then I thought about it. It's like, oh, there's something broken or something kind of noteworthy about that. I just put it out there. So, yeah.

Yeah. For the kinds of things that I'm doing, right? It's like if I'm doing a LinkedIn post, the kind of purpose of it is to kind of put my brain on speakerphone in a way, right? That says, 'Okay, well, that's I think what people are there following me for.' It's like, 'Ah, what's going on in Dharmesh's head?' And if you're not curious about what's going on in my head, then you're likely not following me. That's because I have a grab bag of things that I post out there. And it's like, okay, well, that's sort of why the audience is here. So like, and my kind of purpose

is to kind of serve, you know, serve that audience. And so yeah.

I'll summarize my dozen plus years in kind of working in culture involuntarily. I'll distill it down like the big lesson, big analog I've come... I've only said like maybe three smart things in my life, I think this is one of them, which is culture is a product. And what I mean by that is as companies we build a product for our customers, everybody knows that, whatever your offering happens to be. Culture is the product you build for your team. I'm just talking about human teams right now. Well, you would never build a product without asking your customers how the product was doing. So we get, you know, product feedback, so we do an NPS for our employees in the same way. So we sort of carry as much as we can from our kind of product sensibilities over into building this kind of second product, which is what culture is.

And the other kind of big lesson on culture was in the same way that a product is never really done, culture is never really done. I think the mistake a lot of founders and entrepreneurs make is like, oh, my job is to preserve our startup culture, is like what we had when we were like five people, 10 people, 50 people. As long as I can preserve and maintain that, then like we're good. It's like, no, that's not the answer because the product that the company needs and the people need is going to be different at 50 people, 500 people, 5,000 people.

Yeah.