The Future Of AI Agents With Dharmesh Shah | INBOUND 2024

HubSpot turned 18 years old this past June. And since then, I've had the honor of sharing a number of memorable moments on the Inbound stage with you over the years. Thirteen years ago, I shared the launch of my son, Sohan. Back then, he was a tiny language model. Now he's a middle schooler, so he has a teenage language model, which, I will confess, I have not completely decoded yet. Parents in the room know what I'm talking about. Now I can't show a current image of Sohan because he's revoked my licensing rights to use his image. No cap. No cap. Ten years ago, we launched the HubSpot CRM. Many of you in the audience were here. Applaud if you were here for the launch of Inbound CRM. Yep. My lovely wife Kirsten, she was here. Yep. By the way, she's the one that got me the rocking shoes just for you. I know you can't see them on stage. You'll see them later. Yeah. Thank you, honey. So one year ago, we launched our first products with generative AI, including Chat Spot. Chat Spot started as a labs project and now has graduated to become Breeze Copilot. They grow up so quickly. So this is the year of AI Agents. When we look back at this event years from now, you're going to remember it as the Year of AI Agents. Now, agents are all the rage this year, but you know what was the rage in the 1980s? Corduroy pillows. They were making headlines everywhere. Let me give that one a couple of seconds. So here's what I'd like to do. First, I'll give you a debrief of what has happened in the world of AI since we last spoke. Then I'll try and demystify AI Agents because they can be unnecessarily mysterious. And finally, I'll show you some software that's in development. And as is permitted in my speaker's contract, I'll sneak in some dad jokes. I apologize for nothing. I bring my full dad self to the stage. Speaking of dad jokes, so I'm building a dad Generator agent. This is true story. And I collect all of them. I've got 2,000 in what I call a database. And so this is basically feedback for the training agents. So thank you for laughing for those that laugh as we go along. But anyway, so here we go. So a lot has happened in the past year. I'll focus on just the highlights, things that matter to you and your company. First, AI models are getting bigger. Bigger in terms of the number of Parameters, which sort of defines how big their neural network is, and also their context windows, which defines how much information we can pass to the model at the time we give it a prompt. Bigger models are generally more capable. They can reason better, they can write better, they can follow instructions better. But models aren't just getting better, they're getting better faster. So years ago, Gordon Moore, the CEO of intel, coined what's called Moore's Law. It states that computers double in power every 18 months. That's an exponential curve. That's why a computer that used to fill a room now fits in your pocket. So Moore said computers would double in capacity every 18 months. Then AI came along and said, hold my beer. AI models are doubling every six months. And this growth is partly what's causing all this excitement around AI. And it's not just that the models are getting bigger and better, but we're seeing innovation along multiple dimensions. We now have multiple highly functional frontier models that are good at different tasks. OpenAI's GPT is really good at reasoning. Claude 3.5 is really great at writing. Google. Gemini is really great at supporting huge context windows with millions of tokens. And LLAMA is really great at being open source and having a cute name. And we have some breaking news on the AI model front. For almost a year now, the industry has been abuzz with the anticipation of a new model from OpenAI. Those are the folks that make ChatGPT. The model is codenamed Strawberry. Word on the street was that it was going to be very impressive. What's not very impressive is that joke. I did not write that joke. I will not take credit. OpenAI launched that new model this past Thursday and I've been up late nights, even later than usual every night since. The model is called GPT01. Now, the reason to be excited about GPT01 is because of its ability to reason. It's shockingly good, like graduate student, PhD level reasoning. And it even shares the kind of chain of thought it had in terms of how it came up with the response that it gave you. So it sort of lets you inside its brain. Super cool. Now, this level of reasoning means we can have agents. Now they take on much more sophisticated goals that when we could even do a week ago. That's how fast this stuff is moving. So the lesson here is that when things are moving this fast, we shouldn't go to where the AI is, we should go to where AI is going. Hat tip to Wayne Gretzky, and we're going to be in a very different place three months, six months, 12 months from now. Okay, so we have a variety of great models. They're also becoming increasingly multimodal. So we've moved beyond text to the model supporting images, audio, and video too. This means the models can take image as input, not just the prompts that we give it. This is like we're giving these AI models the sense of sight. That's like when you're a teenager can first start seeing the clothes on their bedroom floor. It's amazing when it happens. It's amazing. So this is what AI models used to see when you sent them to a webpage. This is the webpage for School of rock, long standing HubSpot customer, and I'm the parent of a recent customer of School of Rock. So this is what the models used to see, and this is what the models can see today. They see the same webpage as we do. So, quick recap of the debriefing. Models are getting bigger and better. We have different models for a diversity of tasks, and AI models now have multimodal support. This makes it possible for the first time to create AI agents that can take on a wide variety of assignments. But agents can seem a bit mysterious and complicated. So let's try to demystify them. Let's raise your AIQ now. Agent IQ Speaking iq. I read somewhere that talking to yourself is a sign of intelligence. I've said that to myself hundreds of times now. First of all, you might be wondering, like, what is an AI agent anyway? And there are a lot of definitions out there. The one I prefer is the one Yamini shared. An AI agent is software that uses AI and tools to accomplish a goal that requires multiple steps. That's it. Yes, some agents can have be able to run autonomously. Some have executive planning capabilities. But those are niceties, not necessities, to be an AI agent. And there are a range of agents in terms of the complexity of the goals they can take on. Everything from complex goals like the four agents that Andy just demonstrated, all the way down to very simple goals. The thing I want you to remember is that an agent's an agent, no matter how small. Now, whether big or small, all agents share some things in common. An agent uses one or more AI models. If it doesn't use AI, you can't officially call it an agent. Then it's just sparkling automation software. Thank you. Next, an AI agent has the ability to use tools so it can access things like the HubSpot Smart CRM and Take Action on your behalf. And finally, agents have a memory. They can remember things across tasks. Otherwise, it's like me walking up the refrigerator, opening it, and then forgetting why I opened the refrigerator. All right, so let's dig in and talk about what's in development. I'm going to share with you what I've been cooking on late at night, 2 o' clock in the morning, every night. My wife can attest to this. And by cooking, I mean coding. I can't actually cook, not in a literal sense. And I also don't stand at my standing desk, which doesn't sit well with my personal trainer. I actually don't have a personal trainer. I used GPT, by the way. It's suggested that I maybe do, like, lunges. And I think that would be a big step forward. And so having the honor of being up on stage at Inbound and sharing the stuff I've been working on is the favorite time of my whole year. And I've probably spent 1500 hours working on the next 15 minutes that I'm about to show you. And before I do that, before we dig in, a quick disclosure and warning. The upcoming segment is rated X. X for experimental. Everything Andy showed you are part of real products that we sell at HubSpot. Products you should buy, by the way, if sales team is hiring. That's my sales pitch right there. What I'm about to show you is an experiment. Viewer discretion is advised. And the product I've been staying up late every night working on for many moons this year is Agent A.I. thank you. Agent A.I. is the number one professional network for A.I. agents. It's also the only professional network for A.I. agents. Now, at this point, some of you might be thinking, wait, what? Why do agents need their own professional network? Reasonable question. Because I predict that in the future, our teams are going to be hybrid. They will consist of humans like you and AI agents like me. Just kidding. I'm not an AI agent. Yet here on stage. That likely won't happen until far into the future, like Inbound 25, by the way, inspired by Andy's references to Back to the Future. I was going to tell you guys a joke about time travel, but you didn't like it. So here's the homepage for Agent AI. We see a listing for some of the agents on the network. And every agent has a profile. So we can click into the profile, we can follow that agent, we can hire that agent and start using it. It's free. And we have and we can add the agent to our team, and we have a bunch of agents around marketing, sales, customer service, operations. And a new agent's being added every week, so I invite you to check it out. You can sign up for free, get 100 credits. It's awesome. So it's time to meet some of these agents. Now we're gonna look at agents that help us with marketing, selling, and how to delight customers. Now, the number one challenge for marketers today is generating traffic and qualified leads. I'm not making that up. One of the most clueful people I know in our industry said this circa 43 minutes ago. So it's even more important that we take the website traffic that we have coming right now, those website visitors, and convert them to leads at the highest rate possible. And the way to do that is with conversion rate optimization. Now, here's the website for School of Rock that we saw earlier. It's a great website. Is there anything that can be tweaked or improved? Yes, always. But is it a pain in the butt to figure out what to change? Yes, always, because you have to review the copy, you've got to check the cta, you've got to tweak the headlines, review the testimonials. There's just so much to do. And conversion rate optimization is a very specialized skill. Thankfully, we've reduced this down to a few easy clicks. Here's the conversion rate optimizer agent. It's super easy to get started. All you do is you type in the web link for a webpage, any public web page. It can be your homepage, it can be a landing page, it can be your pricing page, and the agent goes through, looks at that entire webpage and it has decades of domain expertise and conversion rate optimization. And will it come up with a list of suggestions? It's like you're saying something this way on the webpage. I'll bet you you'll get higher conversions if you change it to this or you're missing a call to action, or maybe you should add imagery. It comes up with really, really good suggestions. I encourage you to try it. So it's, it's a simple agent, but it's a profitable one. I encourage you to check it out. Now, Yamini also mentioned how the average sales rep spends only two hours a day with customers, and a large part of their time is spent in prep, and a large part of their prep time is spent researching companies. Now, the reason this takes so much time is you have to go to Google. You go to the company's website, you go to Crunchbase and look up their funding, you go to YouTube, go to the social media channels, you go to Glassdoor. You're jumping all over the Internet on this adventure hunt with all this unstructured data. We built a company research agent that picks up all of those pieces all over the Internet and puts it together for you with just a few easy clicks. Let's look at how easy this is. We start by entering the website's domain. I'm going to enter OpenAI.com, my favorite AI company, and we get a beautiful multi page report broken out into sections. So we can click on any of these sections. So let's say we click on founders to see who the founders of the company are. We can click on the new section and see if there's anything new that's been published about the company. We can click on their web traffic information or their organic keywords. A bunch of different sections brings it all together in one place. But the coolest part is this research question section. You can type in any question you want. It's like let's say you're selling payment software and you want to know whether this company sells a subscription product. The agent will crawl the company's entire website, crawl the Internet and come up with an answer for you. And it says, yes, OpenAI sells a subscription product, but what's really cool is you can add that question to the agent. So every future report that you run will automatically answer all your custom questions. So let's say yes, OpenAI is a good fit. With a single click, we can add this company to the HubSpot Smart CRM and we can see that company record and can do all the things you can do inside HubSpot Smart CRM. And finally, let's say we're interested in this company. We can click the track button and type in English what we want to track. It's like, okay, if they launch a new product, if they raise a new round of funding, if there's an exec change, describe it in English, whatever is you're interested in tracking. And now this agent works on your behalf and will track all of those things automatically for you and notify you if something interesting to you happens. So that's the company research agent, by far the most popular agent on the network right now. Thank you. Now what if I told you agents can use other agents? I'm telling you agents can use other agents. We can use existing agents as building blocks for building new, more powerful agents in. It's like agent composition. I haven't had this much fun since I played with Lego. So this weekend, basically. Okay, so let's say I'm a customer success manager, and Canva is one of my accounts. Love Canva, by the way. Shout out to Canva, folks. Yeah, I know. Wonderful company. So now, as a customer success manager, before hopping on a call, I'd like to kind of update myself what's been going on with this company. Both within the HubSpot Smart CRM, but. But then all over the web news, funding. Now, this sounds familiar. A lot of that work the company research agent already does. So we built a simple agent that takes all of this information and reduces all that work in terms of creating an executive briefing into a few easy clicks. Let's see how easy this is. So here's the executive briefing agent. It knows which company I'm going to, which one's assigned to me, and it will come and it will say, okay, I'm gonna look at the CRM. I'm gonna look at the last calls, any issues. I'm gonna go look at the a customer research agent, and I'm gonna pull it all together into a very brief executive briefing, and I'm gonna convert it to an audio file so you can just kind of listen to it on the commute to work if you want to. Right. Just pulling all that information together and giving it in the form that you want. All right, so given the agents that Andy has demonstrated, given the agents I just showed you and the two dozen other agents on the agent AI network already, and more coming every week, I could make the prediction that most of you will be using agents this year. And that's an easy prediction to make. So I'm going to make that prediction. Most of you will be using agents this year, but I'm not going to stop there. I predict that many of you will be building agents this year, and not just the introverted developers like me. Anyone can build an AI agent. You can build an agent. You can build an agent. You can build an agent. Yes. By the way, if, you know, I could have my, like, Steve Ballmer chant right now, it's like, agent Builder, Agent Builder, Agent Builder. But I do not have the dancing skills of my friend Brian, so we'll just enjoy him for a few seconds. This is him on the inbound stage years ago. Hi, Brian, by the way, thankfully, Brian does not drive a hard bargain in terms of using his image in my presentations, which I'm grateful for. So announcing Agent Builder, which lets you build agents. Why should I have all the fun? So let me show you how easy it is. So here's the agent builder for the agent. I just showed you the customer briefing agent. So it goes through five steps. It says, okay, go talk to the HubSpot CRM and figure out which account has been assigned to me. Now go do the research agent, collect all of that information. Now go pass all of that context to a large language model, generate an executive briefing, and then generate the audio file so I can listen to it on my ipod or my music player. And that's it. It's drag and drop. I'll give you a sense. If we click on the plus button, there are other actions we could add. Let's say we wanted to add another step to this. And we have access to all sorts of data inside the HubSpot CRM. We can build UI, we can do all sorts of things. It's awesome. So the coolest part about Agent Builder has access to all the frontier models. So it has access to GPT to Claude. It has access to all the image generation models, Dall E3, Flux, Dot1, Ideogram, Playground, name it. It has access to all the HubSpot CRM, has access to people who create blog posts, create web pages, has Google, has YouTube, Twitter, everything. You can mix it all up and build an agent with drag and drop in the same way you build workflows in a marketing automation product like HubSpot. All right, I'm gonna take a breath. This is a cue for me to breathe so I don't pass out on stage. Thank you. You're welcome to join in and take a breath as well. This is not a moment of Zen. We have work to do, so we're all right. So that's Agent AI, the number one professional network for AI agents. Also the only professional network for AI agents. Now, normally for labs projects like this, if I get a few hundred users before the launch date, I'm happy. That gives me feedback, gives me signal. If I get a few thousand users before launch date, I'm like, this kid's got potential. This one's gonna grow up and be something special. And I am thrilled to announce agent AI already has 47,000 users. They grow up so quickly. This makes it the second most successful innovation project we've done at HubSpot, the first one being HubSpot CRM. All right. And what's equally exciting is we have 1700 people signed up for the wait list for the agent builder. And I'm approving people literally every night. So I'll be doing that in the hotel room later tonight. So sign up. It's Fun. And remember, you don't have to be a coder to be an agent Builder. You just have to be curious. All it takes is a few easy clicks. So we started by building a ui, a user interface. This part of it is probably a little geekier than it needs to be for a mainstream audience. I'm a little geekier than I should be for a mainstream audience. It'll be okay. This will be quick. Then we layer it on with an application programming interface. This allows applications to be able to talk to HubSpot. The next evolution of this is what I call the aui. That's the agent user interface. And the reason this is important is this is what you need for agents to be able to talk to each other and be able to collaborate. And the cool thing about Agent Builder is every agent that's built on Agent builder automatically has an aui. Now, this may not make sense to some of you. That's okay. But your developers are gonna love it. I promise. A movie reference, that's a really subtle one, but we'll let it go. Okay, so now some of you are wondering, why is a customer platform company like HubSpot building an agent network? It's my belief that the next generation customer platform needs to have three things. It needs a smart CRM that has unified store of structured and unstructured data. Yamini talked about this. And that foundational AI has to be built into the core CRM. The batteries have to be included, right? Otherwise it's not a smart CRM. It needs smart engagement apps. You saw Andy talk about this, where we're infusing AI into all the engagement apps at HubSpot. And finally, you need an agent ecosystem, just like we have an app ecosystem now. Shout out to all the app partners that are out there. Thank you. So at HubSpot, we like to dream big and iterate small. And our dream has always been to help millions of organizations grow better. We do this by demystifying, democratizing, and delivering on the potential of new technology. If you're a Fortune 1000 company with a fortune to spend on AI, there are lots of good options out there for you. But for those of you that just want to get growing with AI, HubSpot's here for you. Our focus is not just to make it exceptionally easy to use agents, but exceptionally easy to build agents and share agents and distribute agents. That's the idea. Now, since our founding 18 years ago, we've been inspired a lot by Apple. Apple changed the world years ago with the introduction of the iPhone. But what was equally impactful was the launch of the App Store. Their vision was embodied in the phrase, there's an app for that. Our vision for the future is there's an agent for that. For every marketing, sales, customer service, use case imaginable, including use cases we can't imagine yet, there will be an agent for that. Agents are the new apps, and we expect someday for there to be thousands of agents available. We want a thousand agents to bloom. Only a small fraction of them will be written and built by HubSpot. Most of them will be built by our partners, by all of you, the community. So we see a world where agents and humans that hire them are connected on a network and can collaborate in this vibrant marketplace. Now, if we think of agents as tools, we will get wonderful tools, but we should skate to where the puck is going. If we think of agents as future digital teammates, we will get wondrous transformation for our businesses. Now, it's natural for this to cause some anxiety, because it's not natural to think of a world where agents are members of our team. Deep down inside, we may be worried that AI will replace us as humans. As a human myself, I will humbly submit that AI does not reduce your value, it raises it. Even though I use AI every day to help me code, help me develop presentations, help me write dad jokes, I think I'm no less valuable now. I'm more valuable with AI and the same is true for all of you. Because humans are more than the sum of the tasks that they do. There is an inexplicable magic in us made from the experiences we've had. AI may be faster than us, AI may be smarter than us, but it can't be more human than us. Can't be, because it has not had human experiences. Like the thrill of starting something new with a friend. Thank you. Or the profound joy of sparking someone new with the love of your life. Or the deep, deep gratitude one feels towards a new friend that is helping continue the dream and vision. AI will not have those feelings because AI cannot have those human experiences. It can read books. It can inhale the entirety of the Internet, but that's not the same as having those human experiences. Having said that, although AI has its limits, it's not going to replace us as humans. AI agents can be wonderfully helpful. So we should let agents automate the mundane and amplify the magic. The magic that makes you you. It is always an honor to be here. Thank you so much for coming out to Inbound. Thank you for all the support on behalf of Yamini Andy and the Global team of 8,000 plus HubSpotters, I appreciate you. Thank.