LinkedIn Is Perfecting AI Agents with Mohak Shroff, SVP of Engineering at LinkedIn

(Intro Narrator at 00:00:00) Today, we're talking to Mohak Shroff, Senior Vice President of Engineering at LinkedIn, about LinkedIn's new AI agents. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:17) Can you believe it's been like five years since we talked?

(Mohak Shroff at 00:00:20) Unbelievable, man. What are you at? 900 episodes almost?

(Joel Beasley at 00:00:24) Almost. Yeah. I started just saying a thousand because I feel like if I count me guessing on other shows too, I've done a thousand.

(Joel Beasley at 00:00:34) Awesome.

(Mohak Shroff at 00:00:36) No, this is great.

(Joel Beasley at 00:00:36) I've been looking forward to this. Yeah. So, you know, one of the things that caught my attention was Alistair saying you guys are building—LinkedIn's building AI agents. You gotta tell me about that. What's going on there?

(Mohak Shroff at 00:00:47) Yeah. Absolutely. So, you know, it's interesting. LinkedIn has—a bit of a history lesson. I feel like I've been here now, what is it? It's coming up on seventeen, sixteen and a half years. It's a long time. And like, LinkedIn, it's this thing which I didn't appreciate about LinkedIn from the outside. And honestly, sometimes even from the inside, you have to sort of step back and take a moment to really appreciate it. But LinkedIn at its core has always been essentially an AI application. Right? What do we do? We're connecting people to opportunity at massive scale. The moment you add connecting and scale, AI sort of peeks into the picture. It's part of this core part of the picture. The way you connect people to things at scale is through AI. And so what do we do? We connect people. When someone shows up on LinkedIn and they want to learn about what's going on in the world, we connect them to other people, to knowledge, to jobs, to learning, to skills, to companies, et cetera. All that forever has been essentially AI powering what LinkedIn does.

(Mohak Shroff at 00:01:49) And so over the last two years, as the world changed radically with this emergence of this new technology with Gen AI and incredibly powerful models, these large language models, it sort of led to this moment of us asking the question of what does this mean both for LinkedIn as well as really what we do on a daily basis in terms of how we manifest the value we deliver in the world of work. I think one thing became really clear. Yes, we can continue to deliver—connect people to opportunity at massive scale. We can use this technology to do so better every single day. We can make our UX more human, more easily accessible. We can make it more intelligent, more interactive. For sure, we can do all of that. But something else became really apparent, that there's a moment coming where these large language models are going to lead to the ability for a lot of what people do from a work perspective to start to actually be automated, to be done by agents. And it led to us asking the question of, as LinkedIn continues to want to be or continues to be the thing professionals look to to make them more productive and successful every single day, how do we do that in the world of agents?

(Mohak Shroff at 00:02:59) So all these companies, everybody out there—a lot of people are thinking about agents as essentially replacing people in the world of work. It's like, what things can these agents do that'll be cheaper or faster or better, whatever it is? And I'm sitting here at LinkedIn going, hold on. We have as a core value and have for the better part of a decade and a half had as a core value at this company the idea that we are members first. We put the needs of our members, the ability to make our members more productive and successful at the forefront of everything we do. And so it's like, how do we take this agentic technology, these large language models applied to autonomous agents, to things that have the ability to make decisions on my behalf, to do work on my behalf? And how do we make them more human-centric? How do we ensure that as that agentic shift is occurring, it's occurring in a way that puts professionals at the center of opportunity rather than displaces them?

(Mohak Shroff at 00:03:35) And so very specifically, focusing on this idea that, hey, we can build agents in a way that—we've sort of had this idea of, what if the agents weren't taking the work we want to do? And instead, there's so much in the world of work that we don't actually want to do. My job as a technology leader is—I love immersing myself into technology. I love spending time talking about cool stuff that's going on in tech. I love building cool tech. If I'm an engineer, if I go back to my days, these ancient days back when I used to actually code—if I think about myself then, what I loved about my job was building software, making some idea come to life in some remarkable way through code. But I spent a lot of time getting stuff to build. And I spent a lot of time testing stuff. And I spent a lot of time scaffolding stuff to work with the way technology wanted me to make it work. And it's like, what if we could build agents that actually take all the stuff I don't want to do and do that in the way that I want them to do it, but let me continue to do these things I really want to do? It's almost like an injection of increasing my agency, enabling me to do more of what I love while the agents go and do the things I don't necessarily want to be doing so I can do the things I got into the job for.

(Mohak Shroff at 00:05:28) So a really good example of this, we just recently launched a recruiting agent. And so when you look at our recruiting agent—we call it the LinkedIn Hiring Assistant. And so when you look at the assistant, the hiring assistant, it's actually really interesting. What we did was we went and spoke to recruiters. Obviously, we recruit a bunch ourselves and we reached out to a bunch of our customers, recruiters, and asked them, you know, when you go and talk to a recruiter about what they do, there's this large list of things. There's a lot of time that they spend searching for candidates on LinkedIn, a bunch of time they spend kind of filtering those candidates, working collaboratively with the engineering leaders or the leaders who they're hiring for to figure out whether or not these are the right candidates. Then they reach out to those candidates to get them interested. And then some number of those people will respond to them and say, I'm interested. And then they engage with those individuals now, where they're like, alright, ready. Now that you've responded, now I can engage and actually do the thing I became a recruiter to do, which is interact with people, talk to candidates, get you really excited about the job, help explain and shepherd you through the process, and ensure that ultimately I'm helping to build a remarkably high-quality team at the company. But there's all this really important work that happens initially or in parallel that I have to do as well. But that's not why I became a recruiter. Nobody took a job so they could run Google searches all day or LinkedIn searches all day. Right? You took the job so that you could actually do this very human thing, the thing you are best suited to do in the world.

(Mohak Shroff at 00:07:04) And so we built the LinkedIn Hiring Assistant with this very specific goal of how can we help recruiters take this stuff that is, frankly, highly algorithmic, but we need to do it the way they do it. You can't just say, hey, LinkedIn is going to tell you how to source candidates, how to search for candidates, because every recruiter has a unique approach. You have your search queries. You have kind of the unique salt that you put in, the unique spice you put into the way you find candidates, the way you reach out to them. And so we thought about building an agent or a set of agents starting with recruiting, where we'd really focus on human-centric approach to agency, to autonomous agency, where it's like, hey, here's what we're going to do. This agent is going to help you with the tasks you don't want to do. It's going to be—initially, it's not going to be super smart. We've built it to be really smart, but relative to you, it's not going to be super smart. It's going to learn from you. You're going to teach it. You're going to train it. You're going to interact with it. It's going to grow with you. And as it does, eventually, it's going to be able to hire the way you hire, to search the way you search, to reach out the way you reach out, and ultimately let you do what you really love and are best at in the world, which is interact with candidates, get them interested in the job, screen them, figure out what the next thing is, shepherd them through the process, and help hire extraordinary talent in the company.

(Mohak Shroff at 00:08:26) So we had this idea of continuing what we've done with AI. For years, we've been a place that takes AI and makes it hyper-personalized, makes it something that helps you be more productive and successful. And we thought, hey. Here's this cool new technology with agents. What can we do with that to do exactly the same thing? How can we help make you more productive and successful? You, the professional, more productive and successful with agents. How can we ensure that we put you at the very center of that? And so it leads to a bunch of really interesting design concepts. There's a bunch of building blocks that come with it and a bunch of really interesting design concepts. But, you know, we started off with recruiting. As you can imagine, there's a set of businesses that LinkedIn is in. Ultimately, we'll have a slew of agents across many of those, and then broader ambitions beyond that.

(Joel Beasley at 00:09:01) I've got a lot of questions. Let's dive in. When—let's go back to the beginning when you said that, you know, you saw the AI happening and then you guys got together and had this conversation. What did that actually look like? Was it an off-site? Was it just somebody knocking on the door of your office saying, hey, Mohak. I want to chat with you about this. Yeah. How did it become something you couldn't ignore to the point where you had discussion about it?

(Mohak Shroff at 00:09:28) Yeah. It's a good question. You know, so I'll start by saying obviously there's aspects of—yeah, it's hard to look at what's happening and not feel inspired by it, at least for me. I sort of look at, you know, you see some of this technology. You see, you know, with the launch of ChatGPT, and you see just how remarkable this tech is. The first time everyone interacted with it—I suspect we all still vividly remember that first interaction moment where we're like, woah. Everything is different. And it's not even that smart yet. We asked to do things, and it's like, it's not even close to there yet where it's like, oh, this is perfect. But it's so clear that the promise exists. And I think a lot of people, sort of simultaneously across the industry, a lot of people had the idea of, there's something there. There's something happening here that's interesting.

(Mohak Shroff at 00:10:09) Then I think comes actually our culture and values. So again, we have as a core value, core cultural tenet, this idea that we put members first. And it's easy in moments like these where there's some cool new technology coming out to say, hey, you know, let's just go and play with it, and let's go build something because it's cool or because it's possible. But to instead say for us, every time there's a technological insight, every time there's a new innovation, we ask the question, what could we do with it that puts our members first? How can we ensure that what we're trying to do is deliver value to our members and our customers? And so that's the second thing that comes in. I think those two things happen simultaneously for us. On one hand, very cool tech. Everyone's blown away. It's making people dream. On the other hand, the immediate question of how will we use this, leverage it to drive value for our members and our customers to continue to create economic opportunity. And it's at the intersection of those two. When vision intersects with what's happening in the world with innovation, I think real magic happens.

(Mohak Shroff at 00:11:21) And so that's kind of what happened for us. It's just sort of been our natural mode of operation for years. You know, mobile revolution is like, hey, we could just build an app. What if this app is something that lets you do X, Y, and Z? It's like, oh, but hold on. How do we ensure we connect you to opportunity? Even now when we—we've recently launched LinkedIn games. For years—

(Joel Beasley at 00:11:47) Game? You—there's games on LinkedIn? Wait.

(Mohak Shroff at 00:11:49) Is this news for you?

(Joel Beasley at 00:11:51) This is news for me.

(Mohak Shroff at 00:11:52) Alright. Alright. Every one of your viewers, every one of your listeners has to go and check out LinkedIn games. They're extraordinary. Honestly, I'm a huge gamer, and LinkedIn games are my new addiction.

(Joel Beasley at 00:12:02) Is it easy? Do you have to do something special, or is it just there?

(Mohak Shroff at 00:12:05) It's just in the app. Yeah. It's just in the app.

(Joel Beasley at 00:12:07) I see video, network, notifications, chat.

(Mohak Shroff at 00:12:09) Click on the network tab, and there should be a tile right down the middle that's LinkedIn games. But so let's take games as an example. So when we launch LinkedIn games—for years, we've talked about the idea of gaming as something that drives a lot of engagement. We're like, but hold on. What does that have to do with making people more productive and successful? And we really took the idea that people connect with each other as a core concept in the workplace. When you think about the professional world, we're—

(Joel Beasley at 00:12:34) So are we playing games with each other, or is it just between you?

(Mohak Shroff at 00:12:36) Well, actually, you're sharing with each other. It becomes a conversation piece. These are games that are all about exercising your brain. They help you sort of shake out the cobwebs. Or for me, every morning, they shake the cobwebs off and get me ready for the work day. But it's also about interacting with people I work with. I see, oh, X, Y, and Z person I work with is playing this game today. Oh, Ryan played it? Alright. Well, I've got to go and try and beat this score. And it becomes a really cool way to connect and interact.

(Joel Beasley at 00:13:05) I gotta try it.

(Mohak Shroff at 00:13:06) No. They're amazing. They're amazing. We recently announced a partnership with Fortune. We're actually—Fortune will be featuring some of our games on their platform as well. So the—but again, the point is all these things we build, every single one of these innovations comes from this very specific perspective. How does it drive value for members and customers? And so this is a great example of games. Everyone's known on the Internet, games are a big deal, and lots of people love playing them. Lots of people use them, and it drives usage. But how do you focus it on ensuring that it's about creating value in the world? Because this core aspect of the way we approach things at LinkedIn, which is we seek to do good and do well. And we sort of have, I think, somewhat uniquely in the world, a platform at extraordinary scale that does good and does well.

(Joel Beasley at 00:13:57) Well, you said everyone kind of remembers their first time using one of these tools, these AI, LLM type tools. What was your first experience with them?

(Mohak Shroff at 00:14:06) I mean, I think it's probably very similar to everybody else. About two years ago, just over two years ago, I had the opportunity to look at GPT and play with it. And I remember one of the first things I did was, you know, I've seen intelligent AI before, and I've seen it be able to answer questions. You ask it a question about what is X, Y, and Z, and it gives you an answer. I'm like, that's not that great. It's a slight improvement over search.

(Mohak Shroff at 00:14:29) I asked it to write me a poem. I asked it to write me a poem about LinkedIn. And it just said, like, hey, here's this thing. Here's some random prompt.

(Mohak Shroff at 00:14:40) I forget what the exact prompt was, but it was, here's this random prompt. Write me a poem about LinkedIn. I remember reading the thing going, this is pretty good. It's not amazing. It's not gonna be in some textbook years from now talking about great poetry, but it was pretty good.

(Mohak Shroff at 00:14:53) I remember thinking to myself, this is better than I could write, and I've got no creative talent. I can't write a poem to save my life. But I remember looking at it and thinking, this is extraordinary that very soon, every single person on Earth will have access to technology with this creative ability. And so what was really special for me about this technology wasn't just that it was able to answer questions. I think that tech has existed before.

(Mohak Shroff at 00:15:18) It was two things. One, it was the fact that it had this creative angle to it. It could seem to be doing things in a way that was more human-centric where it was more relatable to me. And the second was actually the interaction model. What had happened with ChatGPT was that it had made AI super accessible. It went from being technology that was built into very large applications and products to something I could just talk to in a very natural way.

(Joel Beasley at 00:15:45) Yeah. My first was in 2000, I think, '12. One of my buddies was going—he was out at Stanford, and he called me and he said, hey, you won't believe this. We're working on these—I don't know what he called them, but it wasn't large language models. But he said, I wanna show you this. And I said, okay. And he had some type of API endpoint where you could give it a Google search or you give it five web pages and it would, incredibly slow, but it would come back in a couple minutes and it would answer your questions about the dataset that you gave it.

(Joel Beasley at 00:16:25) So I could do a Google search and it's basically what you see today with Google's intelligent response, whatever they call it, the AI insight. And I just thought that was fascinating. I didn't chat with it in that sense or anything like that. That was my first exposure. And then, you know, we're technologists. We see this emerging stuff, these prototypes, and it's just like, okay, that's cute. That's cute. And then one day, it's just not cute, and it's actually usable. It's real.

(Joel Beasley at 00:16:50) It's real. And then there's—

(Mohak Shroff at 00:16:52) But so I think this is important. Right? What you said earlier, that 2012 realization, I think that's existed, to your point, for a very long time. But there's something novel about what we're seeing now where it's not just data extraction. I think we see that, we think of that as tedious. We think of that as a utility. This feels like it's so much more than a utility now. Right? It feels like it has responsiveness. It has reaction to the things we want.

(Mohak Shroff at 00:17:23) It feels far more personalized. And I think that's actually probably the thing that's most magical about it is that level of personalization. It feels like—and, again, at the end of the day, it's just a bunch of code. It's clearly not thinking, but it feels like it understands me.

(Joel Beasley at 00:17:41) Have you heard the GPT Voice thing? Yes. Absolutely. Have you talked to her? Yes.

(Joel Beasley at 00:17:46) It's a her for me. Yeah. Okay. So I used this morning. Maybe I'm gonna try to use it right now just to show you what I did real world.

(Joel Beasley at 00:17:54) So I woke up this morning and we're getting a fence installed. I'll actually tell her while we're doing it. Alright. So I woke up this morning and we're getting a fence installed, and the guys who are out installing the fence don't speak English. They speak Spanish. I do not speak Spanish. The gate is currently 13 feet wide. It's supposed to be 20 feet wide. Can you tell them in Spanish that I need the gate to be 20 feet wide, not 13 feet wide?

(Joel Beasley at 00:18:23) Sure. You can tell them. That's—there was no additional app to install. There was no training. There was no setup and registration and fee. It was just, can you do this?

(Mohak Shroff at 00:18:42) Yeah. So here's, I mean, no offense to you as a podcast host, but here's my drive every morning. Hey, I'd love to pretend that we're in a podcast together, and I want you to educate me on a topic that I don't know enough about. But let's do it in an interactive way. Let's talk about large language models. Can you host a podcast with me about large language models? Oops. Connection failed. Never mind.

(Mohak Shroff at 00:19:11) But we were gonna get there. But here's the crazy thing. Right? I do that every morning on my drive in, and I'm actually having a conversation with this thing. I'm talking to it, learning from it. It's remarkable. And so there's this incredible utility to me, but that same tech is going to get applied to skills, to things we do at work. And it's gonna start being applied to tasks, unique tasks at work. And so the thing is that ability starts to seem so clear.

(Mohak Shroff at 00:19:43) And I think in that context, you know, as a lot of people are out there—tons of venture dollars are being poured into this—the idea of finding ways to fully automate work, we're kinda sitting here every day... I come into work, and I'm like, how do we make sure that this technology gets used for good? Yeah. Yes.

(Mohak Shroff at 00:20:05) Automation is going to happen. Yes. This technology is real, and agents are going to be a thing. How do we make sure that those agents lead to the creation of economic opportunity? And that's what we do every single day. And so that's why we got into the agent game. That's why we're building these, and that's why we're building not just that, but also thinking about a framework. Right? Let's try to envision a future where, as you think about what we do every day in our jobs, there's a hundred things in our jobs. And right now, those hundred things all feel like a hundred different tools that we all have to use.

(Mohak Shroff at 00:20:42) I've got to log in to Word. I've gotta use search. I've gotta use a browser. I've gotta use this other enterprise tool. I've gotta use this thing. All these things sort of complicate our work lives. And I'm like, imagine a future where there are a hundred different agents. Now with tools, it's reasonably simple for me to basically say, I'm not gonna worry about this one right now. I'm just not gonna launch it. With an agent, the thing is gonna demand my attention on a regular basis because it needs attention.

(Mohak Shroff at 00:21:11) These things aren't at the level of precision yet where they can do everything I need them to do. And so across these many different aspects of my job, all these agents are gonna be out there raising their hand, pushing notifications saying, hey, look at me. I need some help. Look at me. I need some help. And so I envision this future where the way you put humans at the center of this is by having the equivalent of a chief of staff, but in a digital sense. You need this digital concept that manages, that organizes all of them. So, you know, I'm gonna hearken back to my early days as an engineer.

(Mohak Shroff at 00:21:48) I was a huge Emacs user. So my first internship in college, I ended up with a mentor who got me on the Emacs train. I've been all vi prior to that, and he got me on the Emacs train. And I created during that internship, I created a dot Emacs file. And I carried that dot Emacs file from task to task, job to job, thing to thing constantly. And what was great about it is that every new thing I picked up, it still felt familiar.

(Mohak Shroff at 00:22:14) My shortcuts were the same. My key bindings were the same. My autocorrect was the same. We knew not to stick a W in my name every time we tried to autocorrect Mohak to M-O-H-A-W-K. And so I'm like, stop doing that. And so it had all of that. And then one day, I stopped using Emacs because it was no longer cool, and I was no longer an engineer.

(Mohak Shroff at 00:22:48) But then I lost that. And since then, every tool I've used, every new tool I picked up wants—insists that it wants to relearn everything about me. And it starts off being really clunky and hard to use. And that sucks. Now just imagine that in the world of agents. This thing is gonna try to send out emails on my behalf. This thing is gonna try to search the way it thinks I want to search. This thing is gonna try and write text the way it thinks I write.

(Mohak Shroff at 00:23:02) And if every single one of these agents has to be retrained and retaught and remanaged and re-, re-, re-, it's just massive cognitive overload. And so we envision a world where you have the equivalent of a digital chief of staff that manages this team of agents for you. We envision this idea that it's personalizing them to you. It's helping them coordinate with each other. It's telling one agent, here's the other agents I work with, and here's what I need you to do. Here's the tools I work with and the ones I need you to work in. I use Office. I use Excel. I need you to use Excel. That's where I need you to create these sheets for me.

(Mohak Shroff at 00:23:39) Here's my credentials, or here's how I access them. Here's how I want you to collaborate. Here's how I work. Here are the hours when I want you to be able to bother me, and here are the hours when I need quiet time, where I'm actually focusing on my family, or I'm focusing on other work and heads down and things. That needs to exist.

(Mohak Shroff at 00:23:57) And so there's a whole bunch of tech that comes with that. So when we talk about the agents we're building at LinkedIn, there's these incredible building blocks that have to be put together. You have to build coordination and collaboration. You have to build shared memory. You have to build personalization, the ability for people to rearrange workflows the way that they work.

(Mohak Shroff at 00:24:15) And this massive agentic stack that gets built, all essentially getting to the point where you're able to use these agents in a way you want to rather than the way they want you to. And I think what this all leads to is, you know, going back to the point of personalization, my experience using this tech, this large language model tech was, wow, this is really personal. And what I think happens with agents is we're gonna return to a level of personalization in tech that we haven't seen in a while. I think work right now, given the technology around us, actually is far less personal than we potentially think it is. We think it's really personal because these are all the tools that I've been able to collect together and use, but they all sort of insist I use them the way they want to be used as opposed to just adapting to the way I work and letting me work the way I want to work.

(Mohak Shroff at 00:25:08) And so we envision kind of an agentic world and the agents we're building work the way you want to work, not the way they wanna work. We have an opinion, but our opinion is that it should be you-centric. And so this highly human-centric approach, this approach in which you are the supervisor of an agentic task force that is working to help you be more productive and successful rather than sort of off on its own.

(Joel Beasley at 00:25:34) You know, four or five years ago, I did an interview with Sir Tim Berners-Lee, and he described something like this before agents were mainstream. He didn't describe them as agents. But if you go back and listen to that episode, the way he describes how as humans we'll use the Internet is in line with how we're gonna use the Internet when the agents—now here's a question for you. Have you—you know David Singleton? He's, he was the CTO of Stripe. Have you seen what he's doing with /dev/agents?

(Mohak Shroff at 00:26:00) Yeah. Absolutely. It's so cool.

(Joel Beasley at 00:26:01) That's so cool, man.

(Mohak Shroff at 00:26:02) So cool. Very cool.

(Joel Beasley at 00:26:03) There's not much information. There's, like, a LinkedIn post, which is our videos.

(Mohak Shroff at 00:26:08) Yeah. A couple of videos. The screenshots. Yeah. But it's cool. Absolutely.

(Joel Beasley at 00:26:11) Are you guys collaborating with them at all on—

(Mohak Shroff at 00:26:13) I mean, I think, you know, it's early days. What I love about this period is—doesn't this just feel like a playground right now? We're all—

(Joel Beasley at 00:26:24) It does.

(Mohak Shroff at 00:26:25) We're all witnessing some remarkable technology being born and some remarkable use cases built on top of that technology. I feel like the moment we're in is kinda akin to the birth of the Internet in the sense of, like, TCP/IP exists. Right? And everyone thinks that's a big deal. But it kind of isn't. Right? The big deal was the browser. That's the thing that suddenly, you know, HTML and the browser and the massive level of accessibility that was created around, oh, people can build web pages and host them on servers. And all of a sudden, this huge explosion of value occurs. I think we're at that point. What we've seen right now is the baseline technology, the ground technology that will power a lot of the incredible value that will come. And so, you know, I've had this insight of, as engineering organizations, probably what's most important for us right now is finding ways to experiment and create flexibility.

(Mohak Shroff at 00:27:19) Right? How do we massively shorten the experimentation cycle? Because what we've all got to recognize is our conviction, while really positive—it's important to have conviction for what we build—is likely wrong. You know, there are very few people who could really predict the future. Right?

(Mohak Shroff at 00:27:37) And those who can, I'd love to meet them. But what we do know is as we experiment and as we build and as we build the right infrastructure that enables that experimentation, some of it is gonna become permanent. Some of it will stick. Some of it will move from, you know, toy app or toy thing to, oh, real thing that needs to scale. And then we'll scale that, and we'll build on it to the next thing and the next thing and the next thing.

(Mohak Shroff at 00:28:06) But you've got to be out there experimenting. So to your question about who are we collaborating—yeah. I'd love, you know, teams that are doing interesting things in the world of professional work to reach out. And let's talk about ways that we can collaborate in the agentic space.

(Joel Beasley at 00:28:20) Have you seen Clay? Clay.com?

(Mohak Shroff at 00:28:22) No. I don't think I have.

(Joel Beasley at 00:28:23) Oh, we love them. Yeah. They're not a sponsor or anything. They should be, but I use them. So you can set up these workflows and I can—essentially, you can do a lot of things with it.

(Joel Beasley at 00:28:35) But what I use it for is cold email at scale, highly personalized. And so I teach it how I want to talk, and then it can go research people and it's got all these fallbacks. So ideally, it's going to find their bio on the website, but if it can't, it'll do this. And so it's got these series of fallbacks, and then it's just a really good way to chain things together when dealing with working with different agents. Again, they have different agents inside of it too that can do different things.

(Mohak Shroff at 00:29:02) That's super cool. Yeah. I've been—impressed is the wrong word. I've been really sort of watching in awe at the way workflows are changing now. I had the opportunity to sit down with a very junior engineer recently and just watch as she worked, just because I wanted to get a sense for what it's like to be a new engineer, not just at LinkedIn, but in the world.

(Mohak Shroff at 00:29:27) And dude, it was such an interesting experience. All the things that I remember doing, she was just breezing through because of Copilot. She was in GitHub using GitHub Copilot, and it was like, "Hey, Copilot. What is the UNIX command for this? Hey, Copilot. How do I do this? Hey, Copilot." And I was just like, "Woah." That would have been a question to some senior engineer that took me six hours because they were too busy to respond, or it would have been me standing up and— but it was just nonstop, unimpeded productivity. And it's really magical.

(Mohak Shroff at 00:29:58) And so to your point, workflows are already changing. People are already starting to adapt, and productivity is just exploding. It's really kind of special.

(Joel Beasley at 00:30:07) And so we talked—the digital chief of staff. That's the future. There's not a digital chief of staff today in LinkedIn. Correct?

(Mohak Shroff at 00:30:14) No. There isn't. And I think the way we think about this is, look, it's early days. We're going to be building agents. We're building this hiring assistant. We're talking about doing the same for a number of our other businesses and areas. And as we do that, ultimately, again, in the point of learning, we will learn what does or doesn't work. What we're building for it, though, is things that feed into that vision of an ultimate digital assistant or chief of staff that helps manage that group of agents. So again, across the different agents we have, we will have shared memory. We want them to know when I work with one agent, and if in one agent I said, "Hey. That's not how I spell something," the other agent better know it because it should. It can. If over here I said, "Hey. Don't bother me at this time," the other agent should. If over here I said, "Hey. This is the way—this is the tone I like to use," this one should as well. And so we'll create that as kind of shared infrastructure across this group of agents.

(Mohak Shroff at 00:31:10) But then ultimately, to your point, going back to what LinkedIn represents— LinkedIn represents kind of your trusted friend, advisor, partner in your professional journey. Can we start to represent some of that in the agentic journey? Right? As you pick up agents, how do we take some of this incredible technology we're building and start to actually make that available in a way that powers all of the agents you work with in a professional setting?

(Joel Beasley at 00:31:38) Are you writing or blogging about this, posting on LinkedIn about it? How can people follow this venture with you?

(Mohak Shroff at 00:31:44) Post coming. Blog post coming shortly. Actually, possibly before this is aired.

(Joel Beasley at 00:31:48) So yeah.

(Mohak Shroff at 00:31:50) Yeah, a blog post talking about some of our agentic ambitions and work. Absolutely.

(Joel Beasley at 00:31:54) Is there anything that you can share that—or that you're uncertain if you can or can't share—that's coming?

(Mohak Shroff at 00:32:00) No. I mean, I think what's certain is lots of assistants. Right? So across our major business lines, I think hiring assistant—you can just imagine we have a sales solutions business line, helping salespeople be more productive. Absolutely something we should do. We've already been doing things with learning, helping learners be more effective. We help job seekers be more effective. We have job seeking assistant. The idea is that in every single professional flow, a place where professionals need help being more productive, being more successful, being more effective—there are things that they should do themselves. There are things they want to do themselves, and there are things that they have to do to get to do the thing they want to do. We want to build an assistant that does all that stuff, all the stuff that you don't want to do but have to do to get to the thing you really want to do. And we will continue, in terms of what our platform does, to build those kinds of assistants.

(Joel Beasley at 00:32:52) Has the AI agents—do you think that's going to change team structures at all? Or do you think we'll still see one-to-five type ratios with teams?

(Mohak Shroff at 00:33:00) It's a good question. You know, I hadn't thought about it. I think—I'll tell you what's on my mind with how the world of work changes, because it's less about team structures. I think it's actually—maybe this relates. I think that there's so much upside to helping individuals be more effective at making ideas real.

(Mohak Shroff at 00:33:26) Right now, the minimum surface area of building anything is, like, what, three engineers, four people, five people? Right? Engineers, designers, product managers, whatever it is. Whatever combination of things. You have an idea. Building that thing requires a combination of people. No one person at any scale or any interesting level of technology can just go build it themselves. But what if they could? What happens then? What cool things do we start to see in the world?

(Mohak Shroff at 00:33:54) Right? Out there somewhere is an entrepreneur who has just an incredible idea and just a ton of grit and determination, desire to go and make something happen. But they're impeded by the fact that the only way they could do it is by going and hiring a team of people who can do the things that they're not great at.

(Joel Beasley at 00:34:14) Mm-hmm.

(Mohak Shroff at 00:34:15) But what if we could actually augment that person with a team of agents that are great at all that other stuff?

(Joel Beasley at 00:34:21) That's what's happening. That is exactly what's happening in my business, Mohak. We've managed to—it's less about the decision of do we need to hire people, and it's more about the increase in quality that we can have because it's so accessible and cheap. It's things we wouldn't otherwise be able to do given our business model that we can now do that make the entire experience better.

(Mohak Shroff at 00:34:50) Yeah. And you know, I think this is really important. Every single time these technological shifts have happened in the past, they've started off feeling disruptive, and they are initially. But ultimately, they lead to massive increases in opportunity because when you make it easier for people to create, opportunity just multiplies. And that's what I think these agents are—that's what this technology is doing, is going to do. And that's what we seek to empower them to do—is how can we have you, if you're an engineer and you love coding, you just love coding, but you spend 30% of your day fixing build issues and writing tests or whatever, 70%, 60% of your day. Right? It's this huge amount of cognitive load. If we eliminate that, you will write more code. You will build more stuff. What cool stuff might you create when unimpeded by a lot of this other thing—all this other stuff that gets in your way?

(Mohak Shroff at 00:35:50) And so that's really the vision here. In the end, what this does is, to your point, it transforms businesses. It transforms individuals. It transforms careers. It transforms companies. As people start to be able to lean into the thing that they're really good at, really passionate about, and leave all the other stuff to agents that can support them.

(Mohak Shroff at 00:36:08) But the trick is it's such an easy thing to say. It's hard in practice because, you know, in a way, what we're doing is we're giving up agency. So let's talk about your business. Right? As you start to give a lot of those tasks to agents and to automation, it comes with concern because you're like, "Will it do it the way I want it to? Will the quality bar be where I need it to be? Will it change my voice? Will it—" You know, as you're going out in tools that will send emails on your behalf, like mass notification or mass outreach on your behalf, and it's like, "Hey. Is that mass outreach my voice? Does it sound like me? What if it says something I don't necessarily agree with?"

(Joel Beasley at 00:36:45) How does that feel? Yeah.

(Mohak Shroff at 00:36:47) How does that feel? It really erodes trust, and trust can be destroyed in just a moment. And so taking a trust-first approach to this is really, really important. And so for us, that trust-first approach is, look, you're in charge. Agency needs to start with you because you grant that agency to the tool, to the automaton, to the bot. And so granting that agency requires you to feel comfortable with the level of support, the level of supervision you can provide. Right? It's actually really no different than—

(Joel Beasley at 00:37:19) Hiring a team. Yeah. It's like— He's going to be— I'm delegating the edits to Josh, hoping that we have an aligned vision and looking at his track record and having enough trust to say, "Okay." And nothing's perfect. You have to balance everything with that.

(Joel Beasley at 00:37:32) When I was making the decision about the LLMs for the outreach, and I was doing batches of hundreds and then reviewing them and then tweaking the prompt and all of that. And I got it dialed in to where I know 175 is going to be not the best way that I could have said it if I were doing it, but the time-benefit cost analysis says run. Run with it. Because now I can do 10,000 a day, and it's just happening. You know?

(Mohak Shroff at 00:38:00) And so I think this is important. The thing I would say to everybody who's playing with agents is there's— Nature has real, uh, what's the word? Real lessons for us. Right? The way we work with each other should be highly informative to how we think about how we want agents to work with us.

(Mohak Shroff at 00:38:26) So I want an agent that works with me the way Josh likely worked with you as you scaled your relationship, as you sort of ramped up on the relationship. Initially, it's highly collaborative. He's like, "Hey. I've got this question. I want to make sure we're aligned. You're okay. This looks good." Lots more communication. As the trust builds, you delegate more. He feels more independent. He's escalating less. As the trust builds, as the trust builds. And over time, the level of actual escalation and the level of collaboration reduces because now he's able to drive a lot of this stuff solo. And I think that actually that design model is really powerful because it builds trust, and it leads to personalization. Josh works the way you need him to work. Now he brings unique flair to it. He brings a unique approach to it, but it's still fundamentally something that you are aligned with. So that idea of alignment, I think, is foundational. And so, again, as people are thinking about building agents, as LinkedIn builds agents, that's been really core to our approach, which is how do we— Talking about the LinkedIn hiring assistant, how do we build an assistant that— Recruiting is a surprisingly— it's an intricate field. There's so much art to how recruiting works. You balance many different constraints. You know, I was working with a recruiter recently on a search, and they were educating me about, "Look. Look. You need to— Your expectations are interesting."

(Joel Beasley at 00:39:50) What a kind way to say it.

(Mohak Shroff at 00:39:51) "Expectations are interesting. But let me educate you about the market, and here's what things are going on, and here's how you need to have these conversations. And you think you could say this, but you've got to say this." And it's such an interesting thing. It was a growth journey for me. That's how recruiting has to work. It always works that way in both directions with candidates as well as with the companies themselves and the leaders who are doing the hiring. That same process needs to make its way into how agents work for recruiters, helping them iterate and build trust over some arc of time where eventually you feel like, "I can trust you to do this work for me in a way that I believe is representative of me." Because in the end, the agent is still representative of the person. And so I think that key human-first design principle is really, really important.

(Mohak Shroff at 00:40:38) So, you know, for us, it's sort of encoded into our principles. We're members first. We start with that. I would love, as everybody out there is working on agents, working on this technology, that they start with that principle of just human first, person first.

(Joel Beasley at 00:40:51) Have you—are you on X at all?

(Mohak Shroff at 00:40:54) Do you spend—

(Joel Beasley at 00:40:55) Any time there? Okay. Yeah. So there's two things. Are we still allowed to say we're on X? Okay. Yeah. Yeah. Yes. We are. It's okay now. Once Elon caught the rocket, everyone was happy again. They're like, "Alright. Alright." Have you asked the GPT the question, say, "Tell me one thing about me that I might not know based off of our past conversations"?

(Mohak Shroff at 00:41:14) I saw that.

(Joel Beasley at 00:41:14) Yeah. Have you done that, though?

(Mohak Shroff at 00:41:16) I haven't. I actually— I think I saw that a day or two ago, and then I was like, "I should do it." I pinned it, and I haven't come back to it. I will.

(Joel Beasley at 00:41:22) So that one was good for me. And the second one that was good for me was "Draw me a picture of what you think my current life looks like." It got my kids. I was podcasting. It had everything.

(Mohak Shroff at 00:41:33) That's cool.

(Joel Beasley at 00:41:33) Yeah. It was really clever. Yeah. So—

(Mohak Shroff at 00:41:36) That's amazing. Yeah. And so, I mean, you imagine a future where these things don't just have the conversations you've been having with them, but they actually have the context of your work. And then you can start having discussions about, like, "Hey. How can I be more effective? How can we be more effective?" I just— I love that idea of collaboration. Right? You know, these tools working where we work, the way we work rather than us bending to them. They're bending to us. And I really look forward to a future that feels less cluttered. You know? In the world of work, I'm able to do the things I love because I love to do them, because that's where I can create most value, and all the clutter is off to this agent.

(Joel Beasley at 00:42:14) I feel like that in the future, I could wake up and I could have a team of five agents that I just talk to all day, as if I were working with them. And they have specialties and we interact because, you know, right now, my company is fully remote. And what is it? What is my interaction? Forget the work being done. What's my interaction with the work being done? It's phone calls. It's conversations about what's happening. It's how to handle a client doing this or that. It's— It essentially boils down to conversations.

(Joel Beasley at 00:42:47) And we can already converse with the AI, and now it's gonna be upscaling over the next five, ten years. I think in the future when that happens, and I want you to correct me or add to what I'm gonna say. But I see us driving a premium on in-person relationships and that there will be high trust between them. So it's like I've got my hoard of agents and I got my skill set that I can accomplish. And then I know you in person.

(Joel Beasley at 00:43:16) Right? Like, we have a relationship. We've shaken hands, had a meal, shared a drink. And then I think that it will put a larger emphasis on in-person human relationships going forward. What do you think?

(Joel Beasley at 00:43:28) Yeah.

(Mohak Shroff at 00:43:29) I mean, I think so let's see. A few thoughts. One, I think that the idea—I don't know how I feel about the idea that I wake up and I'm interacting with a group, with a set of agents all day. I think, actually, my gut tells me, and I think this is probably all of our experience—so there are things that we think agents will be able to do really well, and there are things that agents should have no business doing for the foreseeable future.

(Mohak Shroff at 00:43:53) And so in that model, I actually think it's a hybrid approach where I'm working with a team of people. I'm in a team of people and agents. I'm working with a team of people and agents. And the boundaries of what work goes where is not that different than it is now in the sense that, hey, this is your job. Like, we all have job descriptions. We all have charters. And then within those charters, this is the stuff that the agents do. This is the stuff that the humans do. And it's wonderful because all of us get to do the work we love, and all of the agents get to do the work that we don't love.

(Mohak Shroff at 00:44:29) Now to your point about human relationships, spot on. Like, I think that, undoubtedly, the value of human relationship, the value of personal connection goes up. I don't think it's just in-person connection. I think it's actually still real connection. Like, look. You know, I'm no expert on the actual AI models that we're talking about. But everything I've heard from people who are experts is that describing these models as thinking is a huge stretch. Describing these models as creating is a stretch still because they are fundamentally transformer models. Right? Like, it's a bunch of existing data being recombined in interesting ways to present interesting—

(Joel Beasley at 00:45:09) Which is what humans do. But I won't go there.

(Mohak Shroff at 00:45:11) I won't go there. This is the point. Right? This is the point. Like, we get down a very interesting philosophical path. But I guess my point is I think that there remains a need. Like, we certainly feel like, given the nature of our platform, we certainly feel like, actually—this is actually an important thing to say. Like, I think the need to know who is authentically human goes up. And so maybe what you're describing is in-person, I would maybe phrase differently. I think that it'll be very important to all of us to know that the person we are interacting with, whether that is remotely or in person, whether it is through text, something they said, or something they wrote, or something that I'm texting them, messaging them, whatever mode of interaction it is, it'll be very important for us to know that they're authentically human.

(Mohak Shroff at 00:45:59) And so, you know, in that spirit, again, going back to the point about these shifts in technology and how LinkedIn tends to approach them, we actually built a verification platform. So now LinkedIn is actually the largest professional verification platform on Earth in terms of both the—

(Joel Beasley at 00:46:16) I get the badge ask all the time. Do you wanna be verified? And I'm like, I don't know.

(Mohak Shroff at 00:46:19) Verify yourself, my friend. That's how you prove you're human. No, I'm—but it's actually really important that I think in that world, platforms like ours need to be able to create trust in the people, in the content you're interacting with to say, hey, this is from a human. Not because that content is necessarily more valuable or more insightful to the thing you want, but because there's still value to knowing and believing you're interacting with humans. And there's something about that that the models don't have. There's something about that that the agents and the LLMs and the GAI technology doesn't have, and we all need it. Right? And so that, I think, is actually where—I'd absolutely concur with what you said—human interaction is going to be extraordinarily important, and authenticity is extraordinarily important.

(Joel Beasley at 00:47:06) And I agree, like, especially at your level, the idea of waking up and only working with agents is very, very, very far off. But the idea of a solopreneur that has one employee or whatnot or an assistant—totally. Their reality is not a stretch to imagine. And then I think it's to your point, because I think you did have a good point. And I think what you were kinda saying was it's gonna be a social cultural issue as well. Right? Like, at some point, someone's going to make that an issue of, oh, are you dating an agent or, you know, oh, are you working with only agents? Like, these things are gonna come up. I mean, I'm not saying I'm inventing them, but I am saying I see these things happening out in society.

(Mohak Shroff at 00:47:50) It's almost inevitable these things will be created by someone for some reason. And in some ways, they'll be good for some people and, you know, not great for other people.

(Joel Beasley at 00:47:59) Okay. So, hey, let's wrap it up here. I wanna make sure we get you out on time. As far as engineering leaders, how are they thinking about the future of AI agents? What's one piece of advice that you'd give all engineering leaders?

(Mohak Shroff at 00:48:11) Yeah. Well, can I do two? Yeah. Alright. I'll do two. So two pieces of advice. One is, I think—well, actually, I'm gonna cheat and do three. Just rule of three. It's important. Alright.

(Mohak Shroff at 00:48:21) First is, I think the most important thing is to optimize for flexibility, optionality. Things are gonna change. You know, conviction's fantastic, and building against conviction is important. That's kinda how you invent the future. But at the same time, what's actually also really important is create the optionality that allows you to, as things become more clear, either for yourself or from the world at large, as you see things being invented or built that's gaining traction or that's—there's an interesting idea, sort of modicum of value there—you can double down on it. And so build for optionality is thing number one.

(Mohak Shroff at 00:48:54) Thing number two is, you know, this piece of wisdom I heard from Kevin Scott, actually, which was that in moments like this, it's always really natural for us to lean on doing the thing that we feel like we understand and sort of making it incrementally better. It's like, I can do this thing that I've done for years and make it 20% better, 30% better, 40% better. But the real magic is in doing the thing that was impossible before. Right? Like, this technology makes possible things that previously were truly impossible or impractical. And that's where real innovation is, and that's where real value gets created. And so that's the second thing. It's really push on—you know, way back when we last chatted, I talked about complaining as something that's a core engineer. Like, I think about, in this manifestation, complaining as a path to pushing on, what are the things I cannot stand about the state of tooling or the world or technology that for once, finally, maybe not perfectly, but finally can be solved?

(Mohak Shroff at 00:50:05) I think there's this massive upside and massive value, innovative value in that. And then the last thing I'll say is really focusing on human centricity. Like, I think it's easy with tech like this to just get geeked out as tech for the sake of tech and technology for the sake of technology. And I think that real value ultimately has to tie back to ways that it helps all of us feel more productive, more successful, more effective, more able to do the things we love, more able to be happy and feel, you know, joy. And it's something that we—all three of these are things that we try to bring every day to the work we do at LinkedIn to AI, to just everything we do. And I would just love for every engineering leader to think about that as kind of the core of how they bring technology to life, especially in this moment.

(Joel Beasley at 00:50:54) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.