I ❤️ AI Q&A: What Is RAG, MCP, NL Web & Zero-Trust Identity?

Oh boy, I don't know what's happening right now. What's up? Good to see you all. Welcome. Happy Monday. Normally I do this call on a Friday, but we are live and excited. Oh good, all my AI robots are in the house. Welcome, welcome. Good to see you all. Awesome. Amazing. Let's see here. All right, I'm gonna go to this. Hello. Oh, look at all— that's nice. What's up, Don? I see a bunch of like pictures in your profile. That's cool. It's got like all the photos you take or something like that. Oh my goodness, note taker after note taker. Welcome everyone. This is our I Love AI call. Um, really exciting. I'm gonna do this and hopefully that will be helpful to you all. Welcome in. Let's Oh my goodness. I'm gonna go check. Oh, good. Oh, good. There's more than just note takers in the house. What's up, Joe? Good to see you, Kim. Hello. Hello. And, uh, some, and I'll say hi to the note takers. Hello. Hello. Feel free to, um, bring your own device. I guess now it's like bring your own AI. Um, but we do, we try to take care if you are, are, hi Kim. Um, if you are not, uh, here and your bot is here, like we, we, We actually will kick your bot out. So I hope you come. If not, this is a moment where I will nicely transcribe, I do appreciate you, think you're amazing. But your bot, you know, can't be the attendee, you have to come and hang out with me live and in person. But I do like to give people a few moments to get here. I don't know what I did. But apparently, I turned on a feature, I always turn on AI features. If somebody has one, I always turn them on, see what they do. Apparently there's a new AI note-taking feature inside of, uh, Loom. And many of you know I love the Loom product. It's a video. It allows you to do screen recordings and videos. And, um, anyway, I'm a big fan. And, uh, so apparently there's a new note-taker app, but I didn't know it was like Firefly and Otter and Rev where they inject a, a Loom bot into your meetings. I'm not a huge fan of that. Anyway, but mine are here, so I'm going to actually just remove one of them since I have two. And don't worry if I ever remove your note-taker, unless it did something horrible, I don't typically report them, so it will not damage your reputation at all. What's up, Patrick? Good to see you. Sherilyn is here. Yay. And of course, Don. All right. So welcome. This is typically our Friday call where I kind of give you a roundup of what's happening in AI leadership from my perspective and point of view. Um, but I'll also talk to you a little bit, uh, today about some of the features and capabilities that I've been working on. Of course, give you an update on all the amazing announcements that are happening in the AI Leadership Institute as well as among some of our friends. Don't forget, uh, you can join us live at, in any of these calls. Um, we get together on Zoom. You can ask questions. We actually do a breakout at the end where you get to meet people live. It's a huge part of kind of learning AI is also creating a network of people that not only can help you, but that you can help. And so it's really a wonderful part of our community. So if you would like to do that, you can go to my name, Noelle, N-O-E-L-L-E dot AI. That's the website. And, uh, you can join us. There are two different ways to be a part of our community. One is in I Love AI. That is our free community. And then we also have a premium community that you can be a part of where we do, uh, an AI startup accelerator. We have Agentic AI for Beginners. We run live learning labs every single month to learn new content. As a matter of fact, was it you, Joe, um, that was talking about MCP? I think it was. Uh, so I definitely wanna, um, I think, uh, what we will be doing. Yes. Right. Okay, cool. So we are gonna do a live learning lab and of course, Joe, since it was your idea, I'm definitely gonna make sure you're there with us. Uh, but we're going to do a live learning lab, um, this coming, uh, in, in this month, which June just started. So welcome to June, everyone. But we're going to do an MCP introduction, MCP, I think, uh, which is basically this protocol that's available for you to do agent-to-agent communication. And I think it's really important because so many people, I mean, even some of the people that I feel like should know, uh, are a little confused about what it is and what it's for. Um, and we're, as we often do, overstating, uh, some of its complexity. So I like to just break it down and give you very simple analogies, but most importantly, put your fingers on a keyboard and actually build something. It is hard. It is not, uh, it has a bar— a bit of a barrier to entry. Um, but, but I'm here for it if you are, uh, if you all are. So for those of you in the AI Leadership Institute community, you will see that live. For those of you in the I Love AI community, you will, unless you come up with these cool ideas like Joe, you will see the recording, um, show up in the, the community so you can watch along. Um, but yeah, it'll be fun. So I think that's, uh, that's this month's is going to be Introduction MCP, uh, so that you can understand the difference between MCP and A2A and what, what are some of the other ones like, um, and frameworks like Semantic Kernel and AutoGen and like there's just so many different approaches to solving the problem, and they're all different parts of the architecture. And so sometimes we talk about things as if they solve the whole problem, but really there's quite a few components to solving an agentic, responsibly built agentic system. And so I think it'd be fun to talk through that with all of you. Don't forget, if you are in the AI Leadership Institute, you do have access to our current cohort, which is Agentic AI for Beginners. I've now been told that it, I mean, it's definitely for beginners, but some of the labs are a little bit more intense. But like I said, I'm here for it. Like, let's dive in, let's learn this stuff together. So hopefully you will, uh, have as much fun as I do. Um, all right, so lemme give you a quick update on what's happening, uh, in my world and in the AI Leadership Institute. Number one, uh, this week I will, today's Monday, so tonight I'll be flying to New York City, uh, to join a very good friend of mine, Erin Reddick, uh, with one of her clients. Science to deliver a, a session, um, tomorrow on responsible AI at scale. Uh, and then I'll leave New York and head over to Silicon Valley, uh, to speak to a company called Sandia. I don't know how many of you know them, but pretty cool. Uh, a research engineering organization. Very excited to talk to them and their leadership about responsible AI. Uh, and then I'll fly from there, which is like Wednesday, Thursday timeframe. I will fly from there on Thursday over to Toronto and I will get a wonderful opportunity to join Women in AI North America for their yearly event. I am also nominated for an award, but that's not why I'm going. I get nominated for stuff on occasion. It doesn't drive my participation. I am 100% going because number one, one of my good friends, Renee Cummings, is the kind of like the honorary— I don't know, she basically is the person they're giving kind of this lifetime honorary award to. Um, so I'm there to support her. She's going to be— I'm seeing the event. I also just love hanging out with women on a similar path that as I am on, um, and, and encouraging them and just being like, let's go, how can I help you? Uh, and really being proof in the pudding, right? Like that women supporting women is a thing, that we do do this. Sometimes it doesn't show up in everyone's, uh, career experience, but anyone who's connected to me, I hope, feels supported and cared for. And so that is one of the reasons why I'll be supporting this event. It's in Toronto. One of the other fun things about Toronto, my husband and I are big believers in, um, cold baths. I'm sure some of you follow me on Instagram. You can see one of my pinned posts is us doing an ice bath. And anyway, our, our ice bath chiller is currently broken, so it's stuck at like 35 degrees, which is very cold. Normally we're like 40, 45 people. Um, so we're stuck, uh, doing until we buy a new one. Uh, we're stuck doing kind of super cold and it's funny how these cold plunges work, right? You get in and I'm like, I told my husband, I'm like, I'm only gonna do 3 minutes. And then 3 minutes comes and goes. I'm like, all right, I could do one more minute, right? Like my comp— my competitive nature kicks in. Um, and then I lose feeling in my toes and my fingers and I get out. But, uh, but we are also very big fans of infrared saunas, of which we also have one. And so we typically go back and forth. Well, there is a place in Toronto. Called Othership. We've actually been to the Othership in New York City. Um, but the one in Toronto I think was the first one anyway. And they have a facility where you can go into saunas and then go into, um, a cold plunge and back and forth. They give you like a 2-hour pass. So I can't wait. Um, not AI at all, I know. Uh, but a little fun. I do believe in balancing out my right and left brain. Um, last night, uh, my husband and I spent some really, uh, like just relaxing, quiet time, uh, together, which was beautiful. We just celebrated our anniversary. Some of you saw that amazing post on LinkedIn. If you haven't, go look for it. Uh, I reposted it, um, and give my husband, uh, some props. Yay. Um, yeah, so, uh, so that's exciting. We met each other. Um, we began dating in high school. And, uh, that was like, I told him, when you put those years out there, man, people can like do the math. I like to keep my age a mystery. Um, but, uh, that was 32 years ago. And, uh, yeah, so now we are celebrating our third anniversary, which means we took a while to circle back to each other and get married. But, um, but yeah, it's really cool. We're very, uh, we've always been best friends. And so it's really cool to be married to my best friend and business partner. Um, anyway, yeah, so thank you for that. We went and spent the weekend. We had a little staycation in our local area, so it was really beautiful. So thank you, Annie. We will be traveling all week together as part of that extended anniversary experience. 3 cities, 2 countries. Let's go. If you are in one of those countries, I hope you can meet me for coffee. I'll be having coffee like 8:30, 9 o'clock in New York, in Silicon Valley. Of course, it'll be near my hotel, so I'll let you know where that is, um, as well as in Toronto. So if you are around, hopefully you'll join me for coffee one of those mornings. Um, it would be really great to see you. It is my favorite thing to do is to meet with people that I originally meet online and that we get to actually meet in real life and realize that, uh, our friendships go, you know, beyond just the screen. Uh, so thank you all for being here. Some of the things that I'm excited about these days, uh, I finally have an opportunity to talk about something I've been working on for the last couple months. Uh, which is the new feature, uh, that has recently launched from Google called AI Mode. Um, I was one of their early testers and adopters and, uh, feedback providers, and it's been really an incredible experience. I am now, you know, I, I have, I don't know if y'all probably can't see this, but oh yeah, you can. So I have some like easy access buttons on my phone and one of them has always been ChatGPT. I've been using it the longest. I've been going in and it's trained on me. It knows what I do, what I say, all my products, all my offers, all my keynotes, all my books. It's got all of that stuff, right? That, that I've given it information on. And I own, you know, I have an, I'm an enterprise user, so it's, you know, my personal, or I should say my professional, my business's tenant inside OpenAI. Um, so we're not, you know, what I do doesn't train the public model, all of that good stuff. Uh, but when OpenAI came out, I really had to kind of shift my habits, right? Like these cognitive pathways in my brain that were just like ChatGPT, ChatGPT. And, uh, and once I got into the habit of using AI mode, I realized there's a lot of like, there's a lot of what I currently do now, even with ChatGPT, that is what I got trained to do with Google, right? Like I got trained to search with Google and there's some things that are just very natural for me to ask Google, and I realized that I actually have always been asking Google pretty semantic type searches, right? I don't just type in a keyword. I kind of ask it the whole question. Tell me about blah, blah, blah, blah, blah, or can you find this for me? And so now with AI mode, that becomes even easier. But most importantly, it provides me a lot of really good responses. So, so that's been pretty fun. If you haven't tried it out now, it just GA'd, which is why I can talk to you about it now. It just GA'd last week. I honestly think it's a great opportunity to just do something that I do with everything is just add it to the list of models that you ask questions of, especially when it comes to kind of general knowledge things like a go-to-market, you know, um, ideas or lead magnet ideas or, um, you know, how to solve a problem, like ways to brainstorm about content, like whatever, uh, you're thinking. Is I'm not using it today as a business tool because it's still in beta and I just have feelings about that. But I do think it's worth trying out if you haven't done so already. So AI mode, it's just a— you go into Google app on your phone or on desktop and you can turn on AI mode and it will give you this very Perplexity meets Google experience. And I'm a big believer in like both and, right? Not either or, not exclusive, just like give me the best tool for the product. Like, what am I trying to do? What's the problem I'm trying to solve? I wanna know the best of breed product that's going to solve that problem. So I'm always on the lookout. So, so take a look at that. That, that was announced the same week I was at Microsoft Build. Speaking of which, hopefully some of you saw this video, but one of the cool things that I uncovered at Build and got a chance to actually create, or I was gonna say build, no pun intended, was kind of thinking through the art of multimodal Retrieval Augmentation or creating a RAG architecture in a multimodal fashion. Extending that even further to one of our topics this week in the class that we're currently teaching on Maven, which is super successful, by the way, we have 155 students. It's very cool. But in the Agentic AI for Beginners class on Maven, we, like, we this week were focused on Agentic RAG. And I thought it was really cool. Like realization, right? Is that what I like about the future of agents, and I don't know if you saw, but I recently was, uh, listed as the number one thought leader in agentic AI because I've been talking about this since, like, since Alexa. We didn't call 'em agents, but multitask orchestration has always been the dream of like virtual assistants, right? Like we didn't get an Amazon Alexa device just to ask the weather and music. Like, The vision, the dream of this product was that you would use it as a UI, a user interface to do everything you needed to do. Book travel, solve your code problems, figure out who did the best in sales this week, find out what product of yours is selling the most. Like that you would just be able to use your voice and have a conversation with a system that could know everything. That is why when we built Alexa, right, the name of these apps that we would put on Alexa, it was called skills cuz we wanted to be able to custom build skills that this thing would know. And I don't know how many of you know, but Alexa was originally, you know, one of the, I don't know if I, this is what I was told when I first started there in 2014. This is the story they told me and I didn't, I just presumed it was true. Now I'm not sure cuz you know, you never know. But back then the reason, you know, I was always like, why do we pick Alexa? A lot of people ask me that. Why Alexa? And one of the urban legends behind why Alexa's name was chosen, one is based on its linguistic value. So ask a linguist why Alexa is a good word, um, to use as a wake word. And they'll say, well, the odds of a false wake with like X and A and all of that, like it's a good linguistic word to use. So, but that's not fun, right? Like, yeah. Okay. Um, but on top of that, Jeff Bezos told a story way back in the day when he was the CEO of Amazon. He told a story about it being like the Library of Alexandria, that he wanted to build this, like, li— this library again, hence the word skills being used to say this is the library of all information, but now we get to teach it information. And because of the way that it's built, it can never be burned to the ground. It can never be lost. It will always be there. And now I think as we move into the world of agentic AI, it is, And I, again, I learned that like in 2014, that was the vision of building these systems. I got a little frustrated over the years because I would tell people, hey, you can use your, you know, a voice-enabled conversational AI application and you could have it do anything. And they would have it do the most simplest thing, right? Like search for facts about dogs or cats or tell them, you know, how to work out for 30 minutes. Or give them a recipe for monkey bread instead of like the robustness that was possible. Like you literally could do anything and we're asking it for dog facts. Like, oh my goodness, what is happening right now? But the same has kind of happened with AI and building these initial AI solutions. And I think it's part of the evolution of building these systems is that with these systems, the first thing we think about is how do we build Right? A, a, uh, a solution. Let's say like, what, I mean, you all know this, look at the note-takers in the house. One of the easiest first use cases was like, just listen to a meeting and take notes, please AI, just do that. And then that quickly turned into not only take notes, but then extract out using entity extraction, right? A machine learning algorithm, go and extract out entities we care about, action items, Assign them to people. Now we're starting to see a workflow developed, right? Listen to the conversation, extract out the entities, use those entities to figure out who's responsible for what. And then maybe if we get fancy, we're going to actually go ahead and create an integration, a Zapier, right? To take that information and put it in as tasks inside of our ClickUp or inside of our Asana or inside of Jira. And the, these are now workflows that are happening pretty easily now. But like back in the day, I had, it was really a struggle to get people to think outside of just the single use case. And I think we're a little bit there now too, is that when I start talking about AI to organizations, companies, large, uh, audiences, they're usually thinking about, well, what's the one thing I could do with AI? And, and while it is good to start with one thing, it is incredibly valuable when you start thinking about what these single units of work can do together. And that is the real world behind agentic AI. But it's a bit tricky, right? So one of the early things we learned, uh, in 2022 and 2023, I was an early beta tester on OpenAI's Codex model, uh, because I was a Microsoft MVP and they were testing it out for its integration ultimately into GitHub Copilot, which had not released yet. But during this time, we were leveraging these models and And you all might remember '23, like late 2022, early 2023, people were running into challenges with these models that the models would basically run into issues. Many of, many of us call them hallucinations, right? They would just start making stuff up. And I like to think of it kind of as like when you start building an AI system to answer a question, if you don't really narrowly focus that question and the ability for the model to answer that question, if you don't provide that level of focus to a model, it's kind of like asking me to recite something I learned in 9th grade, right? Like in the beginning I might be like, I'll start off pretty good, pretty believable. Like, oh yeah, I remember, I remember calculus in the 9th grade. Yeah, yeah, I remember. But eventually I'm gonna just start making it up, right? Like, like based on my best memory. Like, and, and it'll sound pretty good cuz you don't exactly know what happened in 9th grade for me. Like, so many years ago, so I'm gonna just make it up. And the challenge, of course, with making it up is that, um, it's not true, though it sounds good. And this is what generative models tend to do, is they look into their huge collection of data and kind of figure out a predictive answer that sounds pretty good, but it may or may not be true. So in 2022, 2023, a bunch of scientists were like, okay, we need to figure out a way to actually make this better. Like, we can't just have models going off their training data and pulling old information to generate new answers. It ends up in model errors, confabulations, what we now may call hallucinations. So the way we solve for that was we were like, okay, we're gonna build an architecture. It's not, it's not like a, you can't buy it, you have to design it, but it's an architecture. Um, and that in this architecture, we're actually gonna tell the model, do not use your training data, which is the world's data, the internet data, right? Do not use your training data to find the answer. I want you to augment your search to a dataset that I provide, right? Retrieval augmented generation. I'm going to augment the retrieval process of that data to a data source I provide. Last week I demonstrated to some of you, I think I did it in the, it might have been this call, might not have been, but Anyway, that you can right now when you go into the playground and open Azure OpenAI, you can choose an option on every model. Once you've deployed it, there's an option that says add your data. It's a very quick and easy way, and this was deployed years, a couple years ago, but a very easy way to augment the retrieval of the right information. Add your data. And now in preview, You have the ability to point it to blob storage, which is new, and also the ability to point it to like files on your file system. But just having that decision that you make to go, don't use your training data to get the answer, use this data. And if I'm solving a problem for, let's say, a, you know, a software as a service company and they're trying to create a tool that's going to help their sellers, I don't want you randomly going out and trying to find like go-to-market information from the internet that you trained on 2 years ago. I'm like, I want you to go to this information that I just built right now. I want them, their enablement to come from this data, this, these PowerPoint presentations, these Excel files, these Word documents. I want to control where those answers come from. So that is just to solve the single units of work. I want to create a single model to do a single thing. And that's important. We have to start there. Here is my, my hope is that you realize Every agent you build should be like all by itself, completely responsible, safe, scalable, explainable, transparent, right? It should, it should have woven into the fabric of what an agent is. It should have these principles because that will ultimately help it scale. It'll all— a scalable AI agent or a scalable AI system is one that allows you like that embodies these principles so that as you scale, you have understanding and explainability of how it's working, but you also are taking the steps to ensure that it's fair, that it's honest, that it's helpful, right? These, they call it the 3 Hs, right? Harmless, honest, helpful. Let's just make sure that it's at least those 3 things. Here's my concern right now, of course, is that we're building a bunch of these little Agent Tick systems, these little models, and we're not being super careful about how we build them. And now we're— I call them baby tigers, right? You can see them right here. A little bruiser, right? Like we're not being careful about them. We're in baby tiger mode. We're like, man, this model's pretty cool. It does pretty cool stuff. You're not stopping to say, wait, is it— does it do pretty cool stuff for everyone? Is it guardrailed? Am I sure that it's— the information it's providing is grounded? Not every company is asking these questions. If you're listening to the sound of my voice, you might be my people already. And so you already do ask these questions. But I think it's important that we tell other people that these are the questions that we need to ask. And some of those questions are like, how, you know, where do you get your data? What does that data look like? How often do you need it? Who does it belong to? Should you be compensating those people for the data that you're using? Like, there's so many important questions to be asking. As you go from kind of baby tiger mode into scalable AI systems. So let's just make a presumption that you're going to build your single little, your models that do note-taking, that do content understanding, that do document intelligence, that read PDFs and put them into, you know, or read paper. Like I actually have some physical paper on my desk, right? Take physical paper, and digitize them into a system that take maybe driver's licenses, passports, government documentation, turns it into assets that we can use, metadata we can use to process workflows. This is all great, but there are single models that do each of these tasks. In the world of agentic AI, we're talking about connecting all these things together, and in many cases autonomously. How are we going to do that? How are we going to ensure the safety and security and robustness of a system if we don't actually understand how every single individual model works in that workflow? So I encourage you, like, Agentic is awesome and you should dive in. But the reason that I was listed as number one thought leader, the reason that my company was listed as the number one innovation, uh, company in 2025 by Thinkers360, is because we're not just solving these problems, but we are encouraging organizations to completely rethink what this looks like. And I just have learned over the years that we will— if you're not careful, you're going to build a small unit of work, and that's as big as you're going to think. And in the future, I, I have— I'm seeing it already. You're actually going to start to see these small agents become collections of agents. And those collections of agents are gonna be basically the workforce of a single human on the job. And who's gonna define what the human does and what the agents do? Like, somebody needs to do that work. I mean, there's so many, so much reinvention that is gonna need to happen over the next few years, uh, to ensure that companies not only take advantage of AI, but most importantly, take advantage of the humans that they have in their organizations to grow and scale at an even more rapid pace. So when we, this weekend, uh, our topics were trustworthy, which I've already talked to you a little bit about, trustworthy agentic systems, how to build a trustworthy agentic system, and then also how to build agentic RAG or a retrieval, uh, how, what does retrieval augmented generation look like when you're building an agentic system? And I'm so excited about this, right? Because it starts with, of course, those little agents. But now every single model that you built to do a certain task, now you can go back to those models and re-architect a set of tools that they can use. There's actually a design pattern called the tool use design pattern, which means how do you think about what a single model is gonna do in your organization? Let's say its job is note-taking. What else could it use in order to do that job even better? Right now, as we mentioned at the beginning of the session, not only is it gonna take notes, but maybe it calls on a database of entities that it can use to categorize what's been spoken about. Maybe it also has a tool of Jira and Asana so that as it sees action items, it can create tasks and assign them to people, right? Like, how do I think about every autonomous model that I'm using, AI system that I'm using, what tools could I add to its capability to increase the richness of what it's able to do? And then of course, how do I make sure that I do that as safe and res— as responsibly as possible? I do think that it's important to kind of think about responsibility as like a thread that's woven through everything, right? It's not a checklist you're gonna do at the end. It's not a design thinking session you do at the beginning. It's literally a principle of development. And you, your organization has to adopt this in order to ensure, most importantly, the ability to scale at the end of the day. But it was really fun. So we, myself, Yuvika Sharma, I don't know if she's on the call today, but Yuvika and I talked deeply about how to build, Agentic RAG, how to think about retrieval augmented generation when it's not just what happens in each individual agent, but what is the data source that I might want the entire workflow to have access to, right? When I have an Agentic system that maybe has 100 different agents, I will want to think about what the information architecture is of that system before I go and build out all of these agents. And so I have, uh, in my book, um, I don't know if I have one around here, but, uh, in my book Scaling Responsible AI, I have, uh, a big box on like, I think it's in chapter 2 that says information architecture before artificial intelligence or IA before AI. And it's a principle, like, it is very important. That's what retrieval augmented generation, it is the function of thinking about what will your model need before you build it so that you can go and collect that data so that when you're training that model, you give it that dataset. And it could be a bunch of things, right? It could be a SharePoint directory, it could be a blob storage, it could be a dynamically changing website. Like, there's so many different ways that you can provide the right information to your model at the right time so it's able to answer questions correctly, that it's able to provide the most accurate and true version of the truth at the time your users, your employees, your customers are asking your model for help and support. So really, really cool. Yeah, exactly. Exactly. So MCP, that it was kind of like a little segue, right? So MCP is this ability, and I'm not going to do our intro to MCP now, but that's the direction we're going in, right, is Now, what do I need in order to ensure that the communication to those data sources occurs, the communication between agents occur? Like, how do I manage that communication? How do I govern that communication? And that's when protocols come in. And MCP is just that. It is a protocol, which means it doesn't do any— like, the behavior doesn't come from MCP other than the fact that it controls how your agents are going to communicate and how they're going to get access to the systems. That they are connected to. So I think it's worthwhile that, that we dive in and take a look at it. But yeah, so our Agentic AI for Beginners class, I say it's for beginners because anyone can take it regardless of your skill level. We do podcast-style videos to share with you the ideas in the course, but the labs are hard. And so not everyone is going to be at the right, at the right technical depth to do the lab exercises. But this is how I got into Amazon Alexa. Like, I just started building, like, I didn't know what I was doing. I just started building. And then over time I gained more and more comfort and confidence. And so I encourage you, even if you disclaim yourself to be non-technical, like, it doesn't hurt to take a class, especially a beginner's class, and see what you can do. It is often about the muscle memory that you build, right? How do you get your mind thinking in this way that will then enable you to learn the more technical aspects if you choose to?, but I definitely think everyone should understand agentic AI in general because it will— I'm starting to see it already, right? At Anthropic, Meta, um, OpenAI, like they're already starting to build agentic org charts to complement their human workers. So if you're gonna get a job at a company and they give you 6 agents to help you do your job better, it's good if you understand how that works. So I encourage you to jump in there., and communicate and use that. Um, awesome. Oh, oh, awesome. Someone just shared their location with me. Thank you. I will not be going there wherever you are. Um, okay, awesome. So let's see, I'm just gonna run through this, make sure— oh, nice. Thank you, Sherilyn, for posting that. Um, amazing. So Eve, you said the wrapper. What do you mean by that? Did I— no, like the MCP. It was the conversation about the MCP, like, yeah, like HTML was when we first started. Yes, analogous. Yes. Yes. I, I mean, you could go like HTML, TCP/IP, right? Like all of these. I mean, HTML was more of the markup language where TCP/IP was the communication protocol, right? So I would say like in our case, Python would be HTML in this case. TCP/IP would be MCP, if that makes sense. Yeah. Thank you. Yeah. A great, great analogy. Awesome. Yeah, so that's what I'm excited about these days. Uh, I have, I created a new framework. I don't know how many of you, I think I mentioned it last week called the POET framework, which is really fun. Uh, it stands for precision, like the ability for us to now create very precise outcomes for our customers. There's never been a better time for us to collect the right data, to control the efficacy of our models. I will say one of the most, I think, interesting things to demo and test out, if you are maybe a little bit more on the technical side, is this capability that Microsoft released. Oh my gosh, I'm gonna have to look it up. I can't. Okay, maybe one of you will know. It's, they just launched another protocol that allows you to wrap any of your website assets as an agentic service. Does anyone know? I think it was called WML or something like that. I don't know. I'm gonna have to look it up. Um, but it was really cool. I was like, wait, what? So the good news is, is you can now take any, any website, meaning like a, your website that you have, any website, any data source, um, any content, anything that you want, and you can wrap it up so that it's an API that can communicate via MCP. Which is pretty awesome. The danger in that is like, okay, all of a sudden now we're going to wrap everything and we don't really understand how to build these things responsibly. So I'm going to say like, take your time and, you know, build intentionality into that. I think it's— I'm going to go back in and just look up really quick to make sure I, you know, I'm an MVP for Microsoft. They'd be really sad if I didn't accurately portray this information. For them. Oh, hold on, let me find it. Okay, here we go. Um, here we go. All right, let me— and if you find out, just unmute yourself and tell me. Um, but, uh, it's like API. Um, I'm gonna type in Microsoft. Let's see what happens. I'm gonna hope— I'm hoping that they did a good job with SEO and it'll just come right up. Um, I don't see it. I don't see it. It's funny, you know what came up? REST, the REST API. Is that MCP wrapper LIT protocol? No, it's like WML or something like that, or MML something. Um, okay, I got, I got another idea. This is what we do, we work live. Um, if Microsoft, one of my Microsoft program managers, they'd be here yelling at me. Um, but let's take a look. I'm gonna just search for Microsoft. It's a new announcement. Uh, let's see. See if Perplexity knows the answer. There we go. There we go. Announcements. Announcing Responses API. I feel like it's got to be on this page. Um, Let's see what NTT DATA, I'm speaking with them at the end of the month. Sorry. Uh, let's see if I can find it. AI Foundry. I already talked to you all about that. GitHub Copilot, making agents more capable. Oh, I feel like I'm getting close. Um, oh, okay. Now I'm getting distracted, but I have another thing I wanna talk to you all about. There it is. NLWeb. That it is. NLWeb. Hold on. I'm gonna grab the link and put it in the chat. I knew it was something like that. Um, Hold on, let me find you all. Where are you? There we go. In the chat, NLWeb is what it's called. It's basically API-fying everything so that everything can be an agent, which is fine if you know what you're doing. Kind of scary if we're just gonna like API-fi everything. Oh, I sent it to someone's note-taker by accident. Hold on, let me redirect that to everyone. Okay, I'm hitting— there we go. Unfortunately, it grabbed the link, but if I promise, if you click here, it'll take you to the new NLWeb. So this probably— I'm going to say July is a good month. We'll do an NLWeb Live Learning Lab so that we can learn a little bit more about this and get our fingers on it. But it's pretty cool. So it's allowing— so let's say, for example, you want, you know, you're building a sales enablement tool and you want that agent to be able to communicate to the website and find out information from it or communicate to an old application written in 2005. Like, it allows you to leverage basically the power of APIs and MCP in order to use it. So it is an MCP server. It's like an instance. So you can see an example of MCP in action. I think it's pretty cool. I think they have a video on there. I'm not sure, but Anyway, go take a look at it. Uh, so good. Yay, I found it. Um, so I had imagined maybe, Eve, that it might be related to the MCP. Maybe it's an example of it. But anyway, it's good to see, uh, an actual, like, this is working right now. They've got customers that are leveraging it. So pretty cool. Let's see here. Um, so that's, that's, that's one of the announcements. The other thing, um, that I think it's hilarious. So I've done it, you know, I did this call with you a couple weeks ago, I think, right after my time at Build. Oh no, I did it a couple weeks before that, right after my time at the RSA conference. And I was in the, in like on stage sharing the stage with like the security officer at Meta, Anthropic, OpenAI. They were all there. Um, and they were, we were having this conversation and they were talking about what just anecdotally how important it was for you to think when you're building agents, how are those agents gonna interact with you in your environment? And so they were talking about that these agents will become like employees, that those employees will have domains, that those domains will not be employee-based domains. They'll be more like contractor domains. They'll be treated like zero trust entities in the infrastructure. They'll get, um, the principle of least privilege, right? Like those agents will not get free access to everything. They'll have, the least amount of privilege that they need in order to do their job. Um, and that all of that has to be architected, right? Somebody has to think through that. Uh, and good news is, is that a lot of us have done that work through our cloud infrastructure over the last 15 years, but some of us haven't. Some of us are jumping into this world. Uh, and so one of the things that Microsoft announced at Build, uh, last week was also around this concept of, um, building— I, I think we mentioned Entra IDs, or creating like— it used to be called Active Directory, but now their, their mechanism for managing users is now also going to allow you to manage agents as users in that system, meaning you can give them role-based access control. You can identify like what they can do, how they can do it, how long they can do it. You'll get much more fine-grained access to controlling what an agent can do and not do inside a system. So I just thought it was really interesting that I have this conversation kind of anecdotally with like the, the executives over at OpenAI, Meta, um, and, um, and Anthropic. And then like 2 weeks later, Microsoft announces, hey, and by the way, we're gonna give you the ability to manage these users inside of your tenant, uh, on your 365 tenant. I was like, this is, this is coming fast and furious, which means we really need to educate people on what this means. And now we've got NL Web that's just gonna identify everything. Like, oh my goodness. So I, I like to take a, take a breath and just kind of think through what it would take to build a safe and responsible system. And that means that you don't have an agent that is super admin, right? That, that any agent is going to have to be And that goes back to how we started. Every single agent, whether you have 100 or a million, should have an independent— by itself should have an intention, a purpose. Remember the 5 Ps of prompt engineering that I always talk about? Like, it should have a purpose, it should have an audience, it should have a goal, it should have a metric that it's driving towards to determine its level of success. All of that should be existing for every individual agent. So when you start to orchestrate them, you have the ability to qualify whether or not they're doing their job. I mean, it's very similar to— I mean, I guess some of us over the years have maybe gotten a little looser in like the idea of what success looks like in our personal jobs. But I find that humans are the most successful when they are driving towards a target that they not only understand that they helped create, but also that is aligned with the business value of the company.. And so we need to think about that at an agent level now as well. And that's why I think human resources needs to be a big part of your agent workflows. Like when you start building agents, human resources needs to look at who you have as humans and think intentionally, how are you gonna complement that work? So I think it's, it's a really exciting time. All right, awesome. Thank you, Sherilyn, but I think everyone can see that message. So, but thank you. Um, all right, cool. So I asked Sherilyn to remind me, um, so that we have time to get to know each other and hop into a call. So I would love for you all to just chat a little bit about what you think about the, you know, agentic web. Uh, what, what are your insights? What do you, what are you, um, seeing right now? Um, and maybe what you're most, you know, I hate to say worried about. So maybe some of you can say what you're worried about and some of you can say what you're excited about. Should I, I could say like, those with short hair say what you're worried about and those with long hair say what you're excited about. But I'm going to give you a few minutes to get together all by yourself. So if you all want to hang out with us and get to know some of the other AI leaders that are here, some amazing people here, all you have to do is join us on the Zoom. You can do that by going to Noelle, Noelle Russell, joining our free community, and you can be part of the Zoom every single week. I, I'm super excited for that. So I'm gonna go ahead and, say bye to everyone. Hopefully you found the information of what I'm learning this week and what I'm implementing and what I'm teaching. Hopefully you found it valuable. We're gonna go ahead and turn off the recording and break out into breakout groups so we can get to know each other a little bit better here in the I Love AI community. And hopefully we'll see you all very, very soon. Go out there, build something awesome. Get your fingers on a keyboard, be a doer. Not a talker. That's what the world needs right now. We need people who actually know what they're doing, so you could be one of them. Um, have some fun. I'm gonna go ahead and turn off our livestream now, but thank you all for being here. It was amazing. And I'm trying to find my, my window. Here we go. All right, bye.