Scaling AI Agents with Model Context Protocol (MCP) [Webinar]

IT Training and Development Resources
Webinars
Scaling AI Agents with Model Context Protocol (MCP) [Webinar]

For the past year, organizations have struggled with "AI silos" and isolated chatbots that can’t talk to internal data or work across different platforms.

The Model Context Protocol (MCP) has emerged as the industry’s answer. By creating a universal "plug-and-play" standard, MCP allows any AI model to securely access your company's unique tools and knowledge without the need for constant, expensive custom coding. Join us for an expert led guided tour of how MCP can work for you (and a live demo!)

Browse our AI for Developers training courses, including our course on Building Agentic AI with Model Context Protocol (MCP) for private, customized team delivery at your site or online.

Model Context Protocol (MCP) Webinar Video

In this 1-hour session, Manu Mulaveesala, AI expert and technical instructor, demonstrates how MCP is transforming AI from a simple interface into a powerful, integrated workforce.

Join us as we explore:

The End of the Integration Nightmare: Why a standardized protocol is replacing the "custom-built" approach to AI connectors.
The New AI Ecosystem: A broad look at how industry giants like Google, OpenAI, and Anthropic are using MCP to create a world of interchangeable AI tools.
The "Plug-and-Play" Enterprise: How to build a centralized library of data and tools that any AI agent in your organization can use instantly.
Strategic Security & Control: How MCP provides a universal "kill switch" and permission layer to keep your proprietary data safe.
The Road Ahead: What the shift to open-source governance (via the Linux Foundation) means for your long-term AI investments.

Target Audience: Business Leaders, IT Directors, Digital Transformation Officers, and Product Managers looking to build a future-proof AI roadmap.

About the presenter:

Manu Mulaveesala is a veteran technical instructor with more than a decade of consulting, development, and teaching experience in Artificial Intelligence, Data Science, and Machine Learning. During this time, he has taught more than 5,000 students spanning more than 100 organizations, 20 countries, and varying technical backgrounds. Manu provides a dynamic and adaptive learning environment and enjoys tailoring trainings to clients' real-world projects and goals.

00:00:03.250 --> 00:00:27.179
Anne Fernandez: Hi, everybody! Thank you so much for joining us. We are so happy to have you here. We really appreciate the time that you're spending with us. I know you're busy, and I'm here with my colleague, Manu Mulavisala, and he will be talking to you today about the Model Context Protocol, right? Because who wants to be writing a bunch of code to connect to your databases and your systems when you could just use the MCP? Am I right?

2
00:00:27.730 --> 00:00:32.479
Anne Fernandez: Yes. Alright, so we're gonna get started here in just a second.

3
00:00:32.900 --> 00:00:55.680
Anne Fernandez: Just wanted to take a moment to welcome you on behalf of Ascendient Learning, part of Accenture. My name is Anne, and I'll be here the whole time if you have any questions about, like, oh, I can't hear very well, or if, you know, you can't see something on the screen, you can definitely put that stuff in the chat. You can direct your technical questions in the chat to Manu, because I will not be able to answer

4
00:00:55.680 --> 00:00:58.839
Anne Fernandez: MCP questions, I'm here to learn along with you.

5
00:00:58.840 --> 00:01:17.819
Anne Fernandez: But just in case you didn't know, we are Ascendient Learning, part of Accenture LearnVantage now. So what does that mean for you? Well, it means that we still do everything that we do best, instructor-led, hands-on training, live online for individuals, or online or on-site for teams.

6
00:01:17.820 --> 00:01:35.560
Anne Fernandez: We can customize any course for you and deliver it on your schedule. But now that we're part of Accenture LearnVantage, we have access to all of Accenture's resources, like consulting and mentoring, certifications from leading vendors, self-paced courses, skills assessments.

7
00:01:35.560 --> 00:01:48.760
Anne Fernandez: So really just for the full life cycle of your… for your training, and we do work with a lot of our other sister companies, like Award Solutions that does telecom, Udacity, they have nanodegrees, and

8
00:01:48.760 --> 00:01:58.699
Anne Fernandez: an IA MBA program, and we can all work together just to bring you better learning. And we still do a lot of the same

9
00:01:58.700 --> 00:02:11.870
Anne Fernandez: Upskilling in AI and machine learning, data science, cloud, agile, application development, business intelligence, and a lot more. And of course, we do a lot of, AI and machine learning.

10
00:02:11.900 --> 00:02:34.569
Anne Fernandez: And we have amazing instructors, like Manu. I don't know if you've been to one of his sessions before. I will put in the chat, one of the last webinars he did that was great and very well attended that you might like. He's been working for over 10 years in the AI, data science and machine learning space. He does have a background in consulting, development, and technical instruction.

11
00:02:34.570 --> 00:02:59.569
Anne Fernandez: I think it's really super cool that he's taught 5,000, over 5,000 learners, over 100 organizations, and in 20 countries, truly multinational. But the very cool thing about him is he can work with diverse technical backgrounds, he can teach beginners, he can teach experts, he's very dynamic, and I'm not just saying that, it's, I can see it in the student evaluations, so I don't want to embarrass you.

12
00:02:59.570 --> 00:03:09.439
Anne Fernandez: new, but they are very good. And if you get him on the phone, he's really great at tailoring training to your client projects and your business goals.

13
00:03:09.500 --> 00:03:23.029
Anne Fernandez: So after this, this webinar, just some things to think about. First steps could be building generative AI applications, then moving on to the MCP that kind of maps to this…

14
00:03:23.030 --> 00:03:35.750
Anne Fernandez: session, but it's multi-day and hands-on, you know, it's more than a webinar, it's a course and interactive. And then if you want to get a little more advanced, there's Crafting Custom Agentic AI Solutions.

15
00:03:35.960 --> 00:03:46.270
Anne Fernandez: All right, so with that, I will go ahead and stop my share, and Manu, I will let you dazzle everybody with MCP.

16
00:03:47.190 --> 00:03:54.420
Manu Mulaveesala: Awesome. While you were presenting, I was diagnosing some stuff in the chat. Looks like some people are having challenges with audio.

17
00:03:54.420 --> 00:04:02.360
Anne Fernandez: Okay, let me look at that. Thank you, thank you for telling me. I will go ahead and kill my screen, and let me look at the audio issues. Can you hear me okay?

18
00:04:03.960 --> 00:04:08.640
Manu Mulaveesala: Yeah, I can hear you okay, and a lot of people can, so that's what's weird.

19
00:04:08.790 --> 00:04:21.550
Manu Mulaveesala: is that, some people can hear you, and some people can. I try to tell people about, okay, coming in loud and clear, says Murray, okay, now, okay. Strange. Okay. Maybe, yeah, just…

20
00:04:22.060 --> 00:04:39.940
Anne Fernandez: All right, let me take a look and see if there's some settings I can advise if people still aren't hearing. Or maybe it's just me, maybe they can hear you… not hear me, but hear you, which, actually, it's better that they hear you. All right, let me… I'm gonna go ahead, stop my video, I will be here the whole time, and I will be monitoring the chat.

21
00:04:40.740 --> 00:04:42.819
Manu Mulaveesala: Okay, perfect.

22
00:04:43.530 --> 00:04:48.830
Manu Mulaveesala: Alright, cool. Looks like… I was looking at…

23
00:04:50.330 --> 00:04:55.130
Manu Mulaveesala: Alright, everyone can see my screen. Please confirm.

24
00:04:55.850 --> 00:04:56.650
Manu Mulaveesala: Right?

25
00:04:57.220 --> 00:04:57.990
Manu Mulaveesala: Cool.

26
00:04:59.530 --> 00:05:02.489
Manu Mulaveesala: Let me pull up the chat, there we go, cool, perfect.

27
00:05:03.080 --> 00:05:13.930
Manu Mulaveesala: Alright, thanks for confirming, folks. Well, welcome, welcome. Welcome to the Enterprise AI Blueprint with Model Context Protocol. And we're gonna talk about,

28
00:05:14.300 --> 00:05:25.470
Manu Mulaveesala: a little bit about why 2026 is the year Genic AI becomes enterprise infrastructure. This is hosted by Ascendient Learning, which is a part of Accenture Group, and so,

29
00:05:25.470 --> 00:05:50.420
Manu Mulaveesala: what we'll cover today is we'll cover a story in five acts, right? We'll talk about the bottleneck, why scaling AI has been difficult for a lot of different organizations and at many different enterprises. There was a key shift that happened yesterday, but not everyone has quite caught on yet, and we'll talk about what that shift is, what's behind it, we'll talk a little bit about an open standard that

30
00:05:50.420 --> 00:06:05.800
Manu Mulaveesala: That is really, that, that is industry-backed and has gained real, a lot of traction in industry, what its value is, and the path forward, how to move safely, from experiment to production,

31
00:06:06.780 --> 00:06:09.240
Manu Mulaveesala: And Karen, yes, I believe the PowerPoint.

32
00:06:09.510 --> 00:06:32.749
Manu Mulaveesala: or at least the recording will be shared, but… and we'll be able to clarify that in the chat. So, what's the bottleneck? Why has AI scaling, stalled, right? So there's this 95% integration blocker where the main challenge is, right, when ChatGBT… think about it when ChatGBT came out, right? A lot of people, and you may have been one of those people, I myself was one of those people.

33
00:06:32.750 --> 00:06:39.409
Manu Mulaveesala: Thinking that, wow, ChatGPT is kind of cool, but wouldn't it be great if we could actually plug this in with our own data?

34
00:06:39.410 --> 00:06:40.150
Manu Mulaveesala: Right?

35
00:06:41.610 --> 00:06:48.970
Manu Mulaveesala: So, this is a key thing. Then there was a problem, right? The problem was that, okay, well, if we wanted to get it…

36
00:06:48.970 --> 00:07:13.839
Manu Mulaveesala: to see our own data, how do we do it in a way that's accurate, that I'm not just having to attach files every single time? Wouldn't it be great if it was already trained on all our data, or knew all of our data? So there's been various iterations that the field has gone through, and I'll talk a little bit about those, so that you can also, think about those, in terms of how do we actually get AI to talk to and learn and know about our data, so that when I ask something.

37
00:07:13.840 --> 00:07:26.900
Manu Mulaveesala: of my data, it actually gives me an accurate and coherent answer, right? And another aspect has been that there's been an ROI gap, right? So there's a lot of…

38
00:07:26.930 --> 00:07:42.909
Manu Mulaveesala: great ideas that folks have at different levels, whether they're technical or non-technical, but there's been an observable financial disparity between localized AI pilots and actually scaled cross-system agentic deployments. So everyone talks about agents, but…

39
00:07:42.920 --> 00:07:59.389
Manu Mulaveesala: the, you know, where the rubber meets the road is where things are a little bit shaky, right? And 2 out of 3 organizations have not yet effectively scaled AI across enterprise, despite 62% already experimenting with agents.

40
00:08:00.020 --> 00:08:07.679
Manu Mulaveesala: So, let's take a step back, and I'll try to explain this in as simple terms as possible.

41
00:08:07.680 --> 00:08:26.779
Manu Mulaveesala: Right? There's a technical explanation for this, but for this webinar, we're gonna think about it in as simple terms as possible. Think about it like this, okay? I have data in my codebase, in my GitHub if I'm technical, I have maybe Jira tickets, I have PowerPoint presentations, or SharePoint, I have email, I have…

42
00:08:26.780 --> 00:08:39.709
Manu Mulaveesala: you know, any number of different systems. I have my own personal database, I have local files, I have enterprise files, secure files, right? Any number of systems that I need to be able to connect my applications and my software to, or any number of agents.

43
00:08:40.270 --> 00:09:05.259
Manu Mulaveesala: The challenge that a lot of people have run into, and this integration problem is not something that's just new because of AI, it's something that's always existed, right? If I have Model 1, or you can think about it as application 1, or Stack 1, whatever it is, and I have to connect to different data pieces, I have to write individual wiring, right? If you guys have ever had, like, a TV receiver, right, and you have to connect every single thing to

44
00:09:05.260 --> 00:09:09.559
Manu Mulaveesala: everything else, it looks like a nightmare back there, because all these cables are

45
00:09:09.560 --> 00:09:34.499
Manu Mulaveesala: kind of, like, cross-connected and tangled up, and sometimes if you have to unplug or shift something, it's very difficult. And the same kind of thing actually really happens, in the digital world, in the digital space, where you have a bunch of these wirings, and they're kind of shoddy, and… and maybe you're able to build something that connects well to one thing, or to a bunch of different things, but then if you have a new idea, if you have a new product, new platform, you have to connect it again to every single thing, right? And so you have to basically build up

46
00:09:34.500 --> 00:09:38.969
Manu Mulaveesala: Each of those wirings individually, so that creates this kind of, like, lattice configuration.

47
00:09:38.970 --> 00:10:02.990
Manu Mulaveesala: And it results in a lot of redundant engineering, fragile architecture, and in some cases, vendor lock-in, especially when we're talking about large language model providers or AI providers, right? So 83% of the engineering time, up to some studies, like even done as recently as 2025, talk about the fact that 83% of the engineering time is consumed by what we call the integration tax.

48
00:10:03.030 --> 00:10:23.670
Manu Mulaveesala: Right? Not about building AI, but rather wiring up the plumbing around it. And so that's where a lot of money goes, which, when I say money, I mean the resources, staff, developer time, etc. Right? Writing these custom connectors, getting the AI to be able to see our data, and…

49
00:10:23.670 --> 00:10:31.750
Manu Mulaveesala: you know, it's a big headache, and it doesn't allow us to be able to flexibly switch between AI models, or data sources,

50
00:10:31.780 --> 00:10:49.460
Manu Mulaveesala: you know, if you have to build a bunch of stuff and you accelerate quickly, which is kind of the name of the game these days, right? You might have a lot of undocumented integration layers. I'm surprised by how much of this exists in the industry as well. And what it results in is really, like, a lot of blocked AI initiatives, which again.

51
00:10:49.760 --> 00:10:55.140
Manu Mulaveesala: refer back to the ROI gap that I mentioned earlier. So, Act 2. So, we talked about

52
00:10:55.140 --> 00:11:19.080
Manu Mulaveesala: the problem scape, the landscape of the problem space. What's the shift, right? So enter Model Context Protocol, where we're able to connect, these various different tools, whether different AI tools, whether it's Claude, Cursor, Copilot, Amazon Q, there I'm really talking about applications, but you might be using Claude, you might be using OpenAI, you might be using Gemini, whatever it is, whatever floats your boat.

53
00:11:19.080 --> 00:11:26.949
Manu Mulaveesala: or whatever's the easiest AI model that works for you all, right? The idea of the model context protocol was essentially thinking about

54
00:11:27.680 --> 00:11:42.690
Manu Mulaveesala: how can we take different AI clients, which are basically AI providers or AI models, AI applications, and how can we connect them to the things that we actually want AI to be able to talk to, right? So instead of having this lattice,

55
00:11:42.690 --> 00:12:06.700
Manu Mulaveesala: kind of lattice configuration, all the crossed wires, right? The idea is that we can simplify things by having one server, or a couple different servers, and many different clients, and essentially, we transform what is referred to as an M times N problem, the number of things on the left side.

56
00:12:06.700 --> 00:12:18.920
Manu Mulaveesala: times the number of things, if you see that, you know, if each one has to be connected to every other one, your result, the number of integrations that you have to build is a product, is a factor as a product, right? And imagine there's

57
00:12:18.960 --> 00:12:28.039
Manu Mulaveesala: Some places where people have, like, 900 different applications in a really complex company, and yet to wire to all your data sets is very, very challenging, right?

58
00:12:28.040 --> 00:12:51.350
Manu Mulaveesala: So, this is… this kind of, has, like, a real kind of, like, mathematical underpinning. It's not really complex, right? You can just think about it as I have to build a bunch of things on the left hand before, versus building less things, and all I have to do is figure out how each of my pieces connect to this unified standard, this one protocol, which is called Model Context Protocol.

59
00:12:51.350 --> 00:13:06.010
Manu Mulaveesala: If you think about that word, model, model refers to the AI, context refers to what is it that I want to connect that model to, what is the data that I want the AI to be able to see, and the things on the right is the actual context, right? Slack, GitHub, etc. So.

60
00:13:06.010 --> 00:13:17.110
Manu Mulaveesala: the idea of collapsing n times m custom integrations to N plus M is really about making it easier for the AI to see our data. That's really the fundamental shift.

61
00:13:17.150 --> 00:13:24.940
Manu Mulaveesala: Right, and again, folks, remember I am able to take questions in the chat. If anyone has any questions, feel free to pop them in.

62
00:13:28.720 --> 00:13:32.080
Manu Mulaveesala: Just blowing my nose here, because I woke up with a bunch of allergies.

63
00:13:34.540 --> 00:13:49.289
Manu Mulaveesala: Sorry about that. Love the clarity on each word in MCP and reducing the N complexity. Awesome, Mark, I'm glad you liked that. Hopefully, yeah, it was designed to be able to speak to the rest of us, because

64
00:13:49.290 --> 00:13:58.999
Manu Mulaveesala: A lot of times, when I see people talking about MCP, you know, one word that is commonly used to describe this is, like, oh, it's the USB-C for AI. But, like…

65
00:13:59.080 --> 00:14:23.909
Manu Mulaveesala: I hate that analogy, honestly, and I tried. Maybe it's in here, maybe it was a stowaway in the slides, but I hate that analogy because, first of all, every single person talks about the same thing, but no one really explains, like, what it actually means, right? And hopefully, my core… one of my key goals was, okay, just wrapping our head around what the heck is MCP, and what does it mean? But essentially, it's about getting AI to be able to talk to your data. That's really what it comes to.

66
00:14:23.910 --> 00:14:26.569
Manu Mulaveesala: comes down to, right?

67
00:14:26.720 --> 00:14:49.599
Manu Mulaveesala: So, the really wild thing about what happened with MCP was kind of a fluke, because it accelerated so fast, right? Anthropic, who's the model provider behind Claude, made a lot of headlines and crashed a lot of SaaS companies, let's say, in terms of stock price, but anyway, the…

68
00:14:49.670 --> 00:15:05.890
Manu Mulaveesala: Anthropic released an open source protocol in about November 2024. And then OpenAI, which is another big provider, adopted it, and then another provider, and the list just went on and on and on. And eventually, within just one year, basically.

69
00:15:05.890 --> 00:15:20.790
Manu Mulaveesala: the Linux Foundation established the Agentic AI Foundation, and then Google Cloud and AWS launched mainstream MCP support, and basically this exploded onto the scene. It exploded onto the open source space, and then it started getting traction in all of the big model providers.

70
00:15:21.580 --> 00:15:46.550
Manu Mulaveesala: So, a lot of standardization, across AI has started to move soft… faster than software, and you see this because Anthropic or another company will release something, and then all of a sudden, all the other companies will come out with the same thing. Cloud Code comes out with something, the next month, or in a couple months later, GitHub Copilot and Amazon Q and all of the other ones have the same feature, like Agentic Skills, or what have you, right? So, there's a bunch of different things. We do try to cover different

71
00:15:46.550 --> 00:15:55.890
Manu Mulaveesala: types of skills in webinars, and we really get deeper into this in our courses as well. But, you know, the MCP is a really good,

72
00:15:55.890 --> 00:16:20.869
Manu Mulaveesala: aspect to explore and study. There's a lot of interest in it. I know a lot of students that I've taught, you know, even coming into the class, they've heard something about MCP, and they're really interested in using it. And I think one of the hardest things is just wrapping your head around the concept, because actually, the implementation is not so hard, because the pieces, again, have already been wired together. So if you're using most AI providers, you're able to set it up.

73
00:16:20.870 --> 00:16:34.050
Manu Mulaveesala: right away. And you know what? I am just gonna stop sharing for just one second and show you guys a quick demo. Would you guys like to see a quick demo of what allows MCPU to be really cool?

74
00:16:34.050 --> 00:16:35.330
Anne Fernandez: Yes!

75
00:16:35.600 --> 00:16:36.870
Manu Mulaveesala: Okay.

76
00:16:36.940 --> 00:16:43.329
Manu Mulaveesala: Yes, of course, great. Alright, cool, let me… let me wire this up on my screen.

77
00:16:43.380 --> 00:17:06.929
Manu Mulaveesala: So I'm gonna put a new task. I could even do it in chat, right? So here's an example with Claude, and I have some things. In Claude, they call them connectors, but underneath the hood, they're really just, they're really just basically MCP servers, right? If I wanted to click a new connector, you know, you might not…

78
00:17:06.930 --> 00:17:07.990
Manu Mulaveesala: at your

79
00:17:07.990 --> 00:17:32.970
Manu Mulaveesala: company have access to all of these just yet, but, there are a ton, and there's a ton more than even what's, what's available here. You know, you see things like HubSpot, Atlassian, that's super popular, Confluence, Figma for design, for example, Zoom, you can search, recap, and act your Zoom meetings. Whoa, that's kind of cool, I want to add that. And, you know, in some cases, it's as easy as pressing a button.

80
00:17:32.970 --> 00:17:56.009
Manu Mulaveesala: Now, whether or not your company policy allows this is a different story, but there's a lot of progress being made in that space as well. So, for example, all I have to do is add permissions. It's kind of like when I, you know, sign into something with Gmail, and it gives you permission status. Now, you do want to be careful, and we'll talk more about security in just a sec, but,

81
00:17:56.340 --> 00:18:20.930
Manu Mulaveesala: It's really that simple, folks, to wire something new. You can ignore that error. So now, okay, connected to Zoom for Claude. Maybe I'll get to see if I can search that, and we'll try it out in a second. But, in the meantime, I prepared, sorry, that thing keeps coming up. That's because I try to do some things behind the scenes, you're not always gonna see that error. So, let me… I'm gonna do some voice chat. So, let me say…

82
00:18:21.400 --> 00:18:35.730
Manu Mulaveesala: Hey, can you check out the emails and things that I have going on in my calendar today? I think I have a webinar and some other stuff going on. Can you help me prioritize my tasks today, and tell me what are the most urgent to-dos I have to respond to in email, and what I got going on today?

83
00:18:37.540 --> 00:18:38.720
Manu Mulaveesala: So…

84
00:18:39.110 --> 00:18:56.389
Manu Mulaveesala: this is, you know, I think, you know, we're entering a new stage. Because AI is now able to connect to the things that are kind of going on in our lives, which is, again, things like our email, our calendar, our data sets, etc.

85
00:18:56.390 --> 00:19:00.359
Manu Mulaveesala: Right? We're able to then be able to…

86
00:19:00.380 --> 00:19:25.350
Manu Mulaveesala: get a lot more utility and usage out of AI, because it's actually able to see our stuff, right? You know, even as recently as, I don't know, a year or two ago? That was maybe, I guess, two years, we're talking about… one and a half year ago, this was not the case, because this was pre… in the pre-MCP days, I couldn't do something like this, right? Where you can see it's saying, let me pull up your calendar and check the inbox, and it's doing it in parallel.

87
00:19:25.350 --> 00:19:50.139
Manu Mulaveesala: And I can click this arrow, for example, and see, okay, I'm searching the calendar more broadly for April 21st events, extending through the end of April 22nd, and catch anything that might span those dates, and then it's searching my email threads. Sometimes it's slow, but I also have a fix for that, right? In certain places, like Claude, what I can do is I can run, like, a regular artifact, so I actually already had this.

88
00:19:50.140 --> 00:20:14.119
Manu Mulaveesala: run in my day-to-day, but I'm just going back and kind of showing you, so you can kind of see how it thinks. It's lagging a little bit slow. Okay, there we go. Okay, so it's, it's, again, all of this is the AI thinking in the background. Like, you don't necessarily have to see that, right? But it's… you can press the arrow here, for example, and see, what types of

89
00:20:14.120 --> 00:20:38.360
Manu Mulaveesala: of things it's doing is looking for the stuff I got going on, and I'll close it in case there's any client-sensitive things there. Alright, morning, Manu, ran the full triage. Here's your prioritized view. Okay, boom, MCP webinar, I'm presenting, ta-da! Here we are. Right, here's Anne's name, right above Ann's name, and our other coworker from our email has even pulled up the Zoom webinar ID. Oh, yes, I'm missing,

90
00:20:38.360 --> 00:21:02.779
Manu Mulaveesala: my AI director's office… courseware office hours at the same time, but that's fine. I don't think he'll mind. So, there we go. So it's able to get, you know, a set of things that are a high priority, etc. And I could ask, let's see if it works. I don't know if it will, but I'll check it out. Hey, I just connected the Zoom connector. Are you able to see if you're able to access my Zoom transcripts, and see

91
00:21:02.780 --> 00:21:08.710
Manu Mulaveesala: what… what things that I recently talked about, or were shown in my meetings, or,

92
00:21:08.710 --> 00:21:10.689
Manu Mulaveesala: The sessions that were recorded.

93
00:21:11.600 --> 00:21:36.240
Manu Mulaveesala: I'm not necessarily going to wait for it to finish, I'm just curious if it's able to connect, because there's just one thing that I want to show you. A lot of times when you first connect to something, it's gonna say, allow, do you want to allow, Claude or the AI to be able to look into it? So it's going to allow, there we go. You don't have to click always allow, but I'm just giving an example. In the ideal case, where you could just click through, that's great. It's not necessarily the case in many corporate environments.

94
00:21:36.240 --> 00:21:38.780
Manu Mulaveesala: But there is hope sooner or later.

95
00:21:38.780 --> 00:21:51.419
Manu Mulaveesala: Right? Okay, cool, there's, this meeting, this meeting was not recorded or transcribed, but okay, that's fine. But, you know, there, there's probably, some other…

96
00:21:51.570 --> 00:22:15.220
Manu Mulaveesala: Okay, there we go. Cool. Awesome. A few more things, no Zoom recording here, but it did have the idea, right? Set a reminder to pull today's MCP webinar transcript after it ends. Okay, great. So, this is a little bit of the idea. I didn't get to see the questions in the chat if they popped up. What is the role of the MCP, Nishantas? If I take an example of Copilot, I'm doing the same thing without MCP deployment.

97
00:22:15.220 --> 00:22:16.270
Manu Mulaveesala: Unless it's…

98
00:22:16.270 --> 00:22:34.280
Manu Mulaveesala: back-end Copilot is using MCP. Well, you might be… a lot of times, MCP is not necessarily mentioned as MCP. Sometimes, for example, in the Cloud desktop app, it's called a connector. I don't know if you're using, Nishant, if you're using GitHub Copilot or Microsoft Copilot, unless…

99
00:22:34.280 --> 00:22:57.160
Manu Mulaveesala: rock solid with, Microsoft Copilot, but, a lot of times, if it's something like connectors, if it's, connecting to some type of external data sources, like it's pulling Teams, then, like, it's pulling your Teams data, or some other stuff, if it's connecting to those things, and the AI is able to talk to it, and you're able to chat with those things, it's probably using MCP.

100
00:22:57.160 --> 00:23:01.450
Manu Mulaveesala: Okay, so that's, what I'll say about that.

101
00:23:01.740 --> 00:23:25.310
Manu Mulaveesala: Alright, I think it's always fun to do a quick demo. So, alright, so that being said, right, there's actually almost over 10,000 active public MCP servers that have been deployed across the ecosystem. There's, you know, maybe 97 million, let's see here, 97 million monthly SDK downloads within just one year of release.

102
00:23:25.310 --> 00:23:27.059
Manu Mulaveesala: So this has exploded. Now.

103
00:23:27.060 --> 00:23:50.210
Manu Mulaveesala: should you use willy-nilly any of those 10,000 MC… public MCP servers is another situation, and this is… gets into security and things like that, that… that we'll talk about. Vikram posted a question in the chat, just a curious question. Have you… have been hearing a lot about OpenClaw lately? Is it another provider of MCP, type of functionality? How is it different?

104
00:23:50.680 --> 00:23:55.660
Manu Mulaveesala: Yeah, a lot of people are really interested in OpenClaw.

105
00:23:55.660 --> 00:24:20.620
Manu Mulaveesala: I'll tell you a story, Vikram. I personally have not installed OpenClaw yet on my computer. I do have a second computer, so I could boot it up on a second computer. I could use what people talk about as a virtual proxy server and remote server. I get a lot of the stuff done already without needing to connect to OpenClaw. From an enterprise standpoint, OpenClaw is a little bit of a security nightmare, although they've gotten better at it. I've worked on a team where we've built out

106
00:24:20.620 --> 00:24:45.599
Manu Mulaveesala: a custom version of OpenClaw. But essentially, what OpenClaw allows you to do is that it's like a type of orchestrator for your agents. It's not something that actually plugs into your data. For example, it's not the plumbing between your AI agents and your data. Rather, it's a way to, for example, send a Telegram message, or a Slack message, or a WhatsApp message to your agent, and then it runs things

107
00:24:45.600 --> 00:25:10.569
Manu Mulaveesala: in the background and comes back to you with results. And it's a way… it's a different way of wiring your agents, but really, it's more about orchestration, which is to say… which is a fancy word of saying, of just kind of, like, directing and managing your agents to operate, with maybe different data sources you have plugged in. But, for example, I would still need to use MCP

108
00:25:10.570 --> 00:25:17.640
Manu Mulaveesala: servers. If I, for example, wanted to have it go read my Gmail and whatever other integrations, I still would need to

109
00:25:17.640 --> 00:25:19.779
Manu Mulaveesala: Give it access to those things.

110
00:25:19.980 --> 00:25:27.539
Manu Mulaveesala: Javed, we're gonna talk about security risks in just a sec, so don't you worry.

111
00:25:27.640 --> 00:25:41.339
Manu Mulaveesala: I love it, I love it, this chat is blowing up, and I think that's awesome. I'll try to get… we have a ton of people here, so I'll try to get to as many of those as I can. Let's see, we're gonna talk about security in a second, does,

112
00:25:41.340 --> 00:26:02.649
Manu Mulaveesala: Yes, Ajay mentioned security as well. Yeah, Frederick said, so OpenClaw acts as a triggering each agent to perform their set of tasks. Yeah, something like that. It's a means of communication. It's a type of communication protocol, which is different than the type of communication protocol for MCP, but yeah, it's kind of like that. Triggering each agent, or coordinating agents you can think about.

113
00:26:02.650 --> 00:26:26.420
Manu Mulaveesala: What are some limitations of MCP? In what case does it fail? I love it. Keep it coming, guys. I'm going to talk about this and maybe come back to some of the questions if I don't address them as I go forward, but I think that I am going to talk about a bunch of these. So, I think that we have already kind of understood that adoption is mainstream, a lot of leaders are already using it, a lot of organizations are on track towards partial or full agentic transformation.

114
00:26:26.470 --> 00:26:39.940
Manu Mulaveesala: So, an example of this is that a day in the life, before, it might take 9 days, or close to a month, in order to be able… oh, 9 days to close the month, this is an example for operations.

115
00:26:39.940 --> 00:26:49.880
Manu Mulaveesala: Right? Where… or in finance, I know, right? You have to do context switching between a bunch of different data sources, email, Excel, SAP, etc. Reconciliation.

116
00:26:50.010 --> 00:27:14.150
Manu Mulaveesala: in finance and business is a huge headache, to make sure all the bills have been paid, all the I's are dotted and the T's are crossed, right? Whereas with MCP, you could do it in a matter of days. Some people say you could even do it in a matter of hours, but I don't want to exaggerate, but that is the reality, because everything is already plugged in, so the AI has access to all of those things, right? So let me speed forward a little bit in order to be able to think about and see

117
00:27:14.150 --> 00:27:39.099
Manu Mulaveesala: how we can think about security and, talk, about different things. Regin asked, let me see if I can rapid-fire answer some of these other questions. Regen asks, can MCP… does MCP interoperate with other AI agents with different LLMs? Your demo is specific to Claude and its connectors. Well, the demo I gave was just an example, Rajan, but you could do this with GitHub Copilot.

118
00:27:39.100 --> 00:28:02.610
Manu Mulaveesala: You could do it with, many other cursor, a lot of different tools. If they have MCP integration, and it's an AI, like, standard or popular, mainstream AI tool, you'd be able to do it. But you can also, you know, like, you could code up an entire agent and have it, you know, you could code up the protocol, actually, to be able to do it. So for those of you who are technical, I'll send

119
00:28:02.610 --> 00:28:19.619
Manu Mulaveesala: the source documentation to check out if you're interested, for those of you who are technical, but for a lot of folks, it might just make more sense to use an app that already has it, like Claude, or Claude Code, or Claude Desktop, and you'd be able to do it that way.

120
00:28:23.920 --> 00:28:48.749
Manu Mulaveesala: I would not say, Dennis, that MCP… Dennis asked if MCP might be outdated with A2A and new protocols. Would be great to get your insights on this, right? So, MCP, I see it in a different place. I actually talked to our PhD, Dan Gran, who's the AI director, about this, and a couple of other instructors. We were talking about, well, where does A2A fit in with MCP?

121
00:28:48.750 --> 00:29:13.689
Manu Mulaveesala: for example. A2A stands for Agent to Agent. MCP allows you to plug into your data and other sources, right? I imagine that they're probably gonna roll out or advance MCP to also help wire and orchestrate agents together, so you can imagine, like, the next version of A to A and OpenClaw and MCP, maybe it all kind of merges together as one type of protocol like HTTP.

122
00:29:13.690 --> 00:29:37.960
Manu Mulaveesala: I'm not sure, we can only speculate, but I, for example, I personally don't necessarily use that much A2A, because a lot of tools that I already use, like Cloud Code, GitHub Copilot, etc, a lot of them already have kind of parallel agent capabilities baked in, so if I… and they already have the ability to, have agents talk to one another. A2A is something that's a little bit more Google-specific.

123
00:29:37.960 --> 00:30:01.759
Manu Mulaveesala: But, you know, you can think about the utility of A2A as having your agents be talked to one another, be able to talk to one another, so you just kind of have to, you know, check whether the application that you're planning to use, or the one that you're already using is able to do that. And then, you know, you'll be able to do that. So, I don't see A2A as outshadowing, or being newer, or replacing MCP,

124
00:30:01.760 --> 00:30:07.969
Manu Mulaveesala: If anything, they would be complementary. Right, so if I'm deploying MCP, a great question.

125
00:30:07.970 --> 00:30:17.009
Manu Mulaveesala: Wow, there's a bunch of questions here in the chat, I love it, I love it.

126
00:30:17.180 --> 00:30:35.729
Manu Mulaveesala: Okay, let me speed through the rest of these slides, and, you know, I think that, you know, one of the things to ask ourselves is, you know, whatever domain that I work in, or whatever priority use cases that I have, ask yourself a very simple question. What data

127
00:30:35.730 --> 00:30:37.740
Manu Mulaveesala: Would enable a human to do this.

128
00:30:38.170 --> 00:30:48.020
Manu Mulaveesala: You know, like, a very simple question. If I am a human expert in this field, in XYZ thing, trying to accomplish this XYZ use case, what's the manual way that we do this right now?

129
00:30:48.160 --> 00:31:13.129
Manu Mulaveesala: Okay? And then you'd be able to, as you explore that thought experiment, you'd start to explore, okay, what do they have to do? Okay, well, they have to go to the email, then they have to go to Excel, then they have to, you know, enter it into XYZ system here, and then they have to cross-reference here, and when you ask yourself that question and go into that investigation, you'll start to see what are all of the quote-unquote external pieces that I have to connect to.

130
00:31:13.130 --> 00:31:23.399
Manu Mulaveesala: What are all the systems or platforms that I have to connect to? What are those software as a service subscriptions that I have? Right? And essentially, the idea would be to

131
00:31:23.430 --> 00:31:47.819
Manu Mulaveesala: map and wire our agent to be able to have MCP access to all of those, right? So not every organization is going to need to connect to every single MCP out there. In fact, it's not recommended. One of the limitations, as people asked about with MCP is that if you have it connect to every system at once, and every single agent at once has access to every other system.

132
00:31:47.820 --> 00:32:11.950
Manu Mulaveesala: you're gonna get what's called context overload, where each agent is just flooded with so much information, it doesn't know what to do, or it gets kind of confused, or it spins in circles, and this is kind of the same thing. If you have one person that has to wear many different hats, it's kind of overwhelming. In the same way, if an agent has connection to a lot of different data sets and has a lot of different functions it has to do, that's kind of challenging. And so this is… if you have situations like that, or you're in a complex enterprise.

133
00:32:12.440 --> 00:32:35.839
Manu Mulaveesala: typically, you're going to want to roll out to a multi-agent system, which means that you have multiple actors wearing different hats, I'm the security hat, I'm the developer hat, I'm the product hat, I'm the platform hat, or whatever. I'm the Slack hat, you know, and they're able to, you know… but it doesn't mean that you have to have… only one agent can only use one NCP, it's not really like that.

134
00:32:35.840 --> 00:32:59.160
Manu Mulaveesala: But I would try to have, like, a limited blast radius or a field of scope for each of the agents or actors that you're gonna use, and try to compartmentalize and, you know, the same way that you would split up a team into different roles and functions. In the same way, split up your agents to have specific roles and functions, even more so than you would with a human, actually.

135
00:32:59.160 --> 00:33:22.779
Manu Mulaveesala: And then get those agents to be able to talk to one another, or specify in your system prompts, or your core baseline instructions, which agents should talk to one another, and what, you know, those agents in order to fulfill their task, what connections do they need to have? What integrations do they need to be, connected to? And that's where you're able to kind of, like, wire up the MCPs that they connect to. But again, the cool thing is.

136
00:33:22.780 --> 00:33:25.690
Manu Mulaveesala: That, because you just have to basically

137
00:33:25.730 --> 00:33:49.190
Manu Mulaveesala: in some cases, as easy as clicking a button, and in other cases, maybe you write code, you know, and the cool thing about MCP is that you don't have to, you know, only use, like, a set of existing things, like, oh, if there's a Slack MCP connector, I can connect to Slack. If there's a Jira connector, I can connect to Jira, but if I have a private database that's in Redis, or Amazon AWS, or whatever.

138
00:33:49.190 --> 00:34:14.169
Manu Mulaveesala: doesn't mean I can't connect to it. No, you can actually create… your developers can actually create your own custom MCP servers for whatever data fits your delight, right? So, that's the other kind of net bonus of MCP. Not only does it provide standardized protocols for everything, and then Google becomes responsible for maintaining and updating its MCP connection protocol system, and Slack for its own, and, well, if it's your own custom.

139
00:34:14.170 --> 00:34:29.529
Manu Mulaveesala: are going to be responsible for it, right? But the idea is that your MCP is depending on yours. I actually think that I reset the timing on the chat. How many minutes do we have? I can't see my screen, let's see.

140
00:34:30.330 --> 00:34:32.649
Manu Mulaveesala: Let's see, we have… what time is it?

141
00:34:33.139 --> 00:34:36.169
Manu Mulaveesala: I'm on full screen everywhere. Okay, cool, we still gotta look…

142
00:34:36.170 --> 00:34:37.580
Anne Fernandez: 24 minutes.

143
00:34:37.580 --> 00:34:38.849
Manu Mulaveesala: Okay, perfect, perfect.

144
00:34:38.850 --> 00:34:45.229
Anne Fernandez: I mean, you know, don't worry, it's not like everything's gonna shut down in exactly 24, but I know people have stuff to do, too, so… anyway.

145
00:34:45.230 --> 00:34:53.110
Manu Mulaveesala: Yeah, yeah, yeah, I was just checking, because I stopped and re-shared, so I realized my presentation timer wasn't going the same way, but I think it's not too bad, okay.

146
00:34:53.170 --> 00:35:16.680
Manu Mulaveesala: So, you know, I think that a lot of organizations, a large majority of organizations, tend to set efficiency as a primary AI objectives, but one of the things we see with AI innovators and high performers, and those seeing the greatest financial impact, aka ROI, are using some of the core secure agent foundations to push past things like cost reduction.

147
00:35:16.680 --> 00:35:41.679
Manu Mulaveesala: By explicitly targeting top-line growth and transformative innovation. And so, and this comes from McKinsey State of AI 2025. So, you know, I think that, you know, think about what are the things that are already a pain in the neck? What are the things that already, you know, require a bunch of things? That's kind of like a baseline, cost reduction. But then, where are we headed? Where are we growing? And as, as you

148
00:35:41.680 --> 00:35:47.440
Manu Mulaveesala: Start to, you know, work through different types of scenarios or use cases.

149
00:35:48.190 --> 00:36:13.159
Manu Mulaveesala: try to build, even from the get-go, with ROI in mind, right? This is, I think, a key distinguishing factor with high ROI mindset. 39% of organizations, though, are already reporting enterprise-wide EBIT impact from their AI initiatives, so that's good news. We're on the right path as an industry and across different industries. One of the things that Gardner mentioned recently is that AI fundamentally

150
00:36:13.160 --> 00:36:38.160
Manu Mulaveesala: changes the interface, but not the fundamentals of security. So now we're going to get into that good little bit about security. So I'm going to spend a little bit more time on this, because I know some people… multiple people mentioned it and had questions on it. Agents, in some cases, may overshare their data through chat tools, and this kind of comes back to this idea of not giving an agent access to too many different MCP servers

151
00:36:38.160 --> 00:37:03.159
Manu Mulaveesala: data sources at once, right? There's a principle in security called Principle of least privilege, if any of you have heard of that, right? And so, you know, the way to mitigate this in the enterprise is that you… we have what's called human in the loop, right? Because an agent, one of the core ways that it functions is through a loop. One of the cool things about an agent, like, for example, when I gave that example of, oh, hey, go search my email, it starts to think, it looks at my

152
00:37:03.160 --> 00:37:28.140
Manu Mulaveesala: or it does things, and it actually continuously is performing actions on my behalf, like going and checking multiple data sources, thinking about what it should do, drawing on a plan, taking the prompt or input that I gave it, and mapping out the tasks and output that it needs to generate for me, and kind of… the way that it works is that it plans, reasons, observes, and acts, and it runs that kind of paradigm in a loop until it's able to

153
00:37:28.140 --> 00:37:36.839
Manu Mulaveesala: achieve whatever task I provided it. That's why it's like an agent, because it's able to do some of the stuff more autonomously, instead of me having to hold it

154
00:37:36.870 --> 00:37:52.789
Manu Mulaveesala: it's hand, right? Like Chachi BT, in the first days, I'd have to say, hey, can you tell me how to do this? Oh, okay, cool, it gives you a bunch of stuff. Okay, but what about this? And then you have to kind of, like, handhold it to do things, whereas in the new agentic paradigm, the goal is that you do more with less.

155
00:37:52.950 --> 00:38:05.300
Manu Mulaveesala: Right? But there are, again, risks of that, right? Is that as we… there's this concept of degrees of autonomy, so as we increase the amount of autonomy that an AI system could potentially have.

156
00:38:05.500 --> 00:38:20.069
Manu Mulaveesala: we increase the risk that the AI does something we don't want it to do, because we are not necessarily, you know, having a fine-tooth comb or a magnifying glass in every single moment that it's doing that task, because the goal is that it

157
00:38:20.070 --> 00:38:44.870
Manu Mulaveesala: it does things and infers the steps that it needs to take in order to be able to achieve some objective, right? We at Ascendi, in some of our courses, define agentic AI, or AI agents as goal-directed adaptive behavior, right? You have a goal that you're trying to achieve, right? And you adapt your own functioning and run in a loop in order to be able to reach and meet that goal, the same way that a human kind of does.

158
00:38:44.890 --> 00:39:08.519
Manu Mulaveesala: Right? If I give a junior engineer, a person on my team, a task, right, I kind of expect if they have some experience, and I don't have to hold their hand, and they've been at the company for a little bit, that they're gonna go figure out all the steps, even some of the things that they don't know in order to be able to solve that thing, and the same type of thing we're trying to replicate with our AI agents. But that being said, if we overshare our data through either chat tools or AI tools.

159
00:39:08.520 --> 00:39:33.390
Manu Mulaveesala: we can potentially risk, things like data sensitivity leakage and things. One of the really interesting things, like, I don't know if people have heard about some of these security leaks recently that happened at big companies like Anthropic, where they leaked some source code. I heard about one recently at Versal, which, again, is another big name in AI. I heard one at Lovable, so… and these are companies where they're, like, AI-first companies, and AI is a

160
00:39:33.390 --> 00:39:36.280
Manu Mulaveesala: quote-unquote, core part of their brand. So…

161
00:39:36.280 --> 00:40:01.170
Manu Mulaveesala: it's one of the interesting things where security has become a really big thing. I don't know what group of hackers out there hacking every single AI platform out there, but, it's something that is happening quite a bit, so security is a real challenge now. And so part of those things is minimizing the blast radius. This is one of the challenges, for example, with OpenClaw. You know, a lot of different cybersecurity firms say OpenClaw is a nightmare, but I like to have diverse

162
00:40:01.170 --> 00:40:25.169
Manu Mulaveesala: perspectives, so I've also, for example, seen conference talks by the founder of OpenGlaw, Peter Steinberger, I think that's his name, or I don't know how to pronounce his last name, but his name is Peter S, and he talks about, oh, well, everyone talks about security, this and this, and he kind of, like, scoffs at the idea of security, but there are some real challenges with security, let's be real. And NVIDIA and OpenAI and a lot of these companies are on board trying to do security.

163
00:40:25.170 --> 00:40:36.799
Manu Mulaveesala: Claude released a new model that detected software that had issues in Firefox browser that are 27 years old. So we're basically, you know, to, like, distill all of that

164
00:40:37.570 --> 00:41:02.209
Manu Mulaveesala: I don't want to say noise, but distill all of the buzz. What I would say is stick to the fundamentals, understand the fundamentals of security, and the way that you already have it, and what are the principles that you already have, and recognize that AI… I mean, security in the age of AI has a couple more elements of complexity. However, the fundamentals about how

165
00:41:02.210 --> 00:41:07.580
Manu Mulaveesala: A lot of people get access to your file systems and things in your,

166
00:41:08.190 --> 00:41:11.879
Manu Mulaveesala: Things in your network, etc, are…

167
00:41:11.880 --> 00:41:36.769
Manu Mulaveesala: are fundamental. But that being said, if I have the choice to build a custom data integration or an MCP, I almost always would use MCP because MCP, because of its open source traction, has a lot of elements that are wired and built for security. Does that mean that it's guaranteed to be secure? No, that's not necessarily the case, but the idea is that the protocol itself is

168
00:41:36.770 --> 00:42:01.670
Manu Mulaveesala: is built for security. Another example of an instant pattern, overprivileged tokens reaching production systems. So, the way that you want to address this is that you want to have file system sandboxing and lease privilege across scoping. I mean, principle, the least privilege was something that was good in security, and it's something that… it's the fruit that keeps on giving, and it's still something that's useful today. So, one of the things that I'll share is, if you…

169
00:42:01.950 --> 00:42:08.330
Manu Mulaveesala: Want to learn more about, Check out these frameworks.

170
00:42:16.240 --> 00:42:24.950
Manu Mulaveesala: there's a couple of common things that we teach, and we also have cybersecurity, AI and cybersecurity courses,

171
00:42:25.830 --> 00:42:50.800
Manu Mulaveesala: But I'll put a couple of these frameworks that you can check out. You can even feed it to your AI and tell it to teach you about it as well. So there's a couple of different types of things, things like data poisoning, prompt injection. So there's those aspects of kind of learning how things have changed. One of the big challenges with security in the world of AI is that when we talk about cybersecurity, for example, there's this idea of

172
00:42:50.800 --> 00:43:15.060
Manu Mulaveesala: threat surface, right? Where does a potential actor or threat actor or attacker have access to? And the challenge is that the threats, what I think about as the threat surface has greatly increased. And so, the way that I think about it is thinking about establishing when I'm consulting with organizations about establishing security at their organization, is thinking about, you know, kind of concentric circles

173
00:43:15.060 --> 00:43:27.310
Manu Mulaveesala: or tiered levels, where you have, like, limited blast radius, right? Right? Every single system, every single wiring and plumbing,

174
00:43:27.310 --> 00:43:46.770
Manu Mulaveesala: and age at the level of even agents and systems, and the people that are using certain tools or platforms that are agentic, etc, try to minimize the blast radius. For example, if your company is just getting started with AI, I would recommend don't just go ahead and build a customer support chatbot with an agentic system that you have.

175
00:43:47.140 --> 00:43:58.070
Manu Mulaveesala: Because maybe you haven't necessarily actually really dived into what problems might exist. So, you know, one of the easiest, simple things that I recommend people to start with is

176
00:43:58.070 --> 00:44:08.719
Manu Mulaveesala: you know, start the exploration of agentic AI and AI systems, especially the newer and latest and greatest things, with internally facing tools first.

177
00:44:08.730 --> 00:44:17.309
Manu Mulaveesala: before you expand to a public interface, right? So that's another type of way to kind of minimize the blast radius.

178
00:44:17.500 --> 00:44:40.799
Manu Mulaveesala: So, you know, there's this integration paradox where most enterprises are funding custom integration work, but still lack a governance framework to scale it safely. So, in some ways, this is a nice thing about MCPs, is that you can have a little bit more of a centralized governance. Right? I had a conversation recently with a mid-sized company, like, a really deep training,

179
00:44:40.800 --> 00:45:03.680
Manu Mulaveesala: where, you know, I was talking… one of the members of the class was the CTO of the company, and he was really worried about security, and we were talking about, okay, you know, how can we set up a system of, you know, approved MCPs and set organizational policies? There's a lot of different ways to do this, whether you use GitHub Copilot, or Cloud Enterprise, or Google, AWS, every… every…

180
00:45:03.680 --> 00:45:26.710
Manu Mulaveesala: system or provider has their own version of it, so part of it is understanding what capabilities each one has, which ones we like, do we have the budget and the scale to be able to avoid vendor lock-in by being… using the best of different worlds, and having a conversation with your teams, right? And understanding not just, oh, we need to

181
00:45:27.130 --> 00:45:51.909
Manu Mulaveesala: our governance is shut down every AI capability. I've heard of that happen at a lot of companies, too. But, you know, I think that, you know, having crosstalk between the folks establishing governance principles and the folks that actually need to use the AI in their day-to-day, being able to open those channels and understand, okay, what is it that we actually need to do? What are core vital business functions? What are the types of things that are going to drive the highest ROI?

182
00:45:51.910 --> 00:45:56.160
Manu Mulaveesala: That we need to get off the ground right away, and building your governance

183
00:45:56.160 --> 00:46:08.909
Manu Mulaveesala: pipeline around your business, rather than your business building around your governance, right? And I think that a lot of folks, when I've talked about that, have resonated with that, or have mentioned the same. So,

184
00:46:08.910 --> 00:46:30.019
Manu Mulaveesala: The path to agentic in enterprise, is, right, you know, we are moving out of age 1, which was just siloed, all AI lives in isolated chatbots, highly manual integration, in some cases higher risk as well, because, you know, what if, Joe Schmoe attaches, you know, some course, some

185
00:46:30.020 --> 00:46:53.940
Manu Mulaveesala: core database or some repo and exposes it publicly on a public API, right? Now, people have started getting enterprise subscriptions of AI, they started standardizing things more, adopting policies, for example, GitHub Copilot, as an example, allows you to do organizational policies. These are organizational policies, these are team-level policies, these are project-level policies, and these are personal policies, and you can

186
00:46:53.940 --> 00:47:03.459
Manu Mulaveesala: Have, like, a tiered, system of policies, as well as, you know, relevant system prompts across the board in each of those things.

187
00:47:03.460 --> 00:47:15.760
Manu Mulaveesala: Right? And we, you know, we're seeing rapid adoption of MCP, but not just MCP, but really agentic AI and tools that leverage different data sets, right? So,

188
00:47:15.900 --> 00:47:40.020
Manu Mulaveesala: autonomous, enterprise architecture, well, Agent Core is an example, from AWS, but there's many different ones, where we think about how do we have centralized identity, isolated runtimes, and also how do we have monitoring? How do we actually, you know, observe… have observability, logging, and monitoring baked into our overall system, so that if an incident happens, if there's a security thing that happens.

189
00:47:40.020 --> 00:47:47.170
Manu Mulaveesala: If our AI model starts hallucinating or starts producing toxic outputs, how are we able to see all those things happen? Right, so there is a lot of stuff.

190
00:47:47.230 --> 00:47:49.139
Manu Mulaveesala: There is a lot of stuff.

191
00:47:49.140 --> 00:48:13.330
Manu Mulaveesala: But what I try to tell people is that whatever new complexity gets added, or new trend comes out in AI, find a way to bring it back to the fundamentals. Find a way to bring it back, and distill it, and distill it, and distill it, so that you're not just expanding your scope and learning new things, and being overwhelmed and flooded, but try to see how it fits back, and how to… try to see how it fits back in a way that works and is secure.

192
00:48:13.330 --> 00:48:33.169
Manu Mulaveesala: try to see how we, you know, we can leverage these new tools and recognize that a lot of these things are new trends. And so one of the challenges with the age of agentic AI and building within an organization is that it's a… everyone wants to build a house because, aka their agentic AI…

193
00:48:33.170 --> 00:48:57.230
Manu Mulaveesala: arsenal or platform, but at the same time, the stuff is changing so quickly and advancing so rapidly, and we don't know if there's any slowdown in sight. And so, one of the challenges is… it's kind of like you're building a house with a constantly changing foundation. And so, this is where agility, flexibility, building your systems to be modular and robust and all of those things

194
00:48:57.230 --> 00:49:05.609
Manu Mulaveesala: become extra important, which already were kind of principles that already existed, but some of those core principles become even more important. So try to

195
00:49:05.610 --> 00:49:28.640
Manu Mulaveesala: you know, why are your architectures, your plans, your use cases around the idea that these technologies are changing, right? Today, Claude, OpenAI, and whoever are maybe the front, right? But maybe it's like the dot-com days, or the search days, where Yahoo Search and ASGs were the top of the town, and then the next day, Google comes along and wipes everything, wipes the floor, right? We don't know.

196
00:49:28.670 --> 00:49:49.970
Manu Mulaveesala: We don't know what we don't know, we don't know what's to come. So, the best thing that I can recommend is try to build workflows and, and understand things based on first principle, and get your teams to understand things based on first principles. Those are some of the… that's how we try to operate here at Ascendian and Accenture, to try to get people to not just

197
00:49:49.970 --> 00:50:14.939
Manu Mulaveesala: understand the concept, or understand the tool, understand the shiny new toy, but actually understand what does it solve under the hood? What does it enable you to do? What are the first principles that are behind this? Why is this needed? Why is it even worth considering or using, and what is the standards and safety around it, right? So, I'll leave you this with a couple of final thoughts. Three questions for your CTO tomorrow. Are we building custom APIs for AI?

198
00:50:14.940 --> 00:50:17.110
Manu Mulaveesala: Or are we standardizing on MCP?

199
00:50:17.810 --> 00:50:42.769
Manu Mulaveesala: Number two, do our agents have identity-aware authorization at every call, right? Do we have access controls baked in? How are agents being managed right now? And then, where's our integration graveyard, and how fast can we pave over it? What is the integration graveyard? Where, you know, hey, you have a bunch of connectors built and GitHub repos that are just sitting there not being used? But someone somewhere built them somehow, right? And they were intended to do

200
00:50:42.770 --> 00:51:07.729
Manu Mulaveesala: something, and they were a good idea at the time, and maybe they still are, but it was just really hard to do before, and it got up in. But now, there's a way to reconsider old conversations and old ideas in a new light. And so, these are a couple of the ideas that we can ask. The fundamental kind of, like, take-home is stop rewiring your integration, and when we talk specifically about MCP, stop rewiring your integrations and shift your enterprise to a

201
00:51:07.730 --> 00:51:18.470
Manu Mulaveesala: N plus M architecture, and, this is, a couple of sources, and thank you all so much for,

202
00:51:18.750 --> 00:51:29.620
Manu Mulaveesala: For your questions, and if you want to take a screen share, and I think that you had a couple things that you wanted to share, in the meantime, I'll take a look at the chat, and so we can do the question and answer round, yeah?

203
00:51:30.270 --> 00:51:40.120
Anne Fernandez: Yeah, oh my gosh, thank you so much for that. That was really informative. But I also really love the questions. These were really good questions. I think this is one of the best

204
00:51:40.240 --> 00:51:42.690
Anne Fernandez: Question groups that we've had.

205
00:51:42.690 --> 00:51:44.710
Manu Mulaveesala: Totally, totally, it was awesome.

206
00:51:44.710 --> 00:51:52.569
Anne Fernandez: And we love that. Yes, I was just gonna… let me see if I could just share my screen here for a second…

207
00:51:52.940 --> 00:51:58.899
Manu Mulaveesala: If I had the time to teach all of Agentic AI in just a… just an hour, that was kind of the vibe.

208
00:51:58.900 --> 00:52:00.179
Anne Fernandez: Oh, that was, like…

209
00:52:00.290 --> 00:52:25.119
Anne Fernandez: And I'm sorry if we don't get to all the questions, you might have more before the time is up, but maybe we could, I don't know, maybe we could have a part two or something. Anyway, can you, can you see my screen? It's the course for Building Agenda Guy. Okay, so the only thing better than one hour shooting questions at Manu is two days to shoot questions at Manu. So this, this course probably maps best

210
00:52:25.120 --> 00:52:39.349
Anne Fernandez: to this webinar. It's all about the MCP. It's not available for public dates, but if you've got a team, we can customize it and teach it for you on your own schedule. And I'm gonna go… I'll go ahead and put this…

211
00:52:39.350 --> 00:52:43.720
Anne Fernandez: URL in the chat. We'll give this to you, plus some other

212
00:52:43.740 --> 00:52:49.039
Anne Fernandez: courses that kind of map to this, and also some free resources. There's another…

213
00:52:49.090 --> 00:53:08.149
Anne Fernandez: webinar that Manu did on the ROI with coding agents. That was super interesting, too, so I'm gonna put that… that's in the chat, and that will be emailed to you. This recording will be emailed to you. Manu, I don't know if you're sharing the slides. If you are, we can also push those out.

214
00:53:08.790 --> 00:53:11.010
Manu Mulaveesala: Yeah, yeah, I'm happy to share the slides, totally.

215
00:53:11.010 --> 00:53:21.959
Anne Fernandez: Okay, okay, so we'll do the slides, this recording, some of the courses, and some of the other free resources. We've got a blog, we've got Manu's other webinar that he's done,

216
00:53:22.140 --> 00:53:37.629
Anne Fernandez: And… yeah, we'll go from there. So let's see, I can go ahead and maybe stop my share and just see if any other quick questions came in. Yeah, no, it was a great session. Everyone's… everyone loved it. I could tell, great participation, so that's awesome.

217
00:53:37.810 --> 00:53:38.660
Anne Fernandez: Let's see…

218
00:53:38.980 --> 00:53:40.349
Manu Mulaveesala: the more CA

219
00:53:41.510 --> 00:54:05.310
Manu Mulaveesala: Yeah, so a couple of them, so can you post the questions that were in the slide? Yes, you'll have access to the slide, so that should be fine. Okay, a lot of people giving love for the session, that's awesome. Real-time application use cases where we can use AI agents. I mean, there's a ton. You know, Kieran, tell me what industry you're in, and I can try to, route some things off.

220
00:54:05.310 --> 00:54:16.499
Manu Mulaveesala: Have you seen any research such as Gartner or Forrester that would provide guidance on how to best start an agentic reference architecture or a platform concept, or is this an area with continuous innovation where we'll continue to see

221
00:54:16.500 --> 00:54:18.260
Manu Mulaveesala: disruption.

222
00:54:18.350 --> 00:54:43.170
Manu Mulaveesala: Yeah, there's a ton of, research with Gartner, Forrester, MuleSoft, you'll actually, in the, in the slides itself, you'll see references, so I would start maybe there, and start with those references. Some of them include things like State of AI by McKinsey in 2025, then there's a bunch of different things. I think Accenture has some things, too. What I'll see is if I can see, if I can pull, since we're Accenture, we'll pull some of those things.

223
00:54:43.170 --> 00:54:45.710
Manu Mulaveesala: By the way, Anne, is my thing backwards?

224
00:54:46.580 --> 00:54:53.149
Anne Fernandez: Oh yeah, it is. I didn't notice the whole time, though. I was so enthralled with MCP, like, I totally didn't even notice that.

225
00:54:53.150 --> 00:54:55.100
Manu Mulaveesala: I'm sorry about that, guys. That's okay.

226
00:54:55.100 --> 00:55:10.589
Anne Fernandez: You know, I'll do some research, too. I'll see what resources I can find from Accenture and from other places, and I'll send them to you, Manu, but we'll also be sure that we put them in the follow-up email with all the resources as well. So that's a… that's a great question, and let me… let me see what I can pull for you.

227
00:55:10.930 --> 00:55:22.139
Manu Mulaveesala: Okay, great. And Frederick mentioned with… that said, creating specialized agents, would it make more sense to create agents through different softwares, native AI functionality, instead of having an orchestration tool to connect all the agents?

228
00:55:22.140 --> 00:55:45.760
Manu Mulaveesala: There's a couple things with that question that I'm not sure I understand, or maybe I'm too deep in the weeds to think through some of the nebulous nature that a lot of people are thinking when they're first kind of entering, and that's not a jab, that's, like, that is me trying to put my head… to put myself in the learner's shoes, which is something I do a lot.

229
00:55:45.760 --> 00:56:10.639
Manu Mulaveesala: But, you know, what… the way… the best way, Frederick, that I can think to answer your question is basically that, think about it like this, okay? So you have a bunch of different softwares, right, and there's also different AI softwares, right? Each… a lot of those, at least, main AI softwares allow you to plug into a lot of your other data sources. So that's why in MCP, we talk about clients and servers.

230
00:56:10.640 --> 00:56:33.689
Manu Mulaveesala: clients are the things like Claude, or Anthropic, or OpenAI, and whatever things that provide AI and allow you to connect to other data sources through MCP. And even your own organization can build custom MCP, so you could build your own clients as well. But the things that you actually connect to are typically data sources. That's the… I think that that's where you get the most utility.

231
00:56:33.690 --> 00:56:58.389
Manu Mulaveesala: But I think that a lot of people have questions on agentic orchestration, and that's actually, you know, a note, and let's take that back as a note. I think a lot of people are interested in learning more about agentic orchestration, so that could be another webinar. That's kind of, like, what I'm seeing as, like, a trend from all these questions in the chat. Also, if you can save the conversation history, I think that would be really cool, because there's a lot of good gems in here about what people are interested in.

232
00:56:58.390 --> 00:57:15.470
Manu Mulaveesala: Is MCP currently too immature to be a universal standard? It has already become a standard. It went viral and became a universal standard very quickly. It's much bigger than A2A and a lot of the other standards that have come out. If there's, like, one standard or protocol to learn first, I would say it's MCP.

233
00:57:16.940 --> 00:57:21.010
Anne Fernandez: And we see that with clients, too. It's one of the number one questions that we get.

234
00:57:21.310 --> 00:57:26.230
Manu Mulaveesala: And then there… was there a Q&A section?

235
00:57:26.230 --> 00:57:30.819
Anne Fernandez: You know, there… I… yeah, I wasn't sure if you had answered that before. There was.

236
00:57:30.820 --> 00:57:54.639
Manu Mulaveesala: Oh, there's a couple here. Okay, I had a couple questions here. Does MCP… we already talked about MCP with security, you know, Oracle Database, one user can read others and write. Yes, you can manage levels of authorizations, you totally can do that. You can provide read access to data, and typically with MCP, actually, and data, you're generally only providing read access, you know, providing write access is another thing, so…

237
00:57:54.640 --> 00:58:13.500
Manu Mulaveesala: The way that MCP works is that it has 3 core primitives that are called re… I didn't get too technical, but they're called resources, prompts, and what's the third one? Gosh, it's, like, slipping my mind right now. Resources, for data, and prompts for, for wiring your prompts.

238
00:58:13.500 --> 00:58:15.899
Manu Mulaveesala: And your workflows.

239
00:58:16.130 --> 00:58:39.519
Manu Mulaveesala: Gosh, it's… it's giving me right now, it'll come back to me, I'm just gonna need to pause for a second and it'll come back to me. If you search the three MCP primitives, you'll be able to find it. But if I ask the same question or prompt, can the MCP server, help with ensuring the same eventual response? I'm not sure I understand the question there. Any thoughts on recent talk about MCP filling up the context window? Yes, Robert!

240
00:58:39.520 --> 00:58:44.169
Manu Mulaveesala: That is one of the core limitations of MCP. That's why I was mentioning you shouldn't necessarily have

241
00:58:44.170 --> 00:59:08.920
Manu Mulaveesala: too many different MCP servers, because you might get a lot of data pulled in. That's probably what I would say one of the core, potential limitations of MCP. When you're first starting out, it might not necessarily be a big problem, and this really happens more when you're kind of, let's say, writing code, or you're using, you know, heavy use of agentic flows, and it depends on what your MCP is and how much data it pulls. It can be what's called context-heavy. What does that mean?

242
00:59:08.920 --> 00:59:32.809
Manu Mulaveesala: that means that your AI gets filled up with a bunch of memory or ideas and thoughts. I know that some people also were interested in comparing MCP to RAG. It's a little bit different, because RAG is kind of about training some of your tools, and we do teach RAG, but actually recently, the really popular thing has been, Agentic Search, and has kind of… it's kind of like RAG 2.0.

243
00:59:32.810 --> 00:59:57.410
Manu Mulaveesala: Where a lot of the AIs, like Claude and other things, are basically just using an agent to write bash scripts, and iterating and looping on writing the appropriate bash scripts, so that's probably a thing that we'll talk about and have. But yeah, MCP has file system access, and so actually, the newest way that people are doing things is that they're using the MCP for file system access.

244
00:59:57.410 --> 01:00:22.389
Manu Mulaveesala: and then having the agent, for example, iterate itself to write the bask scripts to go grep, or go search over their files, or find the right keywords and different things. For beginners, what would you recommend about learning about MCP and IDEs like Cursor or Claude? Well, take a course with us, and if you don't take… and if not taking a course with us.

245
01:00:22.390 --> 01:00:47.309
Manu Mulaveesala: There's a lot of great YouTube videos out there. The challenge I have with YouTube videos is a lot of them repeat the same information, so you don't necessarily get to build out a depth. It can get you a little bit of the way and an entry point, but it doesn't necessarily take you through intermediate and advanced, nor does it… a lot of YouTube is filled with content creators. There's one that I like that is really nice. There's an AI engineer.

246
01:00:47.310 --> 01:00:57.470
Manu Mulaveesala: YouTube channel with conferences, and they have some good stuff, out there, that is a little bit more advanced and is also enterprise-focused. But,

247
01:00:58.510 --> 01:01:04.030
Manu Mulaveesala: Awesome. I think that we hit most of the questions.

248
01:01:04.030 --> 01:01:11.380
Anne Fernandez: Yeah, and I just put a prerequisite to the MCP class in the chat, so if you're a beginner, you may want to

249
01:01:11.690 --> 01:01:12.710
Anne Fernandez: Start there.

250
01:01:13.360 --> 01:01:38.300
Manu Mulaveesala: Yeah, I think if you're a non-technical person, the best recommendation I would give you is think about what data sources you need to connect to. What are the pieces, you know, for different workflows, what are the data that people use, or the systems, or the platforms that people are using? That's the core thing to really answer for MCP, and sometimes the harder thing to think through, because we're just so used to already using it. And for the technical folks, there's a lot…

251
01:01:38.300 --> 01:02:01.139
Manu Mulaveesala: related to how to build your own MCP servers, and learning the stuff in Python, et cetera, and we do teach that in our class, but the first thing to see is, you know, what are you capable of doing with the AI, and learn a little bit more also about what connection is, and try testing it out with a tool like GitHub Copilot, or whatever your company has access to. A lot of companies have access to some tool.

252
01:02:01.140 --> 01:02:24.169
Manu Mulaveesala: More or less, or try it out with your… with a personal account, just to get a sense of it and try it out without using your work environment. That's another way that you can experiment, right? For example, I used it to connect to my Gmail and calendar. I'm not saying… maybe I… maybe you have to be careful of saying that and clarify that you should, you know, if… check your company policy and governance policy first, before wiring up any MCPs.

253
01:02:24.170 --> 01:02:35.690
Manu Mulaveesala: If it's not always allowed, which can happen in some companies, try it out with, with researching what personal tools you can use in order to just learn about the space and stay abreast of what's going on.

254
01:02:35.710 --> 01:02:46.480
Manu Mulaveesala: So, yeah, with that, I think that we're ready to close it out. Thanks so much, everyone. Super great session, one of the most engaged, interactive sessions I've had. It was a lot of fun.

255
01:02:46.680 --> 01:02:58.720
Anne Fernandez: Yeah, it was a lot of fun. Thank you all for joining and for bringing your A-game questions, and thank you, Manu, for answering them all so expertly. We certainly appreciate it and your time. Alright, we'll have to do another one of these very soon. We'll talk.

256
01:02:58.890 --> 01:03:05.529
Anne Fernandez: Okay. Awesome. Alright, thank you so much, everyone. Have a wonderful rest of your day, and hopefully we'll see you next time. Bye-bye.

257
01:03:05.530 --> 01:03:07.639
Manu Mulaveesala: Awesome, thanks, take care, everyone. Bye!