WEBVTT 00:00.120 --> 00:01.760 And welcome to day four. 00:02.040 --> 00:05.360 On this day we are going to look at tools and autonomy. 00:05.480 --> 00:09.720 But before we get there, I want to talk about Agentic AI frameworks. 00:09.720 --> 00:11.840 Maybe something that's front of mind for you. 00:12.040 --> 00:17.640 There are a lot of these frameworks to pick from, and these frameworks are designed to give you kind 00:17.640 --> 00:24.960 of glue code or abstraction code that takes away some of the detail of interacting with Llms and gives 00:24.960 --> 00:30.440 you a nice, elegant framework for building agentic solutions and focusing on the business problem that 00:30.440 --> 00:31.200 you're solving. 00:31.240 --> 00:35.440 And there are a lot of them and new ones come up all the time, so it's quite hard to stay on top of 00:35.440 --> 00:36.080 everything. 00:36.080 --> 00:41.120 But they are happening and I want to just quickly orient you, tell you about the landscape and show 00:41.120 --> 00:44.720 you how the ones we're going to tackle on this course kind of fit into the bigger picture. 00:45.240 --> 00:49.920 So it's worth pointing out that there's sort of different levels of complexity of these frameworks, 00:49.920 --> 00:51.480 and they have pros and cons. 00:51.640 --> 00:59.040 And perhaps the simplest, the bottom layer of the complexity hierarchy, the bottom is actually to 00:59.080 --> 01:00.800 have no framework at all. 01:00.840 --> 01:03.080 Don't use an AI framework. 01:03.080 --> 01:09.510 And the kind of abstractions that come with it Simply connect to LMS directly using the APIs, just 01:09.510 --> 01:13.590 like we did in the last lab, and use that to orchestrate amongst LMS. 01:13.790 --> 01:16.990 And it's probably no surprise to you to learn that that's what we'll be doing this week. 01:16.990 --> 01:18.830 We're not going to be using any genetic framework. 01:18.830 --> 01:21.910 We'll be connecting direct, in fact, anthropic. 01:21.910 --> 01:27.590 In that blog post I mentioned called Building Effective agents make a very compelling case for just 01:27.590 --> 01:31.230 always using no framework connect directly to LMS. 01:31.350 --> 01:37.150 The APIs are relatively simple and straightforward, and the benefit is you get to see exactly what's 01:37.150 --> 01:38.190 going on under the hood. 01:38.190 --> 01:41.710 You control the prompts in detail, and so it's very compelling. 01:41.710 --> 01:44.150 And you'll see this week will be quite successful with it. 01:44.670 --> 01:51.790 Now, alongside having no framework, I've put something called MCP and that stands for the Model Context 01:51.790 --> 01:52.710 Protocol. 01:52.710 --> 01:57.590 And it's something created by anthropic who are believers in no frameworks. 01:57.590 --> 02:00.630 And it's not a framework, it is a protocol. 02:00.630 --> 02:03.590 So it's a way that things can connect together. 02:03.590 --> 02:05.070 It's open source. 02:05.110 --> 02:12.370 And the idea is it allows models to be connected to sources of data and tools in a way that's sort of 02:12.410 --> 02:13.170 open source. 02:13.210 --> 02:14.810 Agreed, established. 02:14.810 --> 02:17.050 So you don't need to use any glue code. 02:17.130 --> 02:19.290 As long as you conform to this protocol. 02:19.330 --> 02:25.130 You can stitch together models and their providers in this very elegant, simple way. 02:25.250 --> 02:29.650 And so for that reason, because although there is a bit of code associated with it, it's really more 02:29.650 --> 02:31.010 about having a protocol. 02:31.130 --> 02:34.130 I've grouped it along with having no framework at all. 02:34.170 --> 02:37.050 And something tells me that anthropic would like to be grouped that way. 02:38.090 --> 02:42.370 So on the next level up in terms of complexity, come these two frameworks. 02:42.530 --> 02:45.290 The first of them, OpenAI agents SDK. 02:45.690 --> 02:46.610 I love it. 02:46.650 --> 02:48.370 It's one of my very favorites. 02:48.570 --> 02:52.730 It's super lightweight and simple and clean and flexible. 02:53.050 --> 02:55.330 And I just really enjoy working with it. 02:55.330 --> 02:58.010 We're going to be using that next week and I'm looking forward to it. 02:58.210 --> 03:00.210 And it's relatively new. 03:00.250 --> 03:05.010 In fact, it's so new that when I was building some of the projects, the API sort of changed. 03:05.010 --> 03:10.210 They released a version like an hour after I'd made something that broke it. 03:10.250 --> 03:12.130 I was I was really caught off guard by that. 03:12.130 --> 03:17.400 So it is very new, but it's really great and I can't wait to show it to you and crew. 03:17.440 --> 03:19.000 I is also one of my favorites. 03:19.000 --> 03:20.080 I really love crew. 03:20.120 --> 03:21.480 It's been around for longer. 03:21.480 --> 03:22.760 It's very easy to use. 03:22.760 --> 03:24.320 It's also quite lightweight. 03:24.320 --> 03:30.200 One difference is that it has a kind of low code angle to it that you can do a lot of putting together 03:30.240 --> 03:35.440 agents to work on a problem through only configuration or through mostly configuration through YAML 03:35.440 --> 03:36.080 files. 03:36.080 --> 03:38.920 So you'll see that that's, that's a little bit more in that direction. 03:38.920 --> 03:42.840 And there is it's a bit heavier weight than, than OpenAI agents SDK. 03:43.480 --> 03:49.680 And then on top of these not built on top of them, but but a sort of next level of complexity, the 03:49.680 --> 03:54.400 top level of complexity for what I want to talk about, these two landgraaf from the people that brought 03:54.400 --> 03:57.440 you Lange chain and Autogen from Microsoft. 03:57.440 --> 04:02.000 And Autogen, as you'll discover, is really a couple of different things, but these two are relatively 04:02.000 --> 04:04.280 heavyweight compared to the others. 04:04.280 --> 04:09.800 They both have a steeper learning curve, particularly Landgraaf, which is quite complex. 04:09.920 --> 04:13.480 And of course, with this kind of extra complexity comes great power. 04:13.760 --> 04:18.480 It's really the idea of of landgraaf that you're building a kind of computational graph out of your, 04:18.520 --> 04:20.110 your agents and their tools. 04:20.110 --> 04:25.030 It's very powerful and it means you can build quite sophisticated things, but that also comes at a 04:25.030 --> 04:31.510 cost in terms of there being quite a heavy learning curve and you're sort of signing up for the ecosystem. 04:31.550 --> 04:37.430 You're signing up for a lot of terminology and concepts and abstractions which you need to buy into. 04:37.670 --> 04:42.030 And so it really sort of that ecosystem takes over your project in a big way. 04:42.150 --> 04:47.910 It becomes much less of a sort of agent agentic AI project and more of a landgraaf project. 04:47.910 --> 04:48.870 That's the thing. 04:49.030 --> 04:54.590 And and I think that's how it really both Landgraaf and Autogen are different to OpenAI agents, SDK 04:54.630 --> 04:58.590 and crew, where even though you're using those frameworks, you still feel like you're just sort of 04:58.630 --> 04:59.950 interacting with Llms. 05:00.110 --> 05:04.590 Whereas for these two at the top, it's very much being part of that ecosystem. 05:05.390 --> 05:08.830 And so I just bring this up to really orient you for the next few weeks. 05:08.830 --> 05:10.630 These are the ones that we're going to cover. 05:10.670 --> 05:13.470 I want to mention that there are many, many more. 05:13.590 --> 05:16.870 I do think that this will give you a good representative understanding. 05:16.870 --> 05:18.750 I think these are also the most popular. 05:18.790 --> 05:22.950 There are a few others that are quite popular that if there's time, I might add in extras at some point 05:22.950 --> 05:27.770 so that you can see some of them at work, but generally speaking, there are a lot to pick from. 05:27.930 --> 05:31.250 And which one you pick depends on a few things. 05:31.250 --> 05:37.210 It depends on the use case, because different types of platform will fit better for different business 05:37.210 --> 05:40.730 objectives, and a lot of it comes down to personal preference. 05:40.810 --> 05:45.850 There are some trade offs here in terms of how much you want to be using existing abstractions. 05:45.850 --> 05:49.490 I have to tell you that my bias is towards downwards. 05:49.490 --> 05:54.290 I like the the ones that stay out of your way and that are lightweight and simple and flexible. 05:54.290 --> 05:58.450 You can probably tell that what I was saying, but I do appreciate the power that you get from from 05:58.450 --> 05:58.690 the top. 05:58.690 --> 06:03.170 And I had a lot of fun with both Landgraf and Autogen on the projects, I have to say. 06:03.210 --> 06:05.090 So you know, I'm not, I'm not. 06:05.330 --> 06:06.970 I don't feel strongly either way. 06:06.970 --> 06:11.570 I definitely believe they have pros and cons, and I'm excited to show you all of them. 06:11.570 --> 06:17.130 I do think that in this course, we're going to get to see quite the spectrum of different frameworks, 06:17.130 --> 06:23.610 and it will really equip you well to be able to pick the one that works best for your work, for your 06:23.610 --> 06:28.930 skill sets, and for your team skill sets, and for the kinds of business problems that you're looking 06:28.930 --> 06:29.610 to tackle.