WEBVTT 00:00.160 --> 00:05.960 Well, if you're watching this video, then you weren't entirely put off by the grueling day one and 00:05.960 --> 00:07.080 you've come back for more. 00:07.120 --> 00:08.000 For day two. 00:08.040 --> 00:08.920 Welcome back. 00:09.360 --> 00:12.560 So this is going to be an interesting day. 00:12.560 --> 00:13.800 It's going to be different to day one. 00:13.800 --> 00:15.400 It's going to be shorter than day one. 00:15.400 --> 00:16.640 You'll be happy to hear. 00:16.800 --> 00:19.160 But also it's going to be pure theory. 00:19.160 --> 00:22.680 We're going to be talking about agents and agent architecture. 00:23.040 --> 00:27.640 Almost all of the days on this entire six weeks program are very practical, heavy. 00:27.640 --> 00:31.600 This is one of the rare days that is going to be a theory day, but it's going to be interesting. 00:31.600 --> 00:33.000 It's going to be really great. 00:33.000 --> 00:36.680 I'm going to tell you all about Agentic architecture. 00:36.880 --> 00:37.800 Let's get to it. 00:38.240 --> 00:44.000 So where better place to start than with the almost obligatory what is an agent? 00:44.040 --> 00:45.240 Kind of question. 00:45.240 --> 00:50.600 And the interesting thing is that there's not actually a great answer to that. 00:51.040 --> 00:55.840 Agent guy is one of these things that's been super hyped and generated a lot of excitement, and people 00:55.840 --> 00:57.920 are using the term to mean a lot of different things. 00:57.920 --> 01:03.690 And in fact, there was a meme going around last year about how agent Guy could mean almost anything. 01:04.090 --> 01:10.490 But having said that, typically there is one simple definition, and actually the definition that I'm 01:10.490 --> 01:12.210 pulling has come from hugging face. 01:12.250 --> 01:18.690 They're small agents project where they give a really nice, crisp, simple definition, and they say 01:18.970 --> 01:24.610 AI agents are programs where LLM outputs control the workflow. 01:25.250 --> 01:33.010 So basically one one output from LLM is able to decide what tasks are carried out in what sequence. 01:33.930 --> 01:36.890 So that is a good crisp definition. 01:36.890 --> 01:39.010 It's something for you to keep in mind. 01:39.050 --> 01:42.010 And it applies in some situations and not in others. 01:42.690 --> 01:49.250 More generally though, you will find that there are different hallmarks of what makes something an 01:49.250 --> 01:57.930 AI solution, and people will often use the term agentic AI or agents, if any one of these five hallmarks 01:57.930 --> 01:58.610 are met. 01:59.010 --> 02:05.460 So first of all, sometimes any solution that involves multiple LLM calls could be called Agentic I 02:05.860 --> 02:10.820 and, you know, a bit like the one that we just built in day one that could be called Agentic I. 02:11.620 --> 02:18.620 Now, this one LMS with the ability to use tools, that's often what people think of as the kind of 02:18.660 --> 02:20.940 litmus test for whether something is agentic. 02:20.940 --> 02:26.620 If you remember when we looked at the N810 idea and the very first, at the beginning of the first lecture, 02:26.620 --> 02:31.460 when I had it turn on the lights that I was wearing, that was an example of tool use. 02:31.540 --> 02:33.620 And we're going to be talking a lot more about tool use. 02:33.620 --> 02:37.100 So that's a that's obviously a hallmark quality as well. 02:37.740 --> 02:44.580 Another is when you have an environment set up which allows different LMS to send information to each 02:44.620 --> 02:47.940 other, a kind of coordination orchestration environment. 02:48.220 --> 02:56.780 That in itself for some people defines a gender I sometimes it's when you have a planner, some process 02:56.780 --> 03:03.820 that's able to coordinate activities and that that itself is an LM, that is and that's quite, quite 03:03.860 --> 03:09.310 similar to the text in blue that that for some people is the hallmark of agentic AI. 03:09.350 --> 03:16.070 But for some people, for a lot of people, there's this word autonomy, and that really captures the 03:16.070 --> 03:17.750 essence of agentic AI. 03:18.030 --> 03:26.590 And this word is saying that we are giving some ability to an LLM to control what order things happen 03:26.590 --> 03:30.790 in, or what happens to sort of choose its own adventure in some way. 03:30.950 --> 03:33.990 And giving it that autonomy, giving it, if you will. 03:33.990 --> 03:37.990 That agency is one of the ways to define agentic AI. 03:38.230 --> 03:40.190 And in some ways, that sounds kind of spooky. 03:40.230 --> 03:46.550 Like, like what does it mean to have an LLM, a statistical language model that is autonomous? 03:46.710 --> 03:50.270 And to be honest, it's just sort of fancy use of language in many ways. 03:50.270 --> 03:52.270 You could argue that exactly. 03:52.310 --> 03:58.350 The project that we just did, the challenge that I set you to have an LLM come up with a business sector 03:58.350 --> 04:00.750 that would then be analyzed for pain points. 04:00.750 --> 04:06.150 In some ways, it was showing autonomy there because it was able to choose what business sector to use. 04:06.310 --> 04:11.640 And even more straightforward was when we asked for a question, and then we had to answer the question. 04:11.640 --> 04:13.200 It had a level of autonomy. 04:13.200 --> 04:20.000 So you can see how any time when we're giving an LLM the opportunity to decide how we will carry out 04:20.000 --> 04:26.560 future actions, you could think of that as giving it autonomy and that being a way to describe Agentic 04:26.600 --> 04:27.000 AI. 04:27.600 --> 04:34.520 And just to go one level deeper, anthropic wrote this brilliant post that's called Building Effective 04:34.520 --> 04:40.680 Agents, and I will refer back to that many times because I think it's just so clear and well written. 04:40.800 --> 04:47.720 And in that that particular paper or that that blog post, they distinguish two types of what they call 04:47.760 --> 04:49.400 agentic system. 04:49.400 --> 04:51.040 So that's this is their terminology. 04:51.040 --> 04:55.400 But they would say that there are, there's this umbrella term agentic system. 04:55.400 --> 04:58.800 And under that umbrella term there are two different categories. 04:59.040 --> 05:01.720 And one category is called workflows. 05:02.000 --> 05:09.760 And these as it says, are systems where models and tools are orchestrated through predefined paths 05:10.200 --> 05:15.410 and they delineate between that and something called agents. 05:15.410 --> 05:22.210 And agents are where models dynamically direct their own processes and tools, maintaining the control 05:22.210 --> 05:25.050 over how tasks get accomplished. 05:25.410 --> 05:29.850 So these are the two different sort of subfields under a genetic system. 05:29.850 --> 05:32.530 There are workflows and there are agents. 05:32.530 --> 05:37.610 And you can see they're sort of hinting that perhaps the things that a lot of people call agents are 05:37.610 --> 05:39.370 in fact workflows. 05:39.370 --> 05:43.770 So they're drawing that distinction, but they do still describe the whole thing as being a genetic 05:43.770 --> 05:44.570 systems. 05:44.570 --> 05:46.130 So it's a bit of wordplay there. 05:46.170 --> 05:49.010 Does that mean that workflows are agentic or not? 05:49.410 --> 05:50.810 I guess I guess they are. 05:50.970 --> 05:53.850 But you can see there's some ambiguity in the terminology. 05:53.850 --> 05:58.850 But I do think this is a helpful framework for distinguishing between these two worlds. 05:58.850 --> 06:01.210 And we're now going to go deeper. 06:01.210 --> 06:06.050 We're going to look at each of these and look at some common design patterns. 06:06.050 --> 06:10.970 And this should hopefully start to bring things to life and show you about how we can apply this kind 06:10.970 --> 06:13.130 of thinking to actual projects.