WEBVTT 00:00.400 --> 00:02.880 All right, so let's get stuck in. 00:02.880 --> 00:05.880 So before we get stuck in, let me set your expectations. 00:06.240 --> 00:07.600 There's going to be terminology. 00:07.600 --> 00:12.840 There's going to be some new concepts coming with Landgraf just as we've had in the past. 00:12.840 --> 00:19.080 And I realize that it's a bit jarring to say put all the crew stuff to one side, just just put that 00:19.080 --> 00:19.400 away. 00:19.440 --> 00:20.760 Compartmentalize. 00:20.760 --> 00:23.080 We're on to something new now, but it will be quick. 00:23.080 --> 00:27.680 We're going to get through the terminology quickly, and I'm going to be drilling it into you by repeating 00:27.680 --> 00:32.600 it several times, and then it's going to be second nature and you'll forget all about crew. 00:32.640 --> 00:33.400 All right. 00:33.520 --> 00:35.280 So what is the terminology. 00:35.320 --> 00:36.040 Here we go. 00:36.200 --> 00:43.680 So agent workflows are described in Landgraf speak as a graph. 00:43.840 --> 00:47.120 A graph is something that'll be very familiar to people in computer science. 00:47.120 --> 00:49.680 But a graph you can think of like a tree structure. 00:49.680 --> 00:54.320 It's something which has, uh, I'm about to say the names of the things I'm about to describe, but 00:54.320 --> 01:00.280 it has things connected together in, in something which looks a bit like a tree and some sort of a 01:00.280 --> 01:04.560 hierarchy where one thing depends on others beneath it. 01:05.040 --> 01:13.040 And so this, this idea of representing workflows in this kind of graph way, that is the the core idea 01:13.040 --> 01:15.520 of land graph as the name gives away. 01:16.440 --> 01:18.800 And then a state. 01:18.800 --> 01:26.440 So the state is something that represents the current snapshot, the current state of affairs, the 01:26.440 --> 01:28.520 status of your whole application. 01:28.640 --> 01:33.680 It's an object that that needs to encapsulate the state of the world. 01:33.920 --> 01:39.720 And that object is something which, which, uh, is is shared across the whole application. 01:39.720 --> 01:42.080 And it's something very fundamental to land graph. 01:42.080 --> 01:44.200 So you need to remember state. 01:44.200 --> 01:45.720 We're going to be using state a lot. 01:45.760 --> 01:48.760 And that's something which is it is uh it's a variable. 01:48.760 --> 01:51.520 It's like um it's, it's a, it's information. 01:51.760 --> 01:53.240 Um, it's not not a function. 01:53.240 --> 01:54.440 It's information. 01:54.480 --> 01:56.560 Nodes are functions. 01:56.760 --> 02:00.080 So a node which when people talk about graphs, they talk about nodes. 02:00.080 --> 02:01.720 They are the points on the graph. 02:01.720 --> 02:05.800 They are the, the, the things that are going to get connected together. 02:06.120 --> 02:09.240 The nodes are actually representing a function. 02:09.440 --> 02:12.360 Every node is a Python function. 02:12.360 --> 02:14.200 And that can be confusing the first time you hear it. 02:14.200 --> 02:19.360 Because when you think about graphs, you sometimes think of nodes as being things, as being data connected 02:19.360 --> 02:20.080 in some way. 02:20.080 --> 02:22.880 But no nodes in a graph are functions. 02:22.880 --> 02:29.360 They are Python functions, and they represent a piece of logic, a piece of ancient logic, a thing, 02:29.400 --> 02:30.960 an operation of some sort. 02:31.520 --> 02:35.040 They receive the current state as an input. 02:35.400 --> 02:36.800 They do something. 02:37.000 --> 02:40.080 They do something that might, might involve calling an LLM. 02:40.120 --> 02:42.320 It might involve some sort of a side effect. 02:42.320 --> 02:47.680 Changing something, writing something to a file, doing something that affects the world. 02:48.040 --> 02:52.440 And then they return an update, an updated state. 02:52.440 --> 02:54.760 That's that's what comes afterwards. 02:55.040 --> 02:59.400 So they take in a state, they do something and they return a new state. 02:59.680 --> 03:04.120 And I should say that the state you should think of as being like an immutable thing, the state object 03:04.120 --> 03:04.680 itself. 03:04.720 --> 03:05.080 You don't. 03:05.080 --> 03:06.040 You don't change it. 03:06.080 --> 03:13.760 You receive a state and then you return a new state, which is a difference from the state you had before. 03:14.840 --> 03:16.000 So far, you're with me. 03:16.280 --> 03:18.880 State and notes that the the two words. 03:19.200 --> 03:20.280 And then edges. 03:20.320 --> 03:24.640 And anyone that's ever looked at graphs will be familiar with the term edges. 03:24.640 --> 03:25.600 The edges, of course. 03:25.600 --> 03:26.360 The lines. 03:26.360 --> 03:30.520 The words I was trying not to say earlier is that graphs consist of nodes and edges. 03:30.720 --> 03:38.040 Edges are the connections between the nodes, and edges in the graph are also Python functions. 03:38.040 --> 03:45.840 They are Python functions that determine what node should be executed next based on the state, so the 03:45.840 --> 03:48.640 state can tell it what happens next. 03:48.680 --> 03:52.320 Now there are simple edge connections, which just means that is the next thing to happen. 03:52.320 --> 03:59.680 So if you have one node that is a function, the Python function that uh, that prints hello to to the 03:59.680 --> 04:00.200 output. 04:00.200 --> 04:02.800 And you have another node that prints goodbye. 04:02.800 --> 04:03.480 And they're connected. 04:03.480 --> 04:05.480 Then it will print hello, and then it will print goodbye. 04:05.480 --> 04:06.240 And that's it. 04:06.240 --> 04:11.600 But you can have, uh, connections which are conditional, which depend on something happening. 04:11.720 --> 04:17.320 And in that case, that would mean that one operation would happen and then subject to the condition, 04:17.320 --> 04:19.920 maybe the other one would happen, maybe not. 04:21.040 --> 04:23.200 So that is an edge. 04:23.560 --> 04:27.520 And then simply put, nodes are the things that do the work. 04:27.560 --> 04:30.400 They are the things that carry out tasks. 04:30.400 --> 04:35.960 They do stuff and edges are the things that determine, okay, that stuff was done, what's the next 04:35.960 --> 04:37.000 stuff that should be done? 04:37.200 --> 04:39.400 And that's all there is to it. 04:39.400 --> 04:41.840 That's the that's the main terminology. 04:41.840 --> 04:44.280 There are a couple of other things, but this is the core stuff. 04:44.440 --> 04:47.400 So if you've got this then you're most of the way there with Landgraf. 04:47.440 --> 04:50.520 And of course it would be nice to show this in some sort of a diagram. 04:50.560 --> 04:52.120 And this is the obvious diagram. 04:52.320 --> 04:55.120 So no surprise this is what it looks like. 04:55.120 --> 04:56.480 These are On nodes. 04:56.840 --> 05:03.760 Uh, the three orange circles there representing the three nodes, three operations happening in this 05:03.760 --> 05:05.280 graph, the edges. 05:05.280 --> 05:08.080 I'm trying to show there that one edge is not conditional. 05:08.080 --> 05:14.720 It just wants the top node is run, then the next one will run and it will have the state that's passed 05:14.720 --> 05:20.280 in will be the result of the, the the output from that first node. 05:21.200 --> 05:27.360 And the other edge is meant to show some sort of conditional edge that in some situations the node on 05:27.360 --> 05:28.840 the bottom right will run. 05:29.480 --> 05:33.400 So again, nodes do the work, edges choose what to do next. 05:34.560 --> 05:37.120 So that again is the terminology. 05:37.400 --> 05:40.520 So we are going to build our first graph. 05:40.520 --> 05:42.080 We're actually not going to do it today. 05:42.120 --> 05:43.280 Today is just a theory. 05:43.280 --> 05:46.040 Only one because I'm going to leave the practical for tomorrow. 05:46.080 --> 05:48.200 I'm going to let you brew over this. 05:48.200 --> 05:53.400 I'm going to let this sort of sink in so that the new terminology is something that's super clear to 05:53.400 --> 05:53.680 you. 05:53.680 --> 05:56.890 But the five steps to building your graph. 05:56.930 --> 05:58.130 Your first graph running it. 05:58.130 --> 06:00.170 And it's going to have some more terminology. 06:00.610 --> 06:06.250 First of all, you have to define your state class, the class which is going to store state. 06:06.290 --> 06:10.890 You don't you don't to create a state object because there'll be new objects created all the time. 06:11.210 --> 06:16.010 Uh, and, and so it will define the state class and you'll find out when we actually do this. 06:16.010 --> 06:19.970 There's something associated with that state class known as the reducer. 06:20.090 --> 06:22.530 And that's a very important concept. 06:22.530 --> 06:24.370 But I'm going to leave that concept to tomorrow. 06:25.370 --> 06:29.610 Secondly you start something called the graph builder. 06:29.650 --> 06:34.730 And the graph builder is a thing you're going to use to lay out all the nodes in your graph. 06:34.890 --> 06:36.690 This is something you're doing up front. 06:36.730 --> 06:38.090 Nothing's actually running. 06:38.090 --> 06:41.130 It's not like we actually have an agent system that's running. 06:41.130 --> 06:41.650 This is all. 06:41.690 --> 06:46.290 Before we run the agent system, we first start a graph builder. 06:47.050 --> 06:52.730 We then create a node that is one of these functions that's going to represent some operation that we're 06:52.730 --> 06:59.490 going to want to happen as part of this Landgraf and then we create some edges and we may repeat steps 06:59.490 --> 06:59.970 three and four. 06:59.970 --> 07:04.010 We may create lots of nodes and lots of edges to lay out the story. 07:04.050 --> 07:09.410 We're laying out the story of what we want our agent system to do, and this is all before it's actually 07:09.730 --> 07:11.610 live, before it's doing anything. 07:11.610 --> 07:14.490 It's not like we create nodes and edges and then things happen. 07:14.570 --> 07:20.170 No, we create nodes and edges as part of defining the graph, describing what this will do. 07:20.370 --> 07:25.010 Think of this as a bit like writing a program, but we're sort of doing it dynamically at runtime. 07:25.810 --> 07:30.850 So once we've done that, once at runtime, we've we've described the node, we've created the nodes, 07:30.850 --> 07:32.010 we've created the edges. 07:32.010 --> 07:35.010 We've laid out our whole agent workflow. 07:35.330 --> 07:39.130 We then run an operation called compiling the graph. 07:39.250 --> 07:42.850 And that will turn this into something that's ready to be executed. 07:42.890 --> 07:44.970 And then we kick it off and it runs. 07:45.970 --> 07:48.770 So it is important to get your head around this. 07:48.810 --> 07:51.010 What I've described there, this 12345. 07:51.050 --> 07:57.170 This is all stuff that that is running when we when we start our our system running, we do this process 07:57.170 --> 08:02.610 of defining the state class, starting the graph builder, creating nodes and edges, and compiling 08:02.610 --> 08:03.370 the graph. 08:03.650 --> 08:10.530 It's almost like part of running our system involves the sort of two phases, a phase when it's like 08:10.530 --> 08:16.210 defining itself, laying out the whole workflow, and then a second phase when that actually runs. 08:16.330 --> 08:19.890 And so that first phase, we're not really used to doing that when we code. 08:19.890 --> 08:25.170 We don't normally have a sort of meta phase when we're describing what it is that we want to do and 08:25.170 --> 08:26.410 then running it. 08:26.410 --> 08:31.250 But that's how it works with graph, and I hope that's made some sense. 08:31.250 --> 08:35.610 I'm going to explain it one more time when we go through this tomorrow. 08:35.610 --> 08:43.050 But this is the, uh, this is this is the overall approach that's used for building agentic workflows 08:43.050 --> 08:43.770 with graph. 08:43.770 --> 08:47.250 And with apologies for belaboring the point, let me say it one more time. 08:47.450 --> 08:53.050 When you run a graph application, when you kick it off, there's two phases. 08:53.050 --> 08:54.170 Two things happen. 08:54.170 --> 09:00.050 The first thing is that it runs some of your code that lays out the graph that defines what it is that 09:00.050 --> 09:01.930 you're trying to achieve with agents. 09:02.090 --> 09:04.930 And then once you've done that and that is these five steps here. 09:04.970 --> 09:07.690 Once you've done that, you then initiate it. 09:07.690 --> 09:09.690 You then run it, you invoke it. 09:09.690 --> 09:12.130 And then off this graph runs. 09:12.170 --> 09:16.250 And those two phases, both of them are part of running your application. 09:16.330 --> 09:20.370 And so what we're going to do next time is go through these five steps. 09:20.370 --> 09:21.890 Write code that will do all of this. 09:21.890 --> 09:23.170 And then call it. 09:23.250 --> 09:26.410 And you will then see the results all in one go. 09:26.970 --> 09:30.730 And if this if this is completely clear to you then sorry for saying it three times. 09:30.890 --> 09:32.410 Uh, hang on in there. 09:32.450 --> 09:38.090 And if this is not completely clicking for you, then then fear not when you see the actual code, I 09:38.090 --> 09:40.770 feel like it's going to come together and it's going to be easy peasy. 09:41.050 --> 09:44.090 And that indeed is what we're going to do tomorrow. 09:44.330 --> 09:47.050 So in day two we will do some actual code. 09:47.050 --> 09:48.010 We'll get to the lab. 09:48.010 --> 09:51.810 We'll build our first land graph agent I can't wait. 09:51.810 --> 09:52.730 I'll see you then.