WEBVTT 00:00.080 --> 00:03.680 Well, I'm so excited to see you back here for week four. 00:03.720 --> 00:04.680 Day one. 00:04.680 --> 00:06.240 It's already week four. 00:06.280 --> 00:06.800 Here we are. 00:06.840 --> 00:12.600 It is Landgraaf week as we really change things up, Landgraaf thinks about the universe differently. 00:12.640 --> 00:18.200 We're going to have a great old time, but it's going to require some different ways of thinking about 00:18.200 --> 00:18.560 things. 00:18.560 --> 00:19.440 But we're going to have fun. 00:19.680 --> 00:22.720 Uh, the project this week is particularly good. 00:22.720 --> 00:27.280 At the end of it, there's something really fun to show you with real business value that I've actually 00:27.320 --> 00:31.960 I've had business value from this already myself from from this project that we built. 00:32.120 --> 00:37.840 Um, but generally I want to say that this week, week four is actually quite a short week because I 00:37.840 --> 00:44.400 feel like we've got deep into OpenAI agents, SDK and crew, and we've done a lot that's covered familiar 00:44.440 --> 00:44.920 ground. 00:44.920 --> 00:50.040 And now the things that we do both actually with Landgraaf and with Autogen, it's going to have a lot 00:50.080 --> 00:50.480 in common. 00:50.480 --> 00:53.560 We're not going to need to go into quite the same amount of detail. 00:53.560 --> 00:59.080 So we'll be moving a bit more briskly through, but I'll be giving you plenty of the briefing and giving 00:59.080 --> 01:04.620 you the ability to go off and build your own projects with Landgraaf if it if it happens to tick the 01:04.620 --> 01:06.140 right boxes for you. 01:06.180 --> 01:07.620 All right, let's get into it. 01:07.620 --> 01:11.660 But before we even get into Landgraaf, I know what you're thinking. 01:11.900 --> 01:12.980 You're confused. 01:13.140 --> 01:18.460 You're confused about Langshan and Landgraaf and maybe also confused about Lange Smith, if you've heard 01:18.460 --> 01:18.900 of that. 01:18.940 --> 01:25.860 It is a trio of products offered by Lange Chain, and you may be unclear about how Landgraaf fits into 01:25.860 --> 01:26.100 it. 01:26.100 --> 01:28.420 And why aren't we going through Lange Chain? 01:28.420 --> 01:33.300 And it would be a great question, and it's one that I intend to clarify for you right now. 01:33.580 --> 01:39.900 Uh, and this this is the Lange chain ecosystem, and I could think of no better way of showing it than 01:39.940 --> 01:42.940 by taking little, little snapshots from their website. 01:43.260 --> 01:44.940 So Lange chain. 01:44.980 --> 01:46.460 Lange chain is where it began. 01:46.460 --> 01:48.940 It's been around for for many years now. 01:49.060 --> 01:54.500 And it was one of the earliest of the abstraction frameworks that that was there. 01:54.540 --> 02:02.580 And its initial kind of raison d'etre was that if you were building very bespoke integrations with different 02:02.670 --> 02:03.510 APIs. 02:03.710 --> 02:04.830 It was painful. 02:05.030 --> 02:07.910 And if you needed to change, say, from using GPT. 02:08.030 --> 02:11.030 To using Claude, you had to redo a lot of work. 02:11.030 --> 02:13.950 And so they had the idea of building abstractions. 02:14.270 --> 02:19.350 Then when it turned out that a lot of people were writing applications, which involved a call to an 02:19.350 --> 02:24.150 LLM, followed by another, followed by another, it sort of turned into this idea of chaining together 02:24.150 --> 02:24.990 your calls. 02:25.190 --> 02:30.190 Um, and, and, you know, Long Chain really took root and became something that's, that's quite advanced 02:30.190 --> 02:36.590 and supports things like rag uh, for people that do my engineering course, we use lang chain for a 02:36.630 --> 02:37.830 Rag implementation. 02:38.190 --> 02:44.870 Uh, it supports things like prompt templates, a sort of higher level construct built on top of, uh, 02:44.870 --> 02:52.230 prompting and supports memory in a very robust way, allowing you to to build memory that you keep in 02:52.230 --> 02:57.350 memory or that you keep in memory, keep it in Ram, or that you persist in a database. 02:57.590 --> 03:00.030 Uh, and it has various memory models. 03:00.230 --> 03:03.230 Uh, I guess not unlike the things that we saw from Crewe. 03:03.590 --> 03:09.090 Uh, but but there's a bit more, uh, stuff to the a few more abstractions and things to learn about. 03:09.130 --> 03:13.170 They also have their own declarative language, LCL as well. 03:13.290 --> 03:17.770 Uh, so there's, there's a, there's a lot of depth to Lang Chain, and it's really building a kind 03:17.770 --> 03:24.610 of engineering discipline around the, the, the art of working with llms and putting some scaffolding 03:24.610 --> 03:31.970 and some, some templates and some, uh, well, solidified code with things like good prompt practices 03:32.130 --> 03:34.330 around calling llms. 03:34.490 --> 03:38.530 And it's been extremely successful, uh, in that regard. 03:38.730 --> 03:42.810 It, it also it allows you to do things like abstract using tools. 03:42.810 --> 03:47.930 And so from that point of view, it does, in fact support building a genetic infrastructure. 03:47.930 --> 03:54.770 So you can use Lang chain and you can use lang chains workflows to build agent platforms. 03:55.090 --> 04:00.730 But it sort of predates the the recent explosion and excitement with, with agents. 04:00.730 --> 04:03.890 And so it's working in a more simplistic level. 04:03.970 --> 04:07.310 It's not their main agent platform offering. 04:07.350 --> 04:13.270 It's more of their glue code for building any application using LMS. 04:13.510 --> 04:18.790 Now, you probably heard me say that I have something of a love hate relationship with Long Chain. 04:18.910 --> 04:24.950 I definitely appreciate its power and the way that with very little code indeed, you can get up and 04:24.950 --> 04:31.630 running with with a lot of functionality like building a Rag pipeline in like four lines of code. 04:32.150 --> 04:36.230 Having said that, I do also see some drawbacks, and it's very similar to the drawbacks I was talking 04:36.230 --> 04:40.470 about with the more, more opinionated aspects of crew. 04:40.630 --> 04:47.110 It's that by signing up for a lot of the abstractions and a lot of the glue code that comes in the box 04:47.110 --> 04:51.870 with long chain, you're signing up for their way of doing things, and you have a bit less visibility 04:51.870 --> 04:54.390 into the actual prompts going on behind the scenes. 04:54.670 --> 04:59.870 And over time, the API's into LMS has become more and more similar. 04:59.910 --> 05:05.110 Anthropic is a little bit of an odd one out, but everybody else has really converged on OpenAI's endpoints 05:05.110 --> 05:06.360 and on their that structure. 05:06.360 --> 05:11.200 And so it's become extremely simple to interact directly with LMS. 05:11.480 --> 05:16.640 And handling memory is something that is also very simple to do yourself, because memory is really 05:16.640 --> 05:20.920 just the JSON blob of the conversations that you've had with the model. 05:20.920 --> 05:23.400 And so you can handle that JSON yourself. 05:23.400 --> 05:25.040 You can persist it as you want. 05:25.040 --> 05:27.280 You can combine memory in different ways. 05:27.320 --> 05:35.720 And so I see in some projects there's less need to sign up for a big ecosystem around, say persisting 05:35.720 --> 05:36.480 memory. 05:36.480 --> 05:38.960 But again there are pros and cons. 05:39.000 --> 05:42.240 There's definitely strong benefits to working with long chain. 05:42.240 --> 05:48.240 And with all of these significant engineering and problems that have already been solved that comes 05:48.240 --> 05:48.920 with it. 05:48.960 --> 05:52.080 Okay, so that's long chain and my mini rant. 05:52.080 --> 05:53.440 Thank you for putting up with that. 05:53.440 --> 05:56.560 Let's go on to talk about what is graph then. 05:56.760 --> 06:02.280 So on the website this is how it's positioned run at scale with graph platform. 06:02.440 --> 06:08.180 And as we'll talk about in a minute, Landgraf platform is actually one of the parts of Landgraaf, 06:08.180 --> 06:13.300 but Landgraaf itself is a bit bigger than that, so it's confusing on the website that they really push 06:13.340 --> 06:15.740 Landgraaf platform in this way. 06:15.740 --> 06:17.780 But let me tell you what I think Landgraaf is. 06:17.780 --> 06:24.740 Landgraaf is a separate offering from the company Lange chain from the same people it. 06:24.780 --> 06:27.540 It actually is independent from Lange chain. 06:27.540 --> 06:33.620 So whilst when you're working with Landgraaf, you can use Lange chain code to actually call LMS and 06:33.620 --> 06:37.300 to do various things with LMS you can do it's optional. 06:37.300 --> 06:42.940 You can really use any framework or you can just call LMS directly with Landgraaf. 06:43.340 --> 06:52.060 Landgraaf is all about a platform that focuses on stability, resiliency, and repeatability in worlds 06:52.060 --> 06:58.540 where you're solving problems that involve a lot of interconnected processes, like an agentic platform. 06:58.780 --> 07:05.140 So it's an abstraction layer that allows you to organize your thinking around a workflow of different 07:05.140 --> 07:07.460 activities that could have feedback loops. 07:07.460 --> 07:10.520 It could have times when humans need to get involved. 07:10.520 --> 07:16.120 It could have moments when you need to keep memory, and it allows you to organize all of that in a 07:16.160 --> 07:22.600 very repeatable and easily monitored and stable and scalable way. 07:22.920 --> 07:29.280 That's what Landgraf is, and the word graph gives some of it away that it's all built around graphs. 07:29.320 --> 07:33.760 Graphs being kind of tree structures of how to think about your workflow. 07:33.920 --> 07:40.640 So it imagines all workflows, anything that you might have going between agents in the form of a tree, 07:40.680 --> 07:46.720 a tree of nodes which are connected together, which represent different things that can happen at different 07:46.720 --> 07:49.000 points in your agentic workflow. 07:49.200 --> 07:55.360 And by thinking of it in this abstract way, and by putting sort of belts and braces around each point 07:55.360 --> 08:02.240 in this graph, they're able to bring stability and resiliency to a world that is a bit unpredictable 08:02.240 --> 08:06.400 and has has, you know, people have resiliency concerns about agentic AI. 08:06.760 --> 08:08.280 So that's really their approach. 08:08.280 --> 08:10.020 That's the problem they're trying to solve. 08:10.020 --> 08:16.580 And you can see if you if you read the detail there that they're saying you use this to design agent 08:16.580 --> 08:22.620 driven user experiences featuring things like human in the loop, multi-agent collaboration, conversation, 08:22.620 --> 08:28.180 history, memory, and what they call time travel, which is all about being able to checkpoint where 08:28.180 --> 08:33.740 you are in the process of being able to step backwards if you need to, to restore where you were as 08:33.740 --> 08:35.060 of at any point in time. 08:35.820 --> 08:42.100 A deploy with tolerant scalability, fault tolerant scalability, meaning that anything can go down 08:42.100 --> 08:43.220 and it will keep running. 08:43.220 --> 08:45.900 And that's a bit of the Landgraf platform thrown in there. 08:46.300 --> 08:48.660 So that's what Landgraf is all about. 08:48.700 --> 08:49.780 It's not related. 08:49.820 --> 08:52.740 It's not necessarily related to Lange chain. 08:52.820 --> 09:01.500 It is a framework for robustly running complex agent workflows, uh, giving you that kind of stability 09:01.500 --> 09:06.580 and monitoring, although I use the word monitoring there, and that was perhaps the wrong word to use 09:06.780 --> 09:07.100 it. 09:07.140 --> 09:12.350 It gives the ability to monitor, but it doesn't actually do the monitoring itself, because Lang Chain 09:12.350 --> 09:17.710 has a third product called Lang Smith, which is their kind of monitoring tooling. 09:17.870 --> 09:19.790 And Landgraf connects with Lang Smith. 09:19.790 --> 09:24.910 So you can use Lang Smith to monitor what's going on in your Lang graph graph. 09:25.110 --> 09:26.870 But Lang Smith is a separate offering. 09:26.870 --> 09:30.990 And Lang Smith can be used when working with lang chain or with Lang graph. 09:30.990 --> 09:32.990 And we will use Lang Smith. 09:32.990 --> 09:36.270 We will use that so that we can see things going on. 09:36.270 --> 09:42.470 And it gives you, as it says here, visibility into your calls and your reasoning to quickly debug 09:42.510 --> 09:43.350 failures. 09:44.190 --> 09:47.950 So that is how the different products line up. 09:48.310 --> 09:53.030 It is a bit confusing because you can use Lang chain to build agent workflows. 09:53.030 --> 09:58.750 It has an abstraction layer over things like tool calling, but Lang Graph is the core offering. 09:58.750 --> 10:05.110 That's the modern offering that's designed to meet the excitement of today's agent AI. 10:05.270 --> 10:11.590 And the particular thing that they're focused on is the kind of scaling in a resilient, robust, repeatable 10:11.590 --> 10:12.030 way.