1
00:00:00,120 --> 00:00:01,720
So there's one more thing to mention.

2
00:00:01,720 --> 00:00:04,720
One more way in which it's just slightly more complicated.

3
00:00:05,240 --> 00:00:13,240
And that is that Landgraf itself is, in fact three different things, as I alluded to this already.

4
00:00:13,240 --> 00:00:20,560
But Landgraf is in fact Landgraf itself, which is also thought of as the Landgraf framework.

5
00:00:21,320 --> 00:00:28,200
It's a user interface tool called Landgraf Studio, which is one of these kinds of ways that you can

6
00:00:28,240 --> 00:00:29,440
hook things up visually.

7
00:00:29,440 --> 00:00:31,520
So like a visual builder.

8
00:00:32,160 --> 00:00:37,600
And it's something called Landgraf platform, which is the thing that they're sort of promoting on their

9
00:00:37,600 --> 00:00:46,280
website, as if that is Landgraf, uh, Landgraf platform, which is their, um, hosted solution for

10
00:00:46,280 --> 00:00:50,160
deploying and running your agents at scale.

11
00:00:50,680 --> 00:00:58,120
So, Landgraf, which is the framework which is analogous to Cray's framework, Landgraf studio, I

12
00:00:58,120 --> 00:01:00,560
think in Korea it was also called Korea Studio, right?

13
00:01:00,880 --> 00:01:01,440
I think so.

14
00:01:01,680 --> 00:01:06,600
And then Landgraf platform is analogous to CRI enterprise.

15
00:01:06,960 --> 00:01:09,550
And so these are the three offerings.

16
00:01:09,750 --> 00:01:14,590
And I mean, I think to me, I would make the same point that I made with crew, which is that what

17
00:01:14,750 --> 00:01:21,070
what perhaps part of what's going on here is, of course, Landgraf is looking for ways to commercialize

18
00:01:21,070 --> 00:01:22,910
and to monetize their offerings.

19
00:01:22,950 --> 00:01:30,270
And so looking for the enterprise play, the Landgraf platform is about deploying and running your Landgraf

20
00:01:30,270 --> 00:01:35,950
your graphs in their environment, uh, taking advantage of the fact that they've they've got hooks

21
00:01:35,950 --> 00:01:37,470
into everything that they've built.

22
00:01:37,470 --> 00:01:42,390
So I'm sure that if you have built all of your software using Landgraf, then it will be very convenient

23
00:01:42,390 --> 00:01:43,990
to use Landgraf platform.

24
00:01:43,990 --> 00:01:48,910
But I suspect that's what's going on, and I imagine that's why it is sort of promoted and framed so

25
00:01:48,910 --> 00:01:52,910
heavily on their site, as if Landgraf and Landgraf platform are the same thing.

26
00:01:53,030 --> 00:01:59,070
Uh, it's probably because this is the core commercial idea, uh, for for Lange chain and building

27
00:01:59,070 --> 00:01:59,630
this out.

28
00:01:59,950 --> 00:02:05,910
But we are going to be focused, of course, uh, as you can imagine on Landgraf, the framework, that

29
00:02:05,910 --> 00:02:07,550
part of it, that's what we're building to.

30
00:02:07,590 --> 00:02:09,070
We're going to be building our own.

31
00:02:09,270 --> 00:02:13,030
And we'll also be using Lange Smith as well to see what's going on.

32
00:02:13,150 --> 00:02:15,310
Uh, that's going to be interesting for us too.

33
00:02:15,350 --> 00:02:21,250
Now, at this point, I also wanted to do something I wanted to show you the blog post about building

34
00:02:21,290 --> 00:02:26,250
effective agents on Anthropic's website that I mentioned way back in week one, when we were looking

35
00:02:26,250 --> 00:02:27,330
at design patterns.

36
00:02:27,330 --> 00:02:33,010
But I think it's interesting and relevant now because it does show Anthropic's sort of positioning on

37
00:02:33,010 --> 00:02:37,090
this, and it's useful for you to keep this in mind when you think about the difference between what

38
00:02:37,090 --> 00:02:41,010
we're going to do today and what we're going to be doing in week six.

39
00:02:41,250 --> 00:02:48,570
So if we switch to building effective agents, then this this is the blog post you'll find on their

40
00:02:48,570 --> 00:02:49,250
website.

41
00:02:49,250 --> 00:02:51,050
And it's a brilliantly written post.

42
00:02:51,050 --> 00:02:56,850
I've got to tell you, it's really clear it contains within it the different design patterns that I

43
00:02:56,850 --> 00:02:57,290
mentioned.

44
00:02:57,330 --> 00:03:03,130
I hope I think I mentioned that I took the anthropic design patterns right from here.

45
00:03:03,170 --> 00:03:05,890
I thought they're so clear, so, so well explained.

46
00:03:06,130 --> 00:03:13,170
Um, and in addition to that, it has some discussion about using abstraction layers.

47
00:03:13,170 --> 00:03:15,210
And that's really what I wanted to show you here.

48
00:03:15,490 --> 00:03:22,250
So the, the, uh, um, it's says when and how to use frameworks.

49
00:03:22,250 --> 00:03:28,320
So they say there are many frameworks that make agentic systems easier to implement, including Landgraaf

50
00:03:28,320 --> 00:03:29,720
from Lange chain.

51
00:03:29,880 --> 00:03:35,440
They mention a few others that are not actually so popular, and not as far as I know, although I know

52
00:03:35,480 --> 00:03:36,320
of vellum the business.

53
00:03:36,320 --> 00:03:44,080
But but I don't think that that quite is as big as things like Cru and Autogen and now OpenAI agents

54
00:03:44,080 --> 00:03:44,600
SDK.

55
00:03:45,840 --> 00:03:48,000
But anyway, this is the point I wanted to make.

56
00:03:48,040 --> 00:03:53,160
These frameworks make it easy to get started by simplifying standard, low level tasks like calling

57
00:03:53,200 --> 00:03:57,960
llms, defining and parsing tools, and chaining calls together.

58
00:03:58,000 --> 00:04:03,640
However, they often create extra layers of abstraction and that can obscure the underlying prompts

59
00:04:03,640 --> 00:04:05,880
and responses, making them harder to debug.

60
00:04:06,280 --> 00:04:11,720
They can also make it tempting to add complexity when a simpler setup would suffice.

61
00:04:12,360 --> 00:04:14,320
You can see what Anthropic's getting at.

62
00:04:14,320 --> 00:04:17,400
You can imagine from their point of view, they've got an API.

63
00:04:17,720 --> 00:04:20,320
They've got something that is relatively simple.

64
00:04:20,360 --> 00:04:22,880
Memory can be handled with JSON objects.

65
00:04:23,280 --> 00:04:26,280
Llms can be hooked up simply by calling multiple times.

66
00:04:26,680 --> 00:04:33,600
And so for them, this this idea of building all of this abstraction around it, taking you further

67
00:04:33,660 --> 00:04:40,380
from actually working with the LM itself isn't necessarily something that resonates strongly.

68
00:04:41,140 --> 00:04:46,780
And so they conclude with we suggest that developers start by using LM APIs directly.

69
00:04:46,940 --> 00:04:49,660
Many patterns can be implemented in a few lines of code.

70
00:04:49,820 --> 00:04:53,900
If you do use a framework, ensure you understand the underlying code.

71
00:04:54,060 --> 00:04:58,220
Incorrect assumptions about what's under the hood are a common source of customer error.

72
00:04:58,460 --> 00:05:02,220
And by customer they mean us customers of anthropic.

73
00:05:02,900 --> 00:05:03,700
So there you go.

74
00:05:03,740 --> 00:05:05,140
I found this really interesting.

75
00:05:05,140 --> 00:05:08,780
It's obviously very clearly written and something that I'm very passionate about.

76
00:05:08,780 --> 00:05:10,940
It is an alternative, a different school of thought.

77
00:05:11,060 --> 00:05:20,460
It is, of course, somewhat, uh, not not in sync with the graph, uh, philosophy thesis of building

78
00:05:20,460 --> 00:05:21,660
this kind of structure.

79
00:05:21,780 --> 00:05:23,260
Uh, so it's interesting to keep that in mind.

80
00:05:23,260 --> 00:05:28,060
And of course, in week six, we're going to see what anthropic brings to the table in the form of MCP,

81
00:05:28,220 --> 00:05:33,460
which is a different way of thinking about it, a protocol, uh, for, for connecting things rather

82
00:05:33,460 --> 00:05:35,780
than building the actual glue itself.

83
00:05:35,820 --> 00:05:40,940
So I wanted to highlight that and give you that perspective, but it's not going to deter us from getting

84
00:05:40,940 --> 00:05:43,100
deep into Landgraaf right now.