WEBVTT

00:00.230 --> 00:03.710
This is just a quick video to let you know that we've added in some additional resources.

00:03.710 --> 00:09.350
If you're interested in learning about different types of cognitive agent architectures, we've covered

00:09.350 --> 00:15.170
the simple mixture of experts aggregation model, but you might be interested in different types of

00:15.170 --> 00:16.310
agent workflows.

00:16.310 --> 00:21.590
One that's quite interesting, for example, is hierarchical agent teams where you have different levels

00:21.590 --> 00:22.340
of agents.

00:22.340 --> 00:27.860
So you might have a research team, you know, a document offering team, a supervisor that supervises

00:27.860 --> 00:31.160
everything, and then various different types of agents.

00:31.160 --> 00:31.850
below.

00:31.880 --> 00:37.850
You've also got things like the plan and execute model, and this has been typically shown for different

00:37.850 --> 00:43.580
types of ones such as baby AGI, and also is very similar to a react style agent, where you can think

00:43.580 --> 00:48.950
of it as one step at a time, where you have a, you know, a request you plan, you generate some tasks

00:48.950 --> 00:54.230
that goes to a single task agent that loops to solve that task, updates those tasks, then maybe have

00:54.230 --> 00:55.250
some replanning.

00:55.250 --> 00:57.530
The response gets sent back to the user.

00:57.530 --> 01:00.690
And so this is basically quite a useful one.

01:00.690 --> 01:07.290
This you know, this is often been the case using this inside of things like code generation and also

01:07.290 --> 01:08.730
advanced task planning.

01:08.730 --> 01:12.690
You've also got more simple architectural patterns that are quite useful to know about.

01:12.690 --> 01:19.620
So for example reflection is a really easy way of basically asking the large language model to produce

01:19.620 --> 01:25.500
an output, and then after it's produced an output to reflect on what is the quality of that output.

01:25.500 --> 01:28.710
And you can see here we have this user request which goes in and generates.

01:28.710 --> 01:31.020
And then it will make an initial response.

01:31.020 --> 01:36.000
We reflect on that initial response with some sort of critique or evaluation criteria.

01:36.000 --> 01:42.360
And then we will then use that critique to then further generate and change the underlying asset.

01:42.360 --> 01:43.560
And this can be quite useful.

01:43.560 --> 01:48.060
For instance, if you're making like a blog post or a social media post, you might want to ask the

01:48.060 --> 01:54.320
LLM to critique itself or self reflect and say, you know, have you covered these things and making

01:54.320 --> 01:56.840
sure that the LLM always incorporates them?

01:56.840 --> 02:01.610
The one thing you want to note here is the number of times you do reflect will obviously increase the

02:01.610 --> 02:02.090
cost.

02:02.090 --> 02:07.610
So don't tell it to like reflect 100 times, but maybe like 1 or 2 times could potentially improve the

02:07.610 --> 02:09.140
outputs of your LMS.

02:09.350 --> 02:12.560
So we've left all these notebooks there inside of the GitHub repository.

02:12.560 --> 02:15.500
And we'll also give you a link on this lesson.

02:15.500 --> 02:18.740
So you can easily dive in and look at these notebooks in further detail.