WEBVTT 00:00.440 --> 00:02.520 Okay, enough with the sidebars. 00:02.520 --> 00:04.240 It's on to the main business. 00:04.280 --> 00:08.040 It is on to introducing OpenAI agents SDK. 00:08.080 --> 00:10.560 Now let me start by by teeing this up. 00:10.760 --> 00:16.120 So this is a framework which as I've already mentioned, is super lightweight, very flexible. 00:16.600 --> 00:21.960 One of the things I love about it is that it's not what they call opinionated. 00:22.000 --> 00:27.520 Opinionated is how some frameworks are described when they come with a lot of constructs, a way of 00:27.520 --> 00:32.160 doing things that you are then sort of expected to adhere to. 00:32.480 --> 00:37.080 Now, that's not always a bad thing because that can help build particular design patterns. 00:37.080 --> 00:40.320 It can get you it can allow you to build things really, really quickly. 00:40.600 --> 00:46.360 But my personal bias is I often prefer things that are less prescriptive and which give you more flexibility 00:46.360 --> 00:48.640 to choose how you want to to work with this. 00:48.680 --> 00:52.240 And very much OpenAI agents SDK is of that ilk. 00:53.040 --> 01:02.120 It also is very good at the same time at making some of the everyday stuff, like using tools much simpler. 01:02.240 --> 01:07.600 So if you think of all of that gunk that we had to do with the JSON around tools, all of that faffing 01:07.600 --> 01:09.960 around with big JSON objects. 01:09.960 --> 01:12.680 Well, OpenAI agents SDK just does all that for you. 01:12.680 --> 01:16.440 So that kind of stuff is very much it is in the way, but in a good way. 01:16.560 --> 01:22.480 It just does all of that JSON stuff and it handles the whole, if you remember the if statement and 01:22.480 --> 01:27.280 then my fancy version of it, it does all of that so that you just don't need to worry about it. 01:27.280 --> 01:33.080 And that's one way where it is good to have some abstraction, because that's the the useless boilerplate 01:33.080 --> 01:37.160 stuff that you have to repeat every single time that you want a framework to do for you. 01:37.600 --> 01:43.440 And as you can probably tell from the way I'm saying this, OpenAI agents SDK is my favorite. 01:43.720 --> 01:48.840 And yeah, I'm really looking forward to this week and I enjoy working with OpenAI agents SDK. 01:49.000 --> 01:51.320 And you might be thinking, editor, why? 01:51.520 --> 01:53.880 Why are you showing us your favorite one? 01:53.880 --> 01:56.480 First, why don't you want to end with your favorite one? 01:56.480 --> 01:57.800 You should work up to it. 01:57.840 --> 01:59.280 And I have two answers for you there. 01:59.320 --> 02:04.760 So first of all, I do want to point out that that whilst OpenAI agents SDK is my favorite, all of 02:04.760 --> 02:10.920 the frameworks are great, and they all have pros and cons, and each of them have use cases where they 02:10.960 --> 02:13.520 are the best for that particular kind of problem. 02:13.520 --> 02:15.520 So in some ways they're all my favorites. 02:15.520 --> 02:17.200 But but uh, no, that's not true. 02:17.240 --> 02:18.400 OpenAI, you're my favorite. 02:18.520 --> 02:20.480 Uh, they're all good. 02:20.480 --> 02:24.920 And so, uh, but they also get that during the course of our journey, they're going to get more and 02:24.920 --> 02:25.760 more complex. 02:25.760 --> 02:28.080 And so it's going to make sense to do it the way we do. 02:28.120 --> 02:30.480 And our projects are going to get beefier and beefier. 02:30.760 --> 02:31.960 So that's the first reason. 02:32.000 --> 02:37.600 The second reason is that whilst this is the framework that we are starting with, we are also going 02:37.600 --> 02:43.800 to finish with it as well, because in week six, when we work with MCP, which is a protocol, not 02:43.800 --> 02:49.600 really a framework, in the same way we are going to be using OpenAI agents SDK again in a more hardcore 02:49.640 --> 02:49.880 way. 02:49.920 --> 02:50.840 So it's going to come back. 02:50.840 --> 02:54.000 It's going to make a triumphant return in week six. 02:54.000 --> 02:55.720 So you you don't need to worry. 02:55.720 --> 02:57.120 The crescendo will be there. 02:58.040 --> 03:02.040 Now, one thing that all of the frameworks have in common is that they come with some terminology, 03:02.080 --> 03:05.160 some constructs, and each one is of course slightly different. 03:05.160 --> 03:07.840 And we will be learning each of the different terminologies as we go. 03:07.840 --> 03:11.310 But of course OpenAI's terminology is pretty minimal. 03:11.310 --> 03:13.030 It's fairly simple stuff. 03:13.350 --> 03:13.990 There is. 03:14.230 --> 03:17.390 There's three, three concepts, three terms I'm going to go through with you. 03:17.390 --> 03:19.550 The first of them is just an agent. 03:19.590 --> 03:27.030 Agents represent their sort of package around calls to llms, which have a particular role, a particular 03:27.030 --> 03:28.950 purpose in this solution. 03:28.990 --> 03:36.150 Obviously handoffs this is their name for the interactions between agents. 03:36.270 --> 03:41.790 So handoff is a term that they use a fair amount and we'll see it come up in different contexts. 03:41.790 --> 03:42.790 But that's what it means. 03:42.790 --> 03:44.150 It is an interaction. 03:44.630 --> 03:52.390 And guardrails is their terminology for the kinds of checks and controls that you put around, making 03:52.390 --> 03:56.830 sure that an agent is doing what you want it to do and isn't going off the rails. 03:57.070 --> 04:00.710 Guardrails is a pretty common word in normal software engineering anyway. 04:00.710 --> 04:06.990 But but it's one of the the three terms to take on board for using OpenAI agents SDK. 04:07.910 --> 04:13.390 So in order to actually run an agent, there are three steps that you have to take. 04:13.590 --> 04:14.390 And here they are. 04:14.430 --> 04:15.630 Here are the three steps. 04:15.830 --> 04:18.110 You create an instance of agent. 04:18.150 --> 04:20.030 That's going to be the thing that you're going to set. 04:20.070 --> 04:23.630 Up to be one of the roles in your solution. 04:24.150 --> 04:27.750 You use something called with trace to be. 04:27.790 --> 04:31.350 Able to keep a log of all of your interactions with that agent. 04:31.350 --> 04:33.790 This isn't required, but but it's going to be the way we will do it. 04:33.790 --> 04:35.030 And we will always do it this way. 04:35.030 --> 04:39.510 And it's going to allow us to look in OpenAI's monitoring tools and see everything that's going on. 04:39.910 --> 04:42.310 And then you call something called runner. 04:43.430 --> 04:49.590 And runner run is the thing that will actually run the agent as you will see. 04:49.790 --> 04:56.870 And runner run, the thing that actually runs the agent is of course an async function. 04:56.870 --> 04:58.750 Or I should say it's a coroutine. 04:58.910 --> 05:03.910 And so as a result, it's something that we will need to await in order to get it to actually run. 05:03.950 --> 05:07.870 So we are going to need to do something that's going to look like await runner run. 05:08.110 --> 05:10.230 And these are the three steps. 05:10.590 --> 05:14.670 And without further ado, let's go and do those three steps right now.