WEBVTT 00:00.160 --> 00:04.640 Welcome back to cursor, where we are ready to go with OpenAI agents SDK. 00:04.960 --> 00:08.240 So you'll see that we have the folders on the left here, the directories. 00:08.240 --> 00:13.680 And we are now going into number two OpenAI and into lab number one. 00:13.840 --> 00:14.760 Here it is. 00:15.440 --> 00:18.080 First look at the OpenAI agents SDK. 00:18.440 --> 00:21.400 I've included a link to the docs for OpenAI. 00:21.520 --> 00:28.080 And I will say that it does almost concern me showing you these docs because they're so clearly written 00:28.080 --> 00:31.400 and so easy that I rather wonder what my job is in all this. 00:31.640 --> 00:34.920 I will say that for the next few weeks, I think I definitely have a job. 00:34.920 --> 00:39.080 It gets a lot harder, but for this one you'd be doing well just to read the docs and it'll be very 00:39.080 --> 00:39.760 clear to you. 00:39.760 --> 00:42.120 So anyway, I'll keep going regardless. 00:42.400 --> 00:44.480 So let's do some imports to get us started. 00:44.480 --> 00:49.360 So we will say from dot env import load dot env. 00:49.520 --> 00:50.680 We know we got to do that. 00:51.000 --> 00:55.320 And now we're going to to do some imports from OpenAI's package. 00:55.480 --> 00:57.760 And their package is called agents. 00:58.080 --> 01:00.840 And from that package we are going to import agent. 01:01.160 --> 01:03.240 We're going to import this thing called runner. 01:03.920 --> 01:06.060 And we're going to import this thing called trace. 01:06.260 --> 01:10.180 These are the three things I told you about that we were going to use today. 01:10.700 --> 01:13.380 And the OpenAI package is called agents. 01:13.380 --> 01:14.900 That's that's the one I do. 01:14.940 --> 01:20.460 I do sometimes think that it's a bit strange that they use such a generic name for it, because I imagine 01:20.460 --> 01:24.260 a lot of people use the name agents for, for their own packages. 01:24.420 --> 01:28.660 I often do, and so you have to then start messing around with, with different kinds of imports. 01:28.660 --> 01:29.620 But still they did. 01:29.660 --> 01:30.500 Its called agents. 01:30.500 --> 01:35.380 And this is, this is how we import these three key things agent runner and Trace. 01:35.900 --> 01:36.300 All right. 01:36.300 --> 01:37.860 So the usual starting point. 01:37.900 --> 01:39.980 Hopefully this is second nature to you. 01:40.020 --> 01:46.860 We need to run load dot env with override is true to bring in our environment variables they are in 01:47.300 --> 01:47.620 okay. 01:47.660 --> 01:49.500 We are now going to make an agent. 01:49.500 --> 01:53.620 We're going to create our first agent, our first agent from a framework of this course. 01:53.780 --> 02:01.740 And so we're going to have an agent and say agent equals it's going to be an instance of the class agent. 02:01.740 --> 02:05.700 So we're creating an instance of agent OpenAI's class. 02:05.700 --> 02:06.980 Now it has some parameters. 02:06.980 --> 02:08.580 We need to use some arguments. 02:08.580 --> 02:10.470 And one of them is name. 02:10.470 --> 02:13.430 We give it a name and we're going to call it the jokester. 02:14.070 --> 02:16.750 And then we give it something called instructions. 02:16.950 --> 02:18.830 I like the way this is just filled in. 02:18.830 --> 02:21.310 I'll just press tab and put me out of my misery. 02:21.430 --> 02:24.070 So we pass in some instructions. 02:24.150 --> 02:30.030 The instructions are basically the system prompt for this call to LM. 02:30.350 --> 02:36.430 So each of these agents you can think of as being an LM with one particular system prompt designed to 02:36.470 --> 02:38.830 organize it around one particular task. 02:39.070 --> 02:41.670 And so our system prompt called the instructions. 02:41.670 --> 02:44.030 In this case you are a joke teller. 02:44.030 --> 02:47.950 And you also pass in a model which in this case is GPT four mini. 02:48.190 --> 02:53.790 And you might be thinking in your mind a very good question, which is AI, does this mean that OpenAI 02:53.830 --> 02:57.870 agents SDK always uses OpenAI's models? 02:58.070 --> 03:02.350 You might, you know, that might sound like it's a sensible conclusion, but no, definitely not. 03:02.390 --> 03:10.070 The agents SDK is very much able to work with many models, as we will later in this week, explore 03:10.110 --> 03:13.950 using different models with the OpenAI Agent Agents SDK. 03:14.090 --> 03:15.210 It's not opinionated. 03:15.210 --> 03:19.970 You can use different models with this same framework, but by default, if you just pass in a model 03:20.010 --> 03:22.330 name, it assumes you mean OpenAI. 03:23.210 --> 03:25.730 Okay, so now we're going to execute this. 03:25.730 --> 03:28.010 And now we have our first agent. 03:28.050 --> 03:29.450 Let's just look at what it looks like. 03:30.290 --> 03:32.930 It is an agent with name jokester instructions. 03:32.930 --> 03:33.930 You're a joke teller. 03:33.930 --> 03:35.250 And it's got some things here. 03:35.250 --> 03:36.090 Hand offs. 03:36.090 --> 03:36.930 You remember. 03:36.970 --> 03:42.290 That was the concept that relates to the interactions between agents. 03:42.410 --> 03:47.730 It's got a model and it's got a bunch of settings and other things that you can, you can look at in 03:47.730 --> 03:50.170 your own time to be worth looking at. 03:50.170 --> 03:54.810 It has guardrails, input guardrails and output guardrails that we will look at later. 03:55.330 --> 03:55.890 Okay. 03:56.050 --> 04:01.010 So I mentioned that runner run is the way to run one of these things. 04:01.010 --> 04:06.290 So we can say perhaps results equals runner dot run. 04:06.450 --> 04:09.730 And when you run an agent you you pass in two things. 04:09.730 --> 04:13.170 You pass in the agent itself to be run by the runner. 04:13.170 --> 04:15.930 And you pass in what is essentially the user prompt. 04:15.930 --> 04:19.030 It is the message that you are sending the agent. 04:19.070 --> 04:24.950 Let's say tell a joke about autonomous agents. 04:25.230 --> 04:27.110 Autonomous AI agents. 04:27.550 --> 04:30.350 Okay, so now we are going to run this. 04:30.390 --> 04:31.470 It seems pretty simple. 04:31.630 --> 04:32.990 What do you think is going to happen? 04:32.990 --> 04:34.470 What's going to be in result? 04:34.830 --> 04:36.550 I will print results. 04:37.110 --> 04:38.470 Let's see what we get. 04:38.470 --> 04:40.310 If we do this and we print result. 04:40.590 --> 04:41.790 Let you have any ideas. 04:42.110 --> 04:42.830 Here it is. 04:43.230 --> 04:52.070 We get a coroutine object because runner run is an async method, which means an async function, which 04:52.070 --> 04:54.630 means that it returns a coroutine object. 04:54.630 --> 04:55.150 It's not. 04:55.190 --> 04:57.190 It's not a function, it's a coroutine. 04:57.550 --> 04:59.910 And so hopefully you're expecting that. 04:59.950 --> 05:02.030 Or you guessed the what. 05:02.030 --> 05:02.790 What does that mean? 05:02.830 --> 05:06.630 That means that we can't just call, run and run and expect that to work. 05:06.630 --> 05:08.470 We have to call await. 05:09.190 --> 05:13.310 And I'm going to I'm going to restart this kernel and start again from the top so that we don't have 05:13.310 --> 05:14.630 one hanging out there. 05:14.630 --> 05:15.670 Not not scheduled. 05:15.670 --> 05:16.710 We will come through. 05:17.150 --> 05:18.190 We will do this. 05:18.230 --> 05:20.430 We will create our agent. 05:21.190 --> 05:23.160 And there it is. 05:23.360 --> 05:26.200 And now we're going to say await. 05:27.000 --> 05:29.040 We're going to say results equals await. 05:29.040 --> 05:30.440 Runner dot run like that. 05:30.440 --> 05:36.480 And so now what we know is that runner dot run, that coroutine is going to be queued by the event loop. 05:36.480 --> 05:38.960 And when we call await it's going to execute it. 05:38.960 --> 05:41.760 And it will hold until it's completed execution. 05:41.760 --> 05:45.280 And the results will go into the variable result. 05:45.680 --> 05:50.880 And then we are going to print result dot final output. 05:51.520 --> 05:55.800 And this is now hopefully going to work and tell us a joke. 05:56.120 --> 05:57.000 See what happens. 05:57.600 --> 06:00.320 Why don't autonomous agents ever get lost. 06:00.320 --> 06:04.520 Because they are always following their own self driving instructions. 06:05.240 --> 06:12.600 Okay, so OpenAI is rather better at dealing with lightweight agents frameworks than it is at telling 06:12.600 --> 06:14.920 jokes, but we'll accept it. 06:15.200 --> 06:16.640 So it works. 06:16.680 --> 06:20.520 Now you can be forgiven for saying, all right, we haven't done anything particularly special here 06:20.520 --> 06:21.560 to do with agents. 06:21.560 --> 06:24.200 All we've done is have a system prompt and a user prompt. 06:24.200 --> 06:27.380 But never fear, there'll be plenty of time for that later in the week. 06:27.380 --> 06:30.740 What we're doing right now is just showing you the basic construct. 06:30.740 --> 06:35.220 And for my final part of this, I'm going to wrap this in a trace. 06:36.140 --> 06:39.820 So for trace you wrap this in a context manager. 06:39.820 --> 06:44.260 So I say with trace telling a joke. 06:44.980 --> 06:48.260 And then I just need to put that like that. 06:48.420 --> 06:51.740 If you're not familiar with context managers then then look in the the guides. 06:51.740 --> 06:53.940 But but I imagine that you are. 06:54.100 --> 07:00.860 And what this does is it's telling OpenAI that we would like it to record these agent interactions, 07:00.860 --> 07:05.380 which is only one, and record it under the kind of heading telling a joke. 07:05.420 --> 07:12.180 And this allows you to have sophisticated agent interactions with many different agent calls involved, 07:12.180 --> 07:17.700 and have it all packaged up and available in their monitoring tools under one kind of heading. 07:17.940 --> 07:19.260 So let's run this. 07:19.660 --> 07:20.780 We'll get another joke. 07:20.980 --> 07:23.740 Why did the autonomous agent break up with its human partner? 07:23.780 --> 07:26.380 It just couldn't handle the emotional bandwidth. 07:26.420 --> 07:26.980 Okay. 07:28.020 --> 07:30.420 I guess it's not got a whole lot to do with agents, except it's. 07:30.460 --> 07:31.870 Put it in there with the human partner. 07:31.870 --> 07:32.710 That's all right. 07:32.990 --> 07:34.430 I'll let you be the judge of it. 07:34.430 --> 07:37.270 But now we can go and look at this trace. 07:37.270 --> 07:39.830 If we go to platform Openai.com. 07:41.390 --> 07:42.310 Here it comes. 07:43.150 --> 07:44.070 Traces. 07:44.070 --> 07:49.070 It's actually you can just go to platform openai.com and then just click on traces in the left, and 07:49.070 --> 07:52.950 you'll see that telling a joke is the top trace up here. 07:53.110 --> 07:56.190 And if I go into this trace we will see. 07:56.190 --> 08:01.910 What happened is that there was one endpoint called the system instructions was you were a joke teller. 08:01.910 --> 08:05.150 That's the system message that was just called instructions. 08:05.470 --> 08:09.150 The user prompt was tell a joke about autonomous agents. 08:09.150 --> 08:12.070 And then this was what the assistant responded. 08:12.070 --> 08:19.350 So true to its word, it's giving us a clear trace of the single call to an LLM that happened as part 08:19.350 --> 08:24.430 of this gigantic workflow, but it's packaged it up under the heading Telling a Joke. 08:24.630 --> 08:28.630 And this has allowed us to come into traces and see this. 08:28.670 --> 08:32.630 And as you can imagine, it's not particularly profound with this very simple example. 08:32.630 --> 08:36.390 But there will come cases in the future when this is going to be invaluable.