WEBVTT 00:00.360 --> 00:05.000 So look, you're here because you like getting hands on and building stuff. 00:05.000 --> 00:06.320 And that's why I'm here, too. 00:06.360 --> 00:07.920 And I got good news for both of us. 00:07.920 --> 00:09.480 That's what we're doing right now. 00:09.520 --> 00:16.960 Today we're going to be building our first OpenAI agents SDK project, building a sales development 00:16.960 --> 00:17.520 rep. 00:18.320 --> 00:22.360 And we're going to be building really three different pieces. 00:22.360 --> 00:26.280 We're going to be building three different layers of Agentic architecture where we do this. 00:26.280 --> 00:30.800 We're going to be starting with something quite simple, a workflow of agent calls. 00:30.800 --> 00:35.200 We're then going to spice it up by adding an agent that can use a tool. 00:35.440 --> 00:40.320 Remembering back to when we did this, the sort of manual way with the the JSON with the boilerplate 00:40.320 --> 00:41.160 in week one. 00:41.480 --> 00:46.360 And then we're going to have agents that can call on other agents, and you're going to discover there 00:46.400 --> 00:48.160 are two different ways of doing this. 00:48.160 --> 00:53.960 You can treat agents as tools, or you can use the thing called handoffs that I mentioned before. 00:54.200 --> 00:55.360 So keep this in mind. 00:55.360 --> 00:56.840 Keep this construct in mind. 00:56.840 --> 00:58.360 This is how we'll be approaching it. 00:58.360 --> 00:59.880 There's a lot of coding to do. 00:59.920 --> 01:00.960 Let's get started. 01:01.520 --> 01:05.080 And so here we are back in cursor ready for the day. 01:05.080 --> 01:07.200 To begin, we open the OpenAI folder. 01:07.200 --> 01:08.760 We go to lab two. 01:09.240 --> 01:10.640 Here we are at the start of week two. 01:10.640 --> 01:12.200 Lab two, day two. 01:12.240 --> 01:13.880 We're going to do an agent workflow. 01:13.880 --> 01:16.080 We're going to use tools to call functions. 01:16.080 --> 01:19.360 And then we're going to collaborate with tools and handoffs. 01:19.640 --> 01:23.200 So before we start we're going to be making use of a tool. 01:23.400 --> 01:24.960 It's called SendGrid. 01:24.960 --> 01:29.640 And it will allow you to send transactional emails from their servers. 01:29.640 --> 01:30.520 It is free. 01:30.520 --> 01:32.720 So it's going to be very easy to set up. 01:32.720 --> 01:34.160 It's actually owned by Twilio. 01:34.160 --> 01:35.760 So it's a very well known company. 01:35.800 --> 01:38.360 It's similar to something like Sparkpost if you've used that before. 01:38.360 --> 01:40.640 And of course you can use Sparkpost instead if you prefer. 01:41.080 --> 01:42.440 So there's a link here. 01:42.440 --> 01:48.120 If you visit SendGrid right there up, it will come and you can start for free by pressing this button 01:48.120 --> 01:49.040 right here. 01:49.240 --> 01:53.560 And once you do that and you've set up an account, I will switch to another window. 01:53.720 --> 01:54.240 Here we go. 01:54.280 --> 01:56.240 This is what I get once I've signed in. 01:56.280 --> 02:00.480 You can see that I've sent 1 or 2 emails, and there's just a few things for you to do. 02:00.480 --> 02:06.120 Once you've done this, you go to the settings menu down here and to the API keys section. 02:06.120 --> 02:10.110 That is where you will set up an API key and copy it to your clipboard. 02:10.350 --> 02:12.350 You can guess where that's going to go in just a second. 02:12.590 --> 02:17.430 The other thing to do is to go to this tab here Sender Authentication. 02:17.630 --> 02:21.390 And you're going to click this verify a single sender button. 02:21.390 --> 02:26.110 And what that's going to do is it's going to allow you to set up your your own email address. 02:26.110 --> 02:28.310 This is my email address right here. 02:28.310 --> 02:32.110 And do a verify email to confirm that you do own that email. 02:32.110 --> 02:36.510 And once you've done that, you'll be able to send emails from that email address. 02:36.670 --> 02:41.950 So that's a nice thing to do, just so that you'll be able to send emails via SendGrid from that email 02:41.950 --> 02:42.390 address. 02:42.390 --> 02:45.670 But don't worry, we're not actually going to be sending any cold emails out to the world. 02:45.670 --> 02:50.430 It will only be emails going back to you again, this would just be experimental, but you can always 02:50.430 --> 02:53.150 use it for more purposes in the future. 02:53.550 --> 02:55.150 So that is SendGrid. 02:55.190 --> 02:56.670 Please go ahead and get that set up. 02:56.670 --> 02:59.350 And shockingly, I'm sure you expected this. 02:59.390 --> 03:06.350 Once you've done that, you'll then edit your env file over there and you will add in a single row SendGrid 03:06.390 --> 03:07.990 API key equals. 03:07.990 --> 03:13.750 And of course you will paste in there the thing that you've copied from the API keys section. 03:13.790 --> 03:16.790 You created a key there in SendGrid itself. 03:16.830 --> 03:17.390 All right. 03:17.390 --> 03:19.710 So go ahead and do that and then we will proceed. 03:20.350 --> 03:20.910 Okay. 03:20.990 --> 03:24.070 And so now we have some imports. 03:24.070 --> 03:25.950 We're going to be doing some usual stuff. 03:25.950 --> 03:28.750 We've got a few more from agents. 03:28.750 --> 03:31.350 We're also importing this thing called function tool. 03:31.630 --> 03:34.830 We're importing something to help us stream back text. 03:34.990 --> 03:37.990 And we're importing some SendGrid stuff. 03:38.310 --> 03:42.430 We're going to load our environment variables as usual. 03:42.710 --> 03:49.270 And now let's do a simple workflow between agents without anything fancy. 03:49.310 --> 03:53.790 Here we're just going to have some agents calling other agents, and we're going to orchestrate it manually 03:53.790 --> 03:55.950 ourselves in Python code. 03:55.950 --> 03:57.670 And it's going to be super simple. 03:58.110 --> 04:00.750 So first of all three instructions. 04:01.350 --> 04:03.870 And these instructions again think of it in the back of your mind. 04:03.870 --> 04:06.110 These are basically system prompts. 04:06.110 --> 04:09.070 But they are the way that we are framing different agents. 04:09.390 --> 04:15.260 All three are about being sales agents working for a company that I'm calling. 04:15.260 --> 04:16.180 Comply. 04:16.500 --> 04:17.180 Comply. 04:17.340 --> 04:20.300 I are the wonderful naming. 04:20.300 --> 04:27.420 This is a Soc2 compliance company that is looking at helping companies prepare for Soc2 audits. 04:27.660 --> 04:31.300 So something to do with compliance if you're not familiar with this stuff. 04:31.300 --> 04:35.620 But, uh, yeah, it's a SaaS tool that will ensure soc2 compliance. 04:35.820 --> 04:38.460 So it's saying you're a sales agent working for this company. 04:38.700 --> 04:42.860 You write professional, serious cold emails. 04:42.900 --> 04:46.420 It's like a cold sales email or email to to get people interested. 04:46.780 --> 04:49.900 The second one is a humorous, engaging sales agent. 04:49.900 --> 04:54.820 You write witty, engaging, cold emails that are likely to get a response. 04:55.460 --> 05:02.100 And the third one is you're a busy sales agent and you write concise to the point cold emails. 05:02.220 --> 05:08.660 So, as you know from good prompting, you use a system prompt instructions to set the context, to 05:08.700 --> 05:10.780 give the tone to sort of set. 05:10.820 --> 05:16.380 Set the mood and the character and give as much background information as you can in a succinct, instructive 05:16.380 --> 05:16.820 way. 05:17.580 --> 05:20.300 Okay, so now we create three agents. 05:20.620 --> 05:22.740 Sales agent one, two and three. 05:22.780 --> 05:24.500 And this should be very simple. 05:24.500 --> 05:26.060 It should make complete sense to you. 05:26.100 --> 05:28.340 Each agent gets given a different name. 05:28.340 --> 05:31.580 They have the three instructions the three system prompts. 05:31.580 --> 05:34.820 And we're going to be using GPT four mini throughout. 05:34.860 --> 05:35.900 Keep it cheap. 05:36.100 --> 05:38.460 We've just created those three agents. 05:38.900 --> 05:39.220 All right. 05:39.220 --> 05:41.100 And now we're just going to start with the first one. 05:41.460 --> 05:43.380 And I just wanted to show you a different way. 05:43.420 --> 05:46.180 You remember we had runner run before. 05:46.300 --> 05:50.660 Well now I'm calling runner run streamed, which is a way that you can do it. 05:50.660 --> 05:53.540 That will allow us to stream back results. 05:53.580 --> 05:59.220 You may be familiar with streaming from from simpler OpenAI examples, but this is how you can do it 05:59.220 --> 06:00.940 with the agents SDK. 06:01.460 --> 06:07.340 So you may spot that there's no ace a wait word there which which may set off alarm bells, because 06:07.340 --> 06:09.940 that means that what we're getting here is not a response. 06:09.940 --> 06:11.820 We're getting a co-routine. 06:12.100 --> 06:13.860 And you can see what happens next. 06:13.900 --> 06:22.690 We use a special construct async for a special way of of then calling this, calling this in a way that 06:22.690 --> 06:27.330 will return a coroutine, and we will then iterate through those answers. 06:27.330 --> 06:32.650 And this is just a little bit of boilerplate code to make sure that what's coming back is some text 06:32.650 --> 06:33.650 that we can print. 06:33.650 --> 06:35.890 And if we get that, then we print it. 06:36.210 --> 06:39.810 And this just stops it printing on a separate line every time. 06:40.090 --> 06:44.010 So if I run this, it's just very similar to what we did last time. 06:44.010 --> 06:49.170 But you'll see that what comes back streams back bit by bit. 06:49.370 --> 06:54.290 And it's a very nice use of the streaming APIs and I show it to you. 06:54.290 --> 06:56.890 So you have an example of how to how to do this. 06:57.170 --> 07:02.130 And of course, just as you would expect, we see a great behavior from our agent. 07:02.330 --> 07:07.450 This is the professional one, I think, and we can see that it's come back with something nice and 07:07.450 --> 07:09.130 professional, nice and slick. 07:09.170 --> 07:11.610 To simplify your soc2 compliance efforts. 07:11.610 --> 07:17.090 And it's no surprise to to everyone that's worked with Llms for a bit that they are really good at this 07:17.130 --> 07:17.650 kind of thing. 07:17.650 --> 07:24.290 They can write great professional, realistic, reasonable sales outreach emails. 07:24.810 --> 07:26.130 So far, so good.