WEBVTT 00:00.120 --> 00:02.920 And welcome to week two, day three. 00:02.960 --> 00:04.440 We have lots to cover today. 00:04.480 --> 00:09.440 Continuing our exploration of OpenAI agents SDK and with our SDR project. 00:09.760 --> 00:14.040 So I want to quickly take a moment to recap what we did last time. 00:14.240 --> 00:22.960 We looked at tools by just using that simple decorator as as a way to allow us to wrap a function into 00:22.960 --> 00:28.480 a tool and to be able to call that tool so easily without the boilerplate JSON that we'd needed to use 00:28.480 --> 00:29.400 in week one. 00:29.680 --> 00:33.680 We then looked at the way that agents can also be turned into tools. 00:33.680 --> 00:39.160 By using the as tool function, we can simply change an agent into being a tool. 00:39.400 --> 00:42.120 And that turned out to be very easy indeed. 00:42.360 --> 00:46.840 And then slightly confusingly, we saw that there is another way that agents can collaborate with each 00:46.880 --> 00:48.920 other, which is using handoffs. 00:48.920 --> 00:54.000 When you're creating a new agent, you can give it any number of tools and you can give it handoffs. 00:54.040 --> 00:58.640 And I explained that the key difference between them is that when you're using tools, you can think 00:58.640 --> 01:02.330 of that as a call to the other agent, which will then return back. 01:02.530 --> 01:06.490 But when you're doing handoffs, you can really think of that as a transfer of control. 01:06.610 --> 01:13.490 The next step on the workflow is giving over control to that agent that is on the receiving end of the 01:13.530 --> 01:14.090 handoff. 01:14.810 --> 01:18.930 We're now going to go and extend what we did last time in three different ways. 01:18.930 --> 01:22.450 Three things that will allow us just to learn a bit more about the framework. 01:22.490 --> 01:27.650 First of all, we're going to do what I promised we'd do and look at models other than OpenAI. 01:28.050 --> 01:34.450 So using the OpenAI agents SDK to drive things like Gemini and deep tech, we're also going to look 01:34.450 --> 01:36.370 at structured outputs. 01:36.410 --> 01:42.530 The way that we can require an agent not to respond just with text, but to populate some kind of an 01:42.530 --> 01:46.490 object where we can specify the fields that are going to get populated. 01:46.810 --> 01:52.890 And then we're going to look importantly at guardrails, which are an approach to making sure that we 01:52.890 --> 01:59.370 have some controls over the information that comes in to our agent setup and what comes out. 02:00.010 --> 02:00.690 All right. 02:00.930 --> 02:04.890 With that intro, let's go back to cursor, back to the lab. 02:05.330 --> 02:07.130 And here we are back in cursor. 02:07.130 --> 02:10.050 And we are looking at the second week. 02:10.050 --> 02:12.130 And we're looking at day three. 02:12.170 --> 02:14.330 Here we are lab three week two day three. 02:14.770 --> 02:18.890 And we're going to start by doing a sort of quick recap of what we did last time. 02:18.930 --> 02:20.690 We're going to do some imports. 02:20.690 --> 02:22.570 We're going to load the env. 02:22.770 --> 02:26.770 We're actually going to bring in a bunch more environment variables than usual. 02:26.810 --> 02:29.330 I'm looking at a bunch of different keys here. 02:29.370 --> 02:30.690 Google deep sea grok. 02:30.730 --> 02:31.970 Now you don't have to do this. 02:31.970 --> 02:34.930 You can just stick with OpenAI if you don't have these other keys. 02:34.930 --> 02:39.130 But if you'd like to experiment with others, then this is the way to do it. 02:39.170 --> 02:43.690 And of course, if you have other things, perhaps you use Open Router as your way to connect to different 02:43.690 --> 02:44.210 models. 02:44.210 --> 02:46.450 You can be using any of these here. 02:47.210 --> 02:47.770 All right. 02:47.810 --> 02:52.450 Now we have the three instructions for our three different sales agents. 02:52.450 --> 02:58.740 You'll remember one is more professional, one is more witty and engaging, and one is more concise. 02:58.940 --> 03:00.980 So income are three instructions. 03:01.380 --> 03:08.060 We're now going to remember that we can use OpenAI's endpoints. 03:08.060 --> 03:15.300 We can use the compatible OpenAI endpoints to talk to other models like Gemini, Deep Seek, and Grok. 03:15.580 --> 03:22.100 And any time that we have this approach, we will be able to easily use the OpenAI agents SDK. 03:22.660 --> 03:26.220 So here we're setting the base URL for Gemini endpoint. 03:26.380 --> 03:29.140 The base URL for the deep sea endpoint and for the grok. 03:29.300 --> 03:32.100 These are the OpenAI compatible endpoints. 03:32.260 --> 03:35.260 So for Gemini you can see it's got OpenAI in there. 03:35.580 --> 03:37.540 And the same for grok for deep secret. 03:37.540 --> 03:38.260 Is there only one. 03:38.260 --> 03:40.100 It's already OpenAI compatible. 03:40.820 --> 03:49.660 Then we create a new instance of the client by instantiating an async OpenAI object, as you can see 03:49.860 --> 03:50.340 here. 03:50.380 --> 03:54.310 Passing in the base URL and passing in the key. 03:54.750 --> 03:58.710 And finally we create three model objects. 03:58.710 --> 04:06.190 These are called OpenAI chat completions model passing in the name of the model and the client that 04:06.190 --> 04:08.750 we just defined in the group above. 04:09.310 --> 04:11.470 So this is a little bit boilerplate. 04:11.470 --> 04:17.470 And if you're using GPT four or mini, if you're using any of OpenAI's models, you don't need to do 04:17.470 --> 04:21.230 any of this because it assumes by default that this is the case. 04:21.350 --> 04:27.310 So what I mean is that when we actually create our agents, we're going to give them a name instructions, 04:27.310 --> 04:29.830 and we're going to pass in one of these models. 04:30.110 --> 04:35.950 Now if for this model you pass in a string you just pass in text, which is what we did before we just 04:35.950 --> 04:38.990 passed in GPT four mini right there. 04:39.230 --> 04:43.470 If it's text, then it assumes that you're talking directly to OpenAI. 04:43.910 --> 04:52.270 But if, by contrast, you pass in a model object an instance of OpenAI chat completions Model, then 04:52.270 --> 04:56.030 it will connect to this model using this endpoint. 04:56.550 --> 05:01.030 And so as you can see we have three agents sales agents one, two and three. 05:01.230 --> 05:03.710 The first one is connecting to deep seek. 05:03.910 --> 05:06.470 The second one is connecting to Gemini. 05:06.710 --> 05:10.510 And the third one is connecting to llama 3.3. 05:10.670 --> 05:16.190 The massive llama model through grok the fast inference platform. 05:16.470 --> 05:17.390 So there they are. 05:17.430 --> 05:19.630 This is how it all connects together. 05:19.870 --> 05:21.070 And we'll run that. 05:21.070 --> 05:25.670 And we now have these three agents connecting to these three different models. 05:26.350 --> 05:31.870 We will now do exactly what we did before our description, write a cold sales email. 05:31.870 --> 05:38.230 And we're now going to repackage each of those three sales agents into tools called tool one. 05:38.270 --> 05:38.950 Tool two. 05:38.990 --> 05:39.710 Tool three. 05:39.950 --> 05:44.310 And passing in this description as the description of the tool. 05:44.550 --> 05:45.710 And that is done. 05:45.990 --> 05:49.560 And we will now have a normal function tool. 05:49.560 --> 05:58.200 We use this decorator to reconstitute wrap this function in the the the boilerplate JSON that describes 05:58.200 --> 05:58.840 this tool. 05:59.600 --> 06:01.760 And it's almost a wrap. 06:01.800 --> 06:04.600 We're zooming through everything we did last time. 06:04.600 --> 06:05.960 We've now got some instructions. 06:05.960 --> 06:08.120 You can write a subject for a cold email. 06:08.520 --> 06:11.200 We've got some HTML instructions. 06:11.200 --> 06:14.120 You can convert a text email to HTML. 06:14.800 --> 06:21.040 We then make a subject writer an agent, which is going to do this. 06:21.040 --> 06:25.160 And as I say, I'm just passing in the string GPT four mini. 06:25.400 --> 06:28.720 So this is an agent that can write the subject of an email. 06:29.080 --> 06:34.880 This is turning that into a tool, a tool that can write the subject of an email. 06:35.320 --> 06:39.680 This is an agent that can convert an email to HTML format. 06:40.040 --> 06:46.000 And here it is as a tool, a tool that can convert a text email body to HTML email. 06:46.160 --> 06:50.040 I know I'm going through this very fast, but we just did it yesterday, so hopefully this is just immediate 06:50.040 --> 06:50.760 revision. 06:50.760 --> 06:54.920 So we run that and we collect together those three tools. 06:55.920 --> 07:01.720 This is our emailer agent the agent which is able to take some instructions. 07:02.160 --> 07:03.480 The email tools. 07:03.600 --> 07:07.000 And it has a hand off description if you remember that. 07:07.360 --> 07:09.360 And now we're almost done. 07:09.680 --> 07:14.720 We now put these three sales tools into one group of tools. 07:14.920 --> 07:19.200 The hand off is now this one hand off agent. 07:19.200 --> 07:20.880 This emailer agent. 07:21.400 --> 07:22.400 We run that. 07:23.080 --> 07:27.200 And finally here we are, the same arrival as before. 07:27.240 --> 07:29.400 The sales manager's instructions. 07:29.400 --> 07:34.320 We pass in the instructions we pass in the tools, we pass in the hand offs. 07:34.520 --> 07:38.600 And we then write the cold sales email. 07:39.000 --> 07:41.440 And with that, I kick this off. 07:41.560 --> 07:45.640 It takes about a minute and I will see you in a second when it completes.