WEBVTT 00:00.360 --> 00:01.840 And now it gets juicy. 00:01.880 --> 00:07.200 Now we're going to go to Crewe and we're going to start building a bunch of different things. 00:07.800 --> 00:11.520 So the first thing is that I want us to use structured outputs. 00:11.520 --> 00:17.520 In other words, we are going to ask our different tasks to be providing information according to a 00:17.520 --> 00:23.160 particular JSON schema as a way of making sure we're getting the information that we want from these 00:23.160 --> 00:26.560 agents in a robust sort of way that is kind of on rails. 00:26.840 --> 00:32.960 And so, as you probably remember, the way to do that is to create classes that are subclasses of base 00:32.960 --> 00:36.760 model and use this as a way of describing what we want. 00:36.840 --> 00:38.080 So let me give you an example. 00:38.080 --> 00:41.120 Let's make a class that's called trending company. 00:41.480 --> 00:43.800 And so this is a subclass of base model. 00:43.800 --> 00:45.120 We give it an explanation. 00:45.120 --> 00:47.760 It's a company in the news attracting attention. 00:48.000 --> 00:51.120 And then we set up name ticker and reason. 00:51.120 --> 00:53.240 And we give them descriptions like this. 00:53.480 --> 00:58.560 And you can see this is a way of, of laying out the information that we're going to want to be gathering. 00:58.560 --> 01:04.280 And it helps guide the agents and the tasks to, to to produce information that we want. 01:04.800 --> 01:07.120 And now I'm going to make a second class. 01:07.120 --> 01:09.880 That is basically just a list of these. 01:09.880 --> 01:11.960 So it's called trading company list. 01:11.960 --> 01:17.120 And it contains a single field companies which is a list of these objects. 01:17.160 --> 01:23.680 It's a way to organize it so that one task can result in a bunch of trading companies. 01:24.800 --> 01:25.640 So far so good. 01:25.800 --> 01:29.440 We're going to do pretty much the same thing again for our research. 01:29.560 --> 01:33.720 So we're going to have a trending company research schema. 01:33.920 --> 01:34.760 Here it is. 01:34.960 --> 01:39.720 So again it's a it's a pedantic class a class that is a subclass of base model. 01:39.920 --> 01:42.000 And it's detailed research on a company. 01:42.000 --> 01:46.800 It has a name, market position, future outlook and investment potential. 01:47.160 --> 01:53.040 And by giving the fields with that name and with that description, we are going to force the agent 01:53.040 --> 01:55.560 to produce that information in its response. 01:55.600 --> 02:01.040 And so it's a really clever way of making sure that we guide the agent's behavior. 02:01.680 --> 02:09.840 And similarly we're going to add a list trending company research list, a detailed list of of all of 02:09.840 --> 02:15.250 the research on the companies, and it's a list of trending company research objects of these objects 02:15.250 --> 02:16.490 with these descriptions. 02:17.170 --> 02:23.050 And again, I find it helpful to be mindful that you use consistent, clear terminology. 02:23.210 --> 02:23.930 I used to have. 02:23.970 --> 02:28.850 I used to my first version of this, I called these newsworthy companies, and I was just introducing 02:28.850 --> 02:32.610 concepts that weren't crystal clear in terms of what I was going for. 02:32.650 --> 02:40.050 And and this has helped make it more reliable using simple and and common terms and using them consistently. 02:40.570 --> 02:44.530 Okay, now it's time to define our crew, our class stock picker. 02:44.770 --> 02:48.570 And as usual we look at the YAML files. 02:48.570 --> 02:49.290 We bring them in here. 02:49.290 --> 02:54.610 And I've deleted the rest of the gunk that it auto generates so that we can build this ourselves from 02:54.610 --> 02:55.250 scratch. 02:55.490 --> 03:03.450 And we're going to start by creating our trending company finder agent, which I'll do like this. 03:03.770 --> 03:08.610 And we need to give it of course the decorator agent like that. 03:08.610 --> 03:10.890 So this is a trending company finder. 03:11.050 --> 03:18.450 It's going to return an agent configured with the config and we're going to use a tool which will be 03:18.450 --> 03:19.930 let me tidy this up for you. 03:20.250 --> 03:20.770 There we go. 03:20.810 --> 03:26.810 It's going to use the super dev tool as before so that it can look for trending companies on the internet. 03:27.250 --> 03:28.810 That seems pretty good. 03:28.850 --> 03:30.770 All right, let's make our second agent. 03:31.210 --> 03:35.930 And this second agent is going to be our financial researcher agent. 03:36.250 --> 03:40.490 And so let's see we're going to want to again use Sirpa. 03:41.050 --> 03:50.050 So we're going to have it look at the config again for the financial researcher and the tools we. 03:50.130 --> 03:51.330 I was wondering what was going on there. 03:51.330 --> 03:58.050 I was missing the deaf deaf financial researcher like so that's going to have an agent. 03:58.690 --> 04:01.130 And there we go. 04:01.170 --> 04:02.490 Now it should work fine. 04:02.890 --> 04:03.370 All right. 04:03.370 --> 04:04.410 That seems good. 04:04.410 --> 04:10.810 And it has suggested cursor has kindly suggested that we do this, which seems good to me, except we 04:10.850 --> 04:12.610 don't really need this. 04:12.610 --> 04:19.850 We don't need to have the Sirpa dev tool in there, because the stock picker has been given all of its 04:20.020 --> 04:22.860 information, it's not going to need that at all. 04:23.340 --> 04:24.260 All right. 04:24.460 --> 04:27.860 So that is the definition of our three agents. 04:28.020 --> 04:30.260 Let's go on to defining our tasks. 04:30.900 --> 04:31.420 Okay. 04:31.460 --> 04:34.060 So it's time to define the first of our tasks. 04:34.060 --> 04:38.940 And the first task is the task to find trending companies using the trending company finder. 04:39.100 --> 04:40.100 And here it is. 04:40.300 --> 04:43.220 Find trending companies is creating a task. 04:43.540 --> 04:44.860 This is the config. 04:45.060 --> 04:48.660 And we also have this field output pedantic. 04:48.980 --> 04:56.700 And that is telling it that this task needs to output some JSON in the schema that conforms to trending 04:56.700 --> 04:57.820 company list. 04:57.820 --> 05:00.580 And that of course is this object right here. 05:00.700 --> 05:06.620 So we are constraining it to make sure that it produces something in that format, which is exactly 05:06.620 --> 05:07.420 what we want. 05:07.460 --> 05:11.180 And so this is the way that you do structured outputs in crew. 05:11.700 --> 05:15.260 And now the research trending companies task. 05:15.420 --> 05:18.420 This is going to again be hooked up to the right config. 05:18.580 --> 05:22.020 And this of course is going to create the trending company's research list. 05:22.060 --> 05:23.260 Pedantic object. 05:23.580 --> 05:24.980 So here we go. 05:25.020 --> 05:28.020 If we look back up, this is, of course, this object right here. 05:28.060 --> 05:31.260 A list of research results. 05:31.380 --> 05:38.660 So we are saying this task to research is going to be forced to produce information that conforms to 05:38.700 --> 05:39.660 that schema. 05:39.780 --> 05:42.300 And that way we know we're going to get research. 05:43.220 --> 05:47.060 Hopefully you're following along with this. 05:47.100 --> 05:50.380 If not then do look at the code and try this out for yourself. 05:50.500 --> 05:51.700 We're almost done. 05:51.700 --> 05:58.100 We're just going to add in the best company picker, which is right here, and which is going to select 05:58.100 --> 05:59.420 the best company. 05:59.620 --> 06:02.660 And that's an easy one, an easy one to end with. 06:02.700 --> 06:05.540 And now we've got left to do is actually create our crew. 06:06.180 --> 06:06.580 All right. 06:06.620 --> 06:07.660 Time to create the crew. 06:07.900 --> 06:09.220 We've made the agents. 06:09.220 --> 06:10.580 We've made the tasks. 06:10.780 --> 06:12.140 Here is the crew. 06:12.580 --> 06:16.060 You may be thinking to yourself, we only created three agents. 06:16.060 --> 06:17.420 Did we have a fourth agent? 06:17.460 --> 06:20.340 Yes, we did, but that fourth agent is a bit different. 06:20.340 --> 06:25.380 It's the manager, and we don't want that to be in the list of the of the, the general agents that 06:25.380 --> 06:27.300 are going to be working on the task at hand. 06:27.340 --> 06:30.630 We're going to want to create this separately and handle it separately. 06:30.630 --> 06:37.710 So we create our manager agent like this just as a, as a as a separate variable manager, an agent 06:37.710 --> 06:41.190 that has the config from that's called manager. 06:41.190 --> 06:44.990 And we also use this field allow delegation equals true. 06:45.230 --> 06:50.230 That is telling crew that we want it to be able to delegate to other agents. 06:50.350 --> 06:54.870 That would be the equivalent of the handoff in the OpenAI agents SDK. 06:55.550 --> 06:55.910 All right. 06:55.950 --> 07:02.670 And so now it just remains for us to actually return our crew from this function here from Def Crew. 07:02.670 --> 07:04.190 And we're going to do it like this. 07:04.670 --> 07:05.950 Return crew. 07:06.070 --> 07:08.270 The agents is self dot agents. 07:08.270 --> 07:11.270 And that is a three agents with the decorators above. 07:11.310 --> 07:13.990 Those are the agents that form this team, this crew. 07:14.190 --> 07:16.710 The tasks are the tasks that we defined. 07:16.750 --> 07:22.030 We're now saying the process is process hierarchical not process sequential. 07:22.030 --> 07:28.590 And that means that we are going to assign an LM to figure out which agent does what task. 07:28.950 --> 07:29.830 Verbose is true. 07:29.830 --> 07:32.350 And this is where we specify the agent. 07:32.350 --> 07:34.510 It's the agent we created right here. 07:34.510 --> 07:36.910 We could just just create that agent right there in the code. 07:36.910 --> 07:41.030 But it's a bit neater to do it this way and that is the end of it. 07:41.390 --> 07:43.310 Now there is actually an alternative here. 07:43.310 --> 07:46.430 You don't need to create a separate manager agent. 07:46.430 --> 07:52.670 You can actually say manager LM equals and just define an LM like GPT four or something. 07:52.750 --> 08:00.990 But um, I found it perform slightly better if I get the opportunity to describe the role of the manager. 08:00.990 --> 08:02.350 But both did work. 08:02.390 --> 08:04.270 But as you'll see, neither works perfectly. 08:04.390 --> 08:04.790 Haha. 08:04.830 --> 08:05.630 It's an adventure. 08:05.630 --> 08:13.150 As I say, it's a very interesting insight into some of the challenges of of autonomous AI, but still 08:13.150 --> 08:16.110 it works better if you define the manager separately. 08:16.150 --> 08:20.190 And one other point I'll make is that in defining that manager, I don't know if you spotted this, 08:20.190 --> 08:23.790 but I'm actually using GPT four, not GPT four mini. 08:23.790 --> 08:28.830 I'm using the bigger version, which means it's a slightly more pricey, so you can make that for a 08:28.830 --> 08:35.790 mini if you'd prefer, but I found that that helped it stay more coherent with the mission at hand. 08:36.110 --> 08:39.190 So anyway, that is defining the crew.