WEBVTT 00:00.280 --> 00:02.720 Well, this is a bittersweet moment for sure. 00:02.760 --> 00:05.720 It is the final day of our journey with crew. 00:06.000 --> 00:07.720 Fifth day of week three. 00:07.760 --> 00:12.440 While we complete the work we've got with crew, and it's going to be a great one. 00:12.880 --> 00:14.000 We're going to end on a high. 00:14.040 --> 00:17.440 We are going to build on the project we did last time. 00:17.480 --> 00:21.880 On day four, we're going to turn our coda into an engineering team. 00:21.920 --> 00:28.120 We're going to have an engineering lead depicted here by this person chilling out because, you know, 00:28.160 --> 00:29.640 that's what what leads do. 00:29.880 --> 00:32.280 Uh, we're going to have a back end engineer. 00:32.280 --> 00:38.840 We're going to have a front end engineer, and we're going to have a test engineer seen here as someone 00:38.840 --> 00:44.600 that's jumping on top, trying to trying to destroy, trying to break everything that they are given. 00:44.920 --> 00:46.440 So this is going to be our team. 00:46.440 --> 00:48.840 We're going to really use the crew and crew. 00:48.880 --> 00:51.400 I let's get to it. 00:51.640 --> 00:53.200 And welcome back to cursor. 00:53.240 --> 00:56.560 Welcome to our final project for crew. 00:57.000 --> 01:01.080 Let's bring up a terminal with control and the back tick. 01:01.360 --> 01:04.080 Let's go into that third directory. 01:04.080 --> 01:06.880 And now, of course, it is time to do crew. 01:06.920 --> 01:11.000 I create crew and it is engineering. 01:11.160 --> 01:14.760 Underscore team is the name of our project. 01:15.200 --> 01:15.880 Let's do it. 01:15.880 --> 01:17.880 Let's choose the options that it tells us. 01:17.880 --> 01:18.920 We want OpenAI. 01:19.320 --> 01:20.800 We want GPT four or mini. 01:20.800 --> 01:23.560 And then we skip making a token. 01:23.560 --> 01:26.720 And then there we have our directory structure. 01:26.920 --> 01:29.880 And now of course we go into engineering team. 01:29.920 --> 01:35.800 We go into source, we go into config and we open our agents YAML file. 01:35.800 --> 01:41.280 As usual we see the boilerplate ones there the researcher and the reporting analyst. 01:41.880 --> 01:43.960 So what are we going to do. 01:43.960 --> 01:46.840 Well we're going to first of all delete what we have here. 01:47.000 --> 01:49.200 And we're going to put in our first agent. 01:49.360 --> 01:54.520 And our first agent is going to be an engineering lead, the engineering lead for the engineering team 01:55.040 --> 01:57.640 directing the work of the engineers. 01:57.680 --> 02:05.480 Let's make that plural, take the high level requirements described here and prepare a detailed design 02:05.480 --> 02:06.920 for the back end developer. 02:06.960 --> 02:08.520 Maybe we'll leave this singular for now. 02:08.560 --> 02:12.600 If it's just one developer, everything should be in one Python module. 02:12.600 --> 02:15.640 Describe the function the method signatures in the module. 02:15.680 --> 02:21.000 The Python module must be completely self-contained and ready so that it can be tested or have a simple 02:21.000 --> 02:22.160 UI built. 02:22.280 --> 02:23.480 Here are the requirements. 02:23.480 --> 02:27.800 So this is going to be one of our input attributes that we can pass in. 02:28.200 --> 02:30.680 The module should be named module name. 02:30.680 --> 02:33.320 And the class should be named class name. 02:33.320 --> 02:36.400 So we are going to bring in those as well. 02:36.400 --> 02:38.720 So we're going to have this be a little bit configurable. 02:39.080 --> 02:43.920 And the backstory you're a seasoned engineering lead with a knack for writing clear and concise designs 02:43.960 --> 02:44.920 okay there we go. 02:44.920 --> 02:50.000 So so look, I'm going to put on an LM now and let's have something different to usual. 02:50.000 --> 02:51.680 I think we're going to go with Gpt4. 02:51.720 --> 02:53.440 We're going to have the big guy you. 02:53.480 --> 02:56.720 I've also run this with Gpt4 on many and it runs just fine. 02:56.720 --> 02:58.280 So you can do that too if you wish. 02:58.480 --> 03:03.560 And you get, you know, you get a bit more comprehensive, extensive solutions if you use its bigger 03:03.560 --> 03:04.840 cousin GPT four. 03:04.880 --> 03:06.280 Oh okay. 03:06.600 --> 03:06.960 so. 03:07.000 --> 03:12.520 And whilst that will work, of course the full way to do it is to show this is OpenAI slash like that 03:12.520 --> 03:13.440 with the provider. 03:14.240 --> 03:16.920 That is the first of our agents. 03:17.600 --> 03:17.960 Okay. 03:18.000 --> 03:22.120 So the next agent is going to be called a backend engineer. 03:22.280 --> 03:29.960 So the backend engineer is going to be a Python engineer who can write code to achieve the design described 03:29.960 --> 03:31.200 by the engineering lead. 03:31.200 --> 03:36.160 So the engineering lead will do the design will lay out the sort of foundations, and the back end engineer 03:36.160 --> 03:42.480 will do the coding write Python module that implements the design described by the engineering lead, 03:42.480 --> 03:44.560 but will also pass in the requirements again. 03:44.560 --> 03:49.480 So it's going to get the requirements along with the design and the module name and class name again. 03:49.720 --> 03:56.000 And it's given a decent backstory saying that it's a Python engineer with a knack for writing clean, 03:56.040 --> 03:57.480 efficient code. 03:57.880 --> 04:00.000 Follow the design instructions carefully. 04:00.400 --> 04:01.760 So again, we'll use a different model. 04:01.760 --> 04:06.480 We'll mix things up, and this time I suggest we use Claude 3.7 sonnet latest. 04:06.520 --> 04:10.240 Now sometimes we get an overload error from Claude? 04:10.280 --> 04:11.080 At least I do. 04:11.120 --> 04:11.880 As of now. 04:12.040 --> 04:17.200 So it's a little bit less stable, but nonetheless, it's a great coding model. 04:17.240 --> 04:18.400 Particularly so. 04:18.400 --> 04:21.160 I think it's a good one to use in this case. 04:21.600 --> 04:22.080 All right. 04:22.080 --> 04:24.880 Let's also add a front end engineer. 04:25.360 --> 04:26.600 And I say front end engineer. 04:26.600 --> 04:32.560 But it's going to be a gradio engineer a gradio expert who can write a simple front end to demonstrate 04:32.560 --> 04:39.360 a back end writer gradio UI that demonstrates the given back end all in one file to be in the same directory 04:39.680 --> 04:43.440 as the back end module, giving the requirements. 04:43.480 --> 04:45.040 Giving the back story. 04:45.200 --> 04:50.520 A seasoned engineer highly skilled at writing simple gradio UIs for a back end class. 04:51.000 --> 04:52.400 So there we have it. 04:53.280 --> 05:00.360 We, uh, explained that the UI needs to be in one module, and we'll use again Claude for this too. 05:00.960 --> 05:01.320 Okay. 05:01.360 --> 05:02.320 And one more agent. 05:02.360 --> 05:08.320 Our final agent is going to be a test engineer, but I'm slightly abusing the word there. 05:08.320 --> 05:08.600 It's not. 05:08.600 --> 05:10.000 It's not like a QA engineer. 05:10.000 --> 05:16.360 This is going to be a Python coder who knows how to write unit tests for a given backend module, because 05:16.360 --> 05:20.000 I want to see unit tests being built out here. 05:20.200 --> 05:22.000 So that's what the test engineer is going to do. 05:22.040 --> 05:26.800 They're going to create a test which is going to be called test underscore. 05:26.800 --> 05:30.800 And the name of the module which will be the backend module. 05:30.960 --> 05:35.720 And they're a seasoned QA engineer and software developer who writes great unit tests for Python. 05:35.720 --> 05:38.000 Code is the backstory. 05:38.160 --> 05:42.600 And you know, for for all of the coding part, we'll use anthropic. 05:42.840 --> 05:43.880 Maybe we should shake it up. 05:43.880 --> 05:44.320 Why don't we? 05:44.360 --> 05:45.760 Why don't we take a risk? 05:45.760 --> 05:47.480 We'll use deep seek instead. 05:47.520 --> 05:48.240 Deep seek. 05:48.480 --> 05:54.600 I'm just going to stick with deep seek chat because deep seek chat is has is also able to write terrific 05:54.600 --> 05:55.440 code as well. 05:55.440 --> 05:59.680 So we'll we'll stick with that model and that will be our test engineer. 05:59.720 --> 06:01.640 Now it's time for us to go on to the tasks. 06:02.440 --> 06:03.680 So look it's a bit repetitive. 06:03.720 --> 06:05.200 We're going to let me get rid of the gunk. 06:05.240 --> 06:08.200 We're going to have a task for each agent. 06:08.200 --> 06:08.680 And it's funny. 06:08.680 --> 06:14.720 This is one of those times when it feels like OpenAI's agents SDK is sort of more straightforward that 06:14.720 --> 06:20.480 they don't have this extra indirection, this idea of a task being a separate first class object, because 06:20.480 --> 06:24.840 there's a lot of times that you will have this 1 to 1 mapping between a task and an agent. 06:24.880 --> 06:29.240 Obviously, crew has built this up because they really want you to think in terms of situations where 06:29.240 --> 06:33.280 you will have series of tasks that will be given to the same agent. 06:33.280 --> 06:35.720 And we did see that and at least one of our examples. 06:35.720 --> 06:41.120 But in this case, we are again going to have the case that there's going to be a task for each of our 06:41.120 --> 06:45.640 agents, and it gives us an opportunity to spell out very clearly what's required. 06:46.360 --> 06:54.120 So the description of the task is to take the high level requirements described and prepare a detailed 06:54.120 --> 06:55.440 design for the engineer. 06:55.440 --> 07:00.520 This is the design task that, of course, is going to be assigned to our engineering lead. 07:00.840 --> 07:03.800 And the output is a detailed design for the engineer. 07:04.320 --> 07:09.280 You'll notice that I make it important that it should only output the design in markdown format. 07:09.320 --> 07:14.120 Otherwise it just wants to start writing the code, which is not exactly what we want. 07:14.360 --> 07:19.440 And here we specify the the output as the module name underscore design. 07:19.440 --> 07:25.440 And it's worth noting that you can have these templated tags in the output file name as well. 07:25.440 --> 07:31.480 You can have it anywhere in this YAML, and it will get substituted in for the actual name of the the 07:31.640 --> 07:33.120 module that we're building here. 07:33.960 --> 07:36.480 So that is the design task. 07:36.520 --> 07:39.120 My next one cursor wants it to be called coding task. 07:39.120 --> 07:41.760 But when I wrote it I wrote code task. 07:41.800 --> 07:42.320 There we go. 07:42.600 --> 07:44.280 Uh, I'd like to be unpredictable. 07:44.520 --> 07:51.280 Uh, so the description is write a Python module that implements the design as described by the engineering 07:51.280 --> 07:53.560 lead in order to achieve the requirements. 07:53.800 --> 07:59.920 And then here I'm saying important output only the Python code without any markdown formatting or code 07:59.960 --> 08:02.080 block delimiters or backticks. 08:02.280 --> 08:05.560 And that's important because these models love to output markdown. 08:05.560 --> 08:09.680 And if you don't put that note in there, it's going to put that tick, tick, tick python at the top 08:09.680 --> 08:12.320 of the file, which means it won't be a valid Python file. 08:12.440 --> 08:14.080 And it just loves doing that. 08:14.080 --> 08:17.240 So unless you really highlight this is important, it's going to do it. 08:17.600 --> 08:19.400 And so that is the task. 08:19.400 --> 08:25.120 And we're saying context design task, which you remember is going to make sure that that information 08:25.120 --> 08:27.000 is made available to it. 08:27.320 --> 08:32.360 And generally speaking you can think of the tasks as being the user prompts and the agents as defining 08:32.360 --> 08:35.520 the system prompts, and that that's kind of how it all comes together. 08:35.520 --> 08:40.400 And this context is showing us what kind of information is going to be included in that prompt from 08:40.440 --> 08:41.040 others. 08:41.720 --> 08:42.200 Okay. 08:42.240 --> 08:44.560 So that is the code task. 08:44.560 --> 08:47.080 How about the front end task. 08:47.480 --> 08:49.560 So this is where the plot thickens. 08:50.000 --> 08:58.360 The front end has got to write a Gradio UI in a module App.py that demonstrates the given back end class. 08:58.400 --> 09:03.280 Assume there's only one user and keep the UI very simple indeed. 09:03.320 --> 09:05.000 Just a prototype or demo. 09:05.160 --> 09:06.520 Here are the requirements. 09:06.680 --> 09:08.720 And so yeah, we keep it. 09:08.760 --> 09:10.200 Keep it nice and simple. 09:10.840 --> 09:12.800 Maybe I'm going to take out the very simple indeed. 09:12.800 --> 09:16.280 Let's just say keep the UI simple and keep the UI clean. 09:17.040 --> 09:18.080 Let's leave it at that. 09:18.960 --> 09:21.120 So, um, important. 09:21.120 --> 09:23.400 Only output the raw Python code again. 09:23.440 --> 09:26.720 The agent is the front end engineer agent, obviously. 09:26.920 --> 09:32.720 And of course, for its context, it needs to get the output from the code task because it's going to 09:32.760 --> 09:38.240 write a user interface, which is going to directly use what the code task has written. 09:38.360 --> 09:43.440 So that's going to be very interesting to see how how it's able to keep that context so well that it's 09:43.480 --> 09:44.840 actually able to do this. 09:45.200 --> 09:51.680 And then cursor is correct that the last one is indeed called test task, but it's got a little bit 09:51.680 --> 09:53.680 more detail than cursor was suggesting. 09:53.880 --> 09:58.560 Write unit tests for the given back end module and create a test in the same directory. 09:58.880 --> 10:00.400 And this is. 10:00.440 --> 10:07.080 Got more information one more time about outputting raw Python without clever tags. 10:07.120 --> 10:09.640 And of course it also depends on the code task. 10:09.680 --> 10:12.960 The unit tests are not going to test the user interface, it's going to test the code. 10:13.200 --> 10:15.280 And that will be the output file. 10:15.280 --> 10:20.800 And that is it for our tasks our 1 to 1 again between the task and the agent.