WEBVTT 00:00.240 --> 00:02.880 And now for something completely different. 00:03.080 --> 00:10.200 We are going to work on making an agent that knows how to code, how to write Python code, how to write 00:10.240 --> 00:14.360 software, and a little bit more than that, it's not only going to know how to write software, but 00:14.400 --> 00:16.000 how to run it as well. 00:16.400 --> 00:25.120 So this is challenging and complex, but it's possible to have an agent in Crewe in the Crewe environment, 00:25.120 --> 00:26.320 which has this ability. 00:26.360 --> 00:32.480 Not only is it able to to take a problem that you set it, it can write code to solve that problem. 00:32.720 --> 00:36.560 It can then execute that code in a Docker container. 00:36.560 --> 00:41.000 It doesn't need to, but it's good to have it run it in a Docker container, which means that it is 00:41.040 --> 00:46.120 a sort of sandboxed, ring fenced environment that doesn't have access to do damage to your computer 00:46.120 --> 00:52.280 because it's in this protected world, and then it can run the code, it can execute it in that container 00:52.280 --> 00:57.120 and look at the results and interpret those results to take on some more action. 00:57.120 --> 01:01.790 So as I say this, this is this is a really advanced Task. 01:01.830 --> 01:06.150 This is something that's taking us to the next level, and it's going to be hard and complex. 01:06.670 --> 01:09.390 Except it's not going to be hard or complex. 01:10.510 --> 01:16.390 You probably saw that coming, but it is going to be as simple as when we create an agent, we're going 01:16.470 --> 01:18.910 to say allow code execution equals true. 01:19.390 --> 01:20.390 And that's going to be it. 01:20.390 --> 01:22.470 And then it's going to be able to execute code. 01:22.590 --> 01:27.310 And if you say code execution mode equals safe, then if you have Docker installed on your system and 01:27.310 --> 01:31.950 I will show you how to install Docker, if you don't, it will run it in a Docker container. 01:32.190 --> 01:34.470 And it seems almost so good. 01:34.510 --> 01:36.030 It's uh, it's crazy. 01:36.030 --> 01:40.670 I had to like close down docker to make sure that that it that it failed when I, when, when uh, when 01:40.670 --> 01:43.910 it was closed because I almost couldn't believe it was happening. 01:43.950 --> 01:45.590 It's that simple. 01:45.590 --> 01:51.990 And this is where frameworks like crew really stand out because they make things like this so very easy 01:51.990 --> 01:52.510 to do. 01:52.910 --> 01:56.910 It's worth mentioning that these this kind of system is sometimes called a Coda agent. 01:56.950 --> 01:58.870 It's various other names for it, you hear. 01:58.870 --> 02:03.610 And it's worth knowing that when people talk about Coda agents, they're not just talking about agents 02:03.610 --> 02:05.250 that can deliver code. 02:05.290 --> 02:11.010 Agents that are able to generate Python, but also this ability to generate Python and then run that 02:11.010 --> 02:17.970 Python as steps towards solving a greater problem so that this is more a means to an end than just saying 02:18.130 --> 02:20.330 it is something which which writes code. 02:20.370 --> 02:23.610 Now, confusingly, our project for this week we are going to build something that writes code. 02:23.650 --> 02:27.290 But but the fact that that it can run it too is kind of cool. 02:27.610 --> 02:29.610 So that is the introduction. 02:29.610 --> 02:32.250 And with that let's go and actually build something like this. 02:32.290 --> 02:33.890 Let's build something that can write it. 02:33.930 --> 02:36.170 We'll make something that can write code and can run it. 02:36.170 --> 02:37.130 Let's do that. 02:37.170 --> 02:38.010 See you in a sec. 02:38.450 --> 02:38.890 Okay. 02:38.890 --> 02:40.370 Welcome back to cursor. 02:40.410 --> 02:42.970 Welcome back to the crew folder. 02:43.210 --> 02:49.250 And I'm going to open up a new terminal and go into that third folder. 02:49.250 --> 02:50.810 And you know how to create a crew project. 02:50.810 --> 02:53.210 So well now that you have it committed to memory. 02:53.330 --> 02:59.770 But you know that it's crew I, which I can't spell crew, I create crew. 03:00.210 --> 03:07.080 And then the name of the project, which we will call coda, and it's going to build the directories, 03:07.080 --> 03:08.520 it's going to do the scaffolding. 03:08.520 --> 03:13.520 All we have to say that we want OpenAI and we want GPT for our mini, and we don't need a key. 03:14.000 --> 03:19.280 And Kabam, we have all of the folder structure. 03:19.320 --> 03:20.040 Great. 03:20.400 --> 03:20.960 Okay. 03:21.000 --> 03:28.120 So as always, the first thing we do is we go into the config and we start by defining our agents. 03:28.440 --> 03:30.160 So what agents are we going to have. 03:30.160 --> 03:32.080 We're only going to have one this time. 03:32.080 --> 03:34.240 It's going to be a very simple agent. 03:34.240 --> 03:37.280 Let me just remove this so we can see what we're doing. 03:37.520 --> 03:44.560 Let's delete the the scaffolding and put in our coda agent. 03:44.800 --> 03:49.080 So we're going to say you are a Python developer. 03:49.080 --> 03:54.240 That's the role you write Python code to achieve this assignment. 03:54.240 --> 04:00.320 And then we pass in the assignment that is the input that will be set in the run method. 04:00.880 --> 04:03.000 First you plan how the code will work. 04:03.260 --> 04:07.460 Then you write the code, then you run it and check the output. 04:08.140 --> 04:08.940 Backstory. 04:08.980 --> 04:12.180 A seasoned Python developer with a knack for writing clean, efficient code. 04:12.220 --> 04:14.700 And we can use GPT four or mini. 04:14.740 --> 04:18.380 Yes, you can do it either way, but we'll do what cursor suggests. 04:18.420 --> 04:22.660 Change it to have the provider and the name of the model which looks a bit better. 04:23.580 --> 04:26.020 Okay, so there we go. 04:26.060 --> 04:26.940 That's pretty simple. 04:26.940 --> 04:29.700 That was our single agent. 04:29.980 --> 04:32.900 And we're going to have one single task. 04:33.300 --> 04:34.580 Let's put it in here. 04:35.020 --> 04:38.460 It is right Python code to achieve this assignment. 04:38.500 --> 04:40.220 A text file that includes the code. 04:40.300 --> 04:45.900 So so the output we want from it is a text file that includes the code and the output of the code. 04:45.900 --> 04:48.060 We want it to put both in the output. 04:48.260 --> 04:52.820 So I've called it code and output text as the output. 04:53.140 --> 04:54.820 So that is what we want it to do. 04:54.860 --> 04:59.820 So because I really want to make the point that it's going to write code, run it, get the output and 04:59.820 --> 05:03.540 it's going to think all of this through and then respond with all of that. 05:03.540 --> 05:08.330 We don't just want it to give us the output because we want to double check what it's doing is right. 05:08.850 --> 05:11.890 Okay, now we know we have to go to the crew module. 05:11.890 --> 05:13.010 Let's do that next. 05:13.530 --> 05:13.970 Okay. 05:13.970 --> 05:19.010 Here we are in the crew module that I've just gone through and deleted the templating stuff that's in 05:19.010 --> 05:20.290 there by default. 05:20.530 --> 05:24.050 And so we're just leaving our config relationships here. 05:24.130 --> 05:26.210 And I'm just going to put a little comment here. 05:26.650 --> 05:27.970 Make that a comment as well. 05:28.010 --> 05:28.450 Whoops. 05:28.810 --> 05:34.850 So this is a link to how you can install Docker if you don't already have it installed. 05:34.850 --> 05:40.370 And this is just literally and this is the Docker desktop web page for you. 05:40.570 --> 05:45.890 And as it does claim, it is a one click install for Mac or Windows or Linux. 05:45.890 --> 05:47.810 And it should be as simple as that. 05:47.970 --> 05:52.130 Many of you, I suspect from an engineering background, already know and love Docker. 05:52.170 --> 05:54.530 If you don't, then welcome to it. 05:54.530 --> 06:00.570 This should be as simple as installing it and then it's installed and you can then leave it be, but 06:00.570 --> 06:06.080 you will need to have it installed for this to work properly in a Docker setting. 06:06.320 --> 06:06.920 Okay. 06:07.120 --> 06:09.000 So now we are going to create our agent. 06:09.000 --> 06:12.840 We start with an agent decorator to make sure it goes in the right place. 06:13.040 --> 06:15.120 And we are going to create a coder. 06:15.320 --> 06:20.520 I'm going to ignore the fact that cursor is already probably several steps ahead of me, rendering my 06:20.520 --> 06:26.280 job here useless, but we'll keep going anyway, I'm going to return an agent. 06:26.280 --> 06:27.760 This is the usual fare. 06:28.120 --> 06:29.520 So what is our agent going to do? 06:29.520 --> 06:31.760 Well, the I will press tab there. 06:31.800 --> 06:36.320 The config is going to be the uh config that we set in the YAML file. 06:36.680 --> 06:39.440 But then we'll have verbose being true as usual. 06:40.160 --> 06:41.400 Now what else do we want. 06:41.480 --> 06:48.120 We now want this super complex, super hard step of making sure that this agent has the power to run 06:48.160 --> 06:53.720 to, to, to execute code, which is as simple as allow code. 06:54.160 --> 06:54.440 Oops. 06:54.480 --> 06:57.000 Allow code execution equals true. 06:57.040 --> 06:57.920 There it is. 06:58.160 --> 06:59.680 Now it can execute code. 07:00.000 --> 07:11.420 There's this step of saying code execution Execution mode is safe, and now that ensures that it runs 07:11.420 --> 07:16.940 it within a Docker container so that it's not just running the code on your platform. 07:17.380 --> 07:19.580 Now we can say Max execution time. 07:19.580 --> 07:21.140 This is a good one to have in there. 07:21.140 --> 07:23.380 And we're going to say 30 30s. 07:23.420 --> 07:25.260 This is going to be a simple task. 07:25.620 --> 07:29.500 And max retry limit. 07:29.620 --> 07:31.700 It's funny that cursor prompted me wrong there. 07:31.740 --> 07:33.380 Let's give it five retries. 07:33.420 --> 07:36.380 It can have up to five times of trying this. 07:36.700 --> 07:39.460 And that is our coder agent. 07:39.700 --> 07:41.580 So now we have a task. 07:42.100 --> 07:46.180 And the task is going to be a coding task. 07:46.460 --> 07:52.620 And I rather suspect that, uh, we, uh, can just use what cursor does, but we can take out expected 07:52.620 --> 07:56.860 output because we already defined the expected output in the task itself. 07:56.860 --> 07:58.060 So it's not needed again. 07:58.180 --> 07:58.540 All right. 07:58.540 --> 07:59.060 So there we go. 07:59.100 --> 08:01.980 We have our agent and our task defined. 08:02.060 --> 08:06.220 Now we just have to do the crew function method just below.