WEBVTT 00:00.200 --> 00:02.480 And now we move on to tasks. 00:02.480 --> 00:08.560 And here you'll see that there is a research task and a reporting task by default. 00:08.560 --> 00:11.800 But we're going to change this and we're going to have a few different tasks. 00:11.800 --> 00:13.760 We're going to have a task called propose. 00:14.320 --> 00:19.160 This is of course going to be the task which is about proposing a motion. 00:19.400 --> 00:22.280 And so the description what is this task. 00:22.720 --> 00:26.480 It's going to be your proposing the motion and then the motion that gets passed in. 00:26.480 --> 00:30.440 When we do that part come up with a clear argument in favor of the motion. 00:30.480 --> 00:31.760 Be very convincing. 00:32.320 --> 00:36.680 Then we're going to have an expected output, which is going to be quite simple. 00:36.680 --> 00:43.560 The expected output is going to be your clear argument in favor of the motion in a concise manner. 00:44.080 --> 00:44.960 The agent. 00:44.960 --> 00:48.360 So this is where you associate a task with an agent. 00:48.360 --> 00:51.760 And obviously this is going to be associated with the debater agent. 00:52.200 --> 00:55.720 And then there's one other thing we can do here which is an output file. 00:55.840 --> 01:02.120 We want to put this in a subdirectory output and we will call it propose.md. 01:02.120 --> 01:06.310 And so there you have our proposal task. 01:06.310 --> 01:11.510 So now we're going to have another task which I'm going to copy and paste. 01:11.510 --> 01:14.750 And this one is going to be oppose. 01:15.110 --> 01:18.470 This is the debater that's saying no description. 01:18.710 --> 01:27.470 Instead of proposing you are in opposition to the motion, come up with a clear argument against the 01:27.470 --> 01:28.390 motion. 01:28.510 --> 01:29.910 Looks like Kostas can help there. 01:29.910 --> 01:30.910 It does help. 01:30.950 --> 01:32.150 Be very convincing. 01:32.190 --> 01:34.430 Your clear argument against the motion. 01:34.670 --> 01:37.510 And that should be opposed. 01:37.550 --> 01:38.590 It's amazing, isn't it? 01:38.590 --> 01:45.750 So cursive does all the work for us, and now we have our final task, which is going to be called decide. 01:46.350 --> 01:49.230 And uh, the there's a little tip here. 01:49.270 --> 01:54.430 Something which is worth knowing is that you cannot call your tasks the same thing as you call your 01:54.430 --> 01:58.550 agents, or you will have a problem with conflicting names. 01:58.550 --> 02:02.110 So sometimes it's better to call this like proposed task, opposed task. 02:02.430 --> 02:03.270 But I've kept it short. 02:03.270 --> 02:06.790 But we couldn't, for example, call this task judge. 02:06.830 --> 02:09.950 Otherwise that would conflict with the agent that we called judge. 02:10.430 --> 02:11.070 All right. 02:11.310 --> 02:14.550 So description we will say review. 02:14.870 --> 02:15.310 Let me see. 02:15.350 --> 02:18.310 Now review the arguments. 02:18.910 --> 02:19.710 Here we go. 02:20.990 --> 02:25.710 Review the arguments presented by the debaters and decide which side is more convincing. 02:26.110 --> 02:29.950 And we'll just change this to delete all of this. 02:30.790 --> 02:31.870 We don't need any of that. 02:31.910 --> 02:38.150 We just need to say the expected output is your decision on which side is more convincing. 02:38.550 --> 02:41.590 And let's say and why. 02:42.230 --> 02:42.870 All right. 02:42.870 --> 02:47.390 And then the this should go please to a folder. 02:47.430 --> 02:52.950 Outputs and decide for the output. 02:52.990 --> 02:53.350 All right. 02:53.350 --> 02:54.110 That seems good. 02:54.150 --> 02:56.150 So there are our tasks. 02:56.830 --> 02:57.510 Okay. 02:57.870 --> 03:04.430 So the final step really is is the final steps is setting up pi and Main.py. 03:04.710 --> 03:06.180 And then we'll be ready to go. 03:06.900 --> 03:09.820 So this is the default module crew dot pi. 03:09.820 --> 03:13.980 And you can see it's got some stuff in here based on the standard scaffolding. 03:14.100 --> 03:16.140 It has created a class. 03:16.340 --> 03:19.660 And it's got this crew base decorator around it. 03:19.700 --> 03:23.780 And this class is named the same as the name of our project debate. 03:23.780 --> 03:27.540 So it's set that for us and it called it Debate Crew, which is exactly right. 03:27.780 --> 03:33.460 Now one of the things I dislike a bit is that it does generate all of this scaffolding code, all this 03:33.460 --> 03:37.660 standard code with lots of comments, and you can read some of the comments and follow the links. 03:37.660 --> 03:41.740 And it's got stuff like there's, there's ways that you can add in functions that get called at the 03:41.740 --> 03:43.740 beginning and end and stuff like that. 03:43.740 --> 03:49.180 But I do find that these comments are get in the way a bit, and I usually start by coming through and 03:49.180 --> 03:52.340 deleting everything in these just to keep it nice and clean. 03:52.340 --> 03:58.140 But what you'll see that it's done that is nice is that it brings in the agent's config and the tasks 03:58.140 --> 03:58.820 config. 03:58.980 --> 04:02.980 It just brings them in from the config folder right here. 04:02.980 --> 04:06.380 So those are set as variables for this class. 04:06.380 --> 04:09.460 And you can see how it refers directly to our configuration. 04:09.460 --> 04:13.740 And of course, that means that if you had different configuration files, then you could bring them 04:13.740 --> 04:15.020 in just like that. 04:15.060 --> 04:17.780 Then this is stuff about how you can bring in your own tools. 04:17.980 --> 04:21.740 And then this is where we set up our agents. 04:21.820 --> 04:24.060 So obviously we don't have an agent called researcher. 04:24.060 --> 04:26.140 We have an agent called debater. 04:26.500 --> 04:31.740 And the At agent decorator is telling crew that this is an agent. 04:31.980 --> 04:34.140 It's going to return an agent. 04:34.180 --> 04:37.460 The config we're going to want to change that to debater. 04:37.460 --> 04:38.820 And we'll leave it verbose. 04:38.980 --> 04:39.220 True. 04:39.260 --> 04:42.740 Which means that we'll get a nice sort of print as it's running with what's going on. 04:43.260 --> 04:46.820 And then we'll have another agent and remember the name of this guy. 04:47.060 --> 04:52.220 Well, even if you don't, luckily cursor does and it's filled it in in both places. 04:52.220 --> 04:52.980 There we go. 04:53.020 --> 04:55.180 The judge is our other agent. 04:55.740 --> 04:56.500 All right. 04:56.940 --> 05:00.340 And now I'm getting rid of the comments about tasks. 05:00.340 --> 05:04.940 But it tells you about how you can do things like track outputs and dependencies and the like. 05:04.940 --> 05:07.580 But we're going to go on and define our our tasks. 05:07.620 --> 05:10.410 They have a again the decorator task. 05:10.650 --> 05:13.450 And now we're going to have a task called propose. 05:14.250 --> 05:15.530 And there we go. 05:15.570 --> 05:16.770 Cursor fills it in. 05:16.810 --> 05:19.250 Now it has an output file listed there. 05:19.250 --> 05:23.650 But we don't need that because it's already specified in our config file. 05:23.650 --> 05:25.130 So that's not actually needed. 05:25.450 --> 05:29.890 And you know now that I look at this I do I have to say I think this would be cleaner if we have this 05:29.890 --> 05:32.890 all on one line like this, we don't need to be on multiple lines. 05:33.010 --> 05:37.690 And now you can see how very simple this whole function is going to look. 05:37.730 --> 05:39.330 And look cursor is going to do it for us. 05:39.330 --> 05:40.810 And then we can just delete that. 05:41.250 --> 05:45.490 Sorry I'm tidying this all up when you see how much better it is when this is in JupyterLab. 05:45.490 --> 05:47.290 And you don't have to watch me doing all this. 05:47.650 --> 05:48.690 So don't hate me. 05:48.690 --> 05:52.850 I'll be quick, but I think it's nicer if we have it looking nice and sharp like that. 05:52.850 --> 05:55.290 You see how nice and simple and clean this is? 05:55.450 --> 06:01.090 And that's the benefit of using the the config that we don't need to, to have too much here. 06:01.090 --> 06:03.490 And then the other task is oppose. 06:04.250 --> 06:06.970 And obviously it brings in the config for oppose. 06:06.970 --> 06:09.650 And then the final one is decide. 06:09.930 --> 06:11.120 And there we go. 06:11.160 --> 06:13.520 Thank you cursor for filling that in for us. 06:13.920 --> 06:18.080 So those are our two agents and three tasks. 06:18.080 --> 06:20.640 And we are down to the final section here. 06:21.280 --> 06:21.920 Okay. 06:22.160 --> 06:25.920 So now we now create this the crew. 06:26.360 --> 06:29.040 And this is, this is in the the function crew. 06:29.040 --> 06:30.640 And it has the decorator crew. 06:31.000 --> 06:36.120 And I'm going to take out the, the piece there which we will be coming to later. 06:36.160 --> 06:39.520 So we just simply have to create an instance of crew. 06:39.760 --> 06:47.680 We populate the agents with self dot agents and as cry helpfully tells us, this is automatically created 06:47.680 --> 06:49.880 by the agent decorator. 06:50.440 --> 06:53.440 We populate it with the tasks that are similarly created. 06:53.640 --> 06:59.560 This is where we choose to be sequential rather than hierarchical in our process, and that we want 06:59.560 --> 07:00.920 to be verbose. 07:00.960 --> 07:04.880 And then there's some stuff about being hierarchical if you want to. 07:05.160 --> 07:10.600 And uh, look, if you want to use that nice little comment from the crew team in there. 07:11.040 --> 07:11.640 Okay. 07:11.760 --> 07:14.000 that's it for our crew object. 07:14.000 --> 07:15.280 We're almost finished. 07:15.560 --> 07:16.560 That was our crew module. 07:16.560 --> 07:20.520 We're now going to the main module where we finish things off. 07:21.040 --> 07:26.080 So this this is the again there's all of this stuff that you can read. 07:26.480 --> 07:29.840 This main file is intended to run your crew locally. 07:29.840 --> 07:32.920 So refrain from adding unnecessary logic. 07:33.240 --> 07:33.920 Gotcha. 07:34.320 --> 07:41.280 So inputs when we're running the crew, this is where we choose those template values that we put in 07:41.280 --> 07:42.320 our YAML file. 07:42.320 --> 07:44.400 This is where we choose what we want them to be. 07:44.440 --> 07:50.960 And so what we want them to be here is we want to have a motion and we want to give ourselves a nice 07:50.960 --> 07:51.320 motion. 07:51.320 --> 07:52.680 We don't we don't have this. 07:52.720 --> 07:54.840 We only have the one field motion. 07:55.160 --> 07:56.720 And what motion? 07:56.720 --> 07:59.520 What do we think that we would like them to debate? 07:59.600 --> 08:08.960 Let's have the motion is there should be a that was that was where I was heading with this uh, cursor. 08:09.240 --> 08:13.710 I must be super predictable, but it does seem like a good one to have our eyes A's debate. 08:13.750 --> 08:14.830 Let's not say there should be. 08:14.830 --> 08:22.030 Let's say there needs to be strict laws to regulate llms. 08:22.470 --> 08:22.990 There we go. 08:23.030 --> 08:25.870 That's a nice meaty topic for them to debate. 08:26.190 --> 08:29.230 And then we've got our stuff down here. 08:29.230 --> 08:30.310 We will come back to this stuff. 08:30.310 --> 08:31.870 We don't care about any of that. 08:32.030 --> 08:34.310 That is all it should take. 08:34.310 --> 08:39.550 We should now be ready to run our first crew just based on that. 08:40.190 --> 08:46.270 Actually, just before we run, let me make another little change that in this main module. 08:46.270 --> 08:51.710 I'm just going to make sure we print the result to the screen just to be satisfying. 08:51.750 --> 08:53.350 It will be saved to a file as well. 08:53.670 --> 08:55.590 Results equals debate. 08:55.710 --> 08:58.750 And then we will print result dot raw. 08:58.870 --> 09:05.030 That will print the raw output from the final task sent to the final agent. 09:05.630 --> 09:11.510 And then maybe one other little change I'll make is I realize we're right now sending both the debater 09:11.510 --> 09:15.950 and the judge to be OpenAI, but it would be more interesting to switch this one up. 09:15.990 --> 09:16.950 Let's let's flip it. 09:16.950 --> 09:18.030 As cursor suggests. 09:18.070 --> 09:21.270 Let's use Claude 37 sonnet latest. 09:21.310 --> 09:27.950 The latest Claude 3.7 model so that we have anthropic judging OpenAI's debates. 09:27.950 --> 09:34.110 And this also means that we'll be using anthropic this time, which was harder for us last time. 09:34.550 --> 09:37.670 Okay, so with that now I think we should be ready. 09:37.710 --> 09:39.790 We can bring up a terminal. 09:40.030 --> 09:40.830 Here it is. 09:41.070 --> 09:47.350 We will want to change into the directory for crew and then the directory for debate. 09:47.590 --> 09:49.590 And now we simply type crew. 09:49.590 --> 09:52.870 I run to kick off the run. 09:52.870 --> 09:56.630 And the first time you do this, it's going to do some UV stuff to build the environment. 09:56.630 --> 09:58.870 But I've already done it, so it's running right away. 09:59.030 --> 10:01.150 You can see it's executing tasks. 10:01.390 --> 10:05.430 The debater, each of the agents is running it's debate right now. 10:05.430 --> 10:10.630 It's the the debater for the proposer has already said the against is now debating. 10:10.950 --> 10:15.830 And now we're deciding on the winner of the debate based on the arguments. 10:15.830 --> 10:16.900 It is anthropic. 10:16.940 --> 10:17.860 That is thinking. 10:18.340 --> 10:24.860 And with any luck, we will soon see whether anthropic decides for or against the motion. 10:25.260 --> 10:25.900 Here we go. 10:25.980 --> 10:26.820 Let me see. 10:27.500 --> 10:33.220 In evaluating, I found the arguments in favor of the motion to be more convincing. 10:33.420 --> 10:39.100 So the view is that there should be strict laws to regulate large language models, such as the very 10:39.100 --> 10:41.420 one that is making the judgment here. 10:41.780 --> 10:42.380 Okay. 10:42.420 --> 10:43.020 So that ran. 10:43.020 --> 10:46.180 And look, that's quite a wordy response that came from anthropic. 10:46.220 --> 10:50.500 There you get to see in here the trace of the thinking that happened. 10:50.500 --> 10:53.700 And you'll see that an output directory has been created. 10:53.700 --> 10:59.220 And if we open this up and if I close this terminal so it's not in the way we can go into the output, 10:59.220 --> 11:05.580 and you'll see that there is weighty argument in favor of LM law regulation that you can read about 11:05.580 --> 11:12.740 about the ethical, safety and social challenges necessitating strict regulatory frameworks and then 11:12.740 --> 11:15.300 the opposing thinking as well. 11:15.780 --> 11:19.020 It's no doubt going to be about hindering creativity. 11:19.260 --> 11:19.900 Yes. 11:19.940 --> 11:20.580 Very good. 11:20.620 --> 11:23.540 And I'm sure there's some compelling arguments there, too. 11:23.780 --> 11:28.900 And then the decision is the thing that we already saw printed a moment ago that came from anthropic. 11:29.100 --> 11:33.460 And so that is our first experiment into the world of crew. 11:33.500 --> 11:35.180 We set up our yamls. 11:35.180 --> 11:40.820 We set up our overall our our module for crew, the crew module. 11:40.980 --> 11:47.620 And then the main.py was where we set the motion, which was the thing that was templated in the various 11:47.620 --> 11:48.260 YAML documents. 11:48.260 --> 11:57.580 And we actually ran our debate with debate crew kickoff passing in the inputs this dictionary of the 11:57.940 --> 12:00.820 templated keys with their values. 12:01.300 --> 12:07.660 So I hope you enjoyed our first foray into crew AI, and I very much encourage you to do the same thing. 12:07.700 --> 12:12.740 Well, of course you will actually see this debate project in there, but you can go and make a second 12:12.740 --> 12:16.860 one or just run this by typing query run and get a handle. 12:16.900 --> 12:20.140 Get a good sense for how crew AI works.