WEBVTT 00:00.040 --> 00:02.200 Did I just say that I'd stop this pantomime? 00:02.360 --> 00:03.400 Well, I take that back. 00:03.400 --> 00:05.360 We'll do the pantomime one more time. 00:05.360 --> 00:09.840 I've just added in another response from the nice LM agent. 00:09.840 --> 00:12.400 So I'm going to have the LM agent, uh, respond to. 00:12.440 --> 00:13.040 Hi there. 00:13.080 --> 00:18.400 I'm going to have the disagreeable agent disagree, and then I'll put it back to, uh, to GPT for a 00:18.400 --> 00:21.400 minute to see how it handles the disagreement. 00:21.760 --> 00:23.680 Let's see how this little conversation goes. 00:23.840 --> 00:24.520 Hi, there. 00:24.840 --> 00:26.040 Uh, hello. 00:26.080 --> 00:27.320 How can I assist you today? 00:27.960 --> 00:29.840 Uh, you said hello. 00:29.840 --> 00:31.960 How can I assist you today, and I disagree. 00:33.240 --> 00:37.600 Then GPT for him, and he says, I appreciate your feedback. 00:37.800 --> 00:40.120 How would you prefer I greet you or assist you? 00:41.480 --> 00:46.080 So if you want to be cruel, you can keep this going and just have it continually disagreeing and see 00:46.080 --> 00:48.760 how, uh, how GPT four and mini handles that. 00:48.800 --> 00:54.840 And once you've had your fun, then you should do this, this, this stops the runtime and then closes 00:54.840 --> 01:00.210 the runtime, which the auto gen core people insist that we do before we start another. 01:00.210 --> 01:03.490 So we will we will do that and be and be good to it. 01:03.530 --> 01:03.890 Okay. 01:03.890 --> 01:08.010 But for one more example, and I'm afraid this time we're not going to have sort of true commercial 01:08.010 --> 01:13.770 examples because, uh, I think this is just one to show the platform at work, and you can take this 01:13.770 --> 01:15.770 away and figure out what you'd like to do with it. 01:15.770 --> 01:20.890 But for this example, uh, we're going to play rock, paper, scissors between a few agents to show, 01:20.930 --> 01:25.530 because I just want to illustrate the fact that this collaboration, this this interaction is what it's 01:25.530 --> 01:26.250 all about. 01:26.450 --> 01:33.450 So very simply, we have a player, one agent that is a subclass of rooted agent. 01:33.690 --> 01:41.730 And it, uh, it has a, um, um, open AI chat completion client as its underscored delegate. 01:41.730 --> 01:50.570 It has an assistant agent that uses that as the model client, and then it has, uh, handle my my message 01:50.570 --> 01:54.500 type that takes messages and and returns the response. 01:54.500 --> 02:02.500 So this is a super vanilla kind of LM routing that will just simply do respond to the message that is 02:02.500 --> 02:04.300 given in the message object. 02:04.500 --> 02:05.020 Okay. 02:05.060 --> 02:09.020 We got a player two which is exactly the same thing, but with one tiny difference. 02:09.020 --> 02:09.900 Can you spot it? 02:10.140 --> 02:11.860 Can you spot the difference between those two? 02:12.860 --> 02:19.540 Well, the difference is that this has a llama track completion client instead of OpenAI track submission 02:19.540 --> 02:19.780 client. 02:19.780 --> 02:21.100 So we're going to use a llama. 02:21.100 --> 02:22.780 We're going to use a local model. 02:23.060 --> 02:27.940 And we are going to see which model that we pick. 02:28.180 --> 02:33.660 For this one we're going to pick llama 3.2, the 3 billion variant of llama 3.2. 02:33.860 --> 02:38.220 So these are are a couple of agents. 02:38.620 --> 02:43.700 And now we're going to have a third one a third agent. 02:43.940 --> 02:47.740 And this is going to be uh rock paper scissors agent. 02:47.740 --> 02:52.150 So this one is going to have a little bit more of a it's going to have an instruction. 02:52.390 --> 02:54.390 You are playing rock paper scissors. 02:55.030 --> 02:55.630 First of all. 02:55.670 --> 02:57.590 So first of all it comes up with this instruction. 02:57.590 --> 02:59.030 You're playing rock paper scissors. 02:59.030 --> 03:01.710 Respond only with the one word one of the following. 03:01.750 --> 03:03.350 Rock, paper or scissors. 03:03.870 --> 03:05.830 And it puts that together into a message. 03:06.470 --> 03:13.630 It looks up the default ID, the default, uh, key for player one. 03:13.630 --> 03:18.750 That type, and it looks up the default for player two, that type. 03:19.030 --> 03:20.550 And it's looking them up. 03:20.550 --> 03:27.350 And it then as part of handling its message, it dispatches off a message to these other agents. 03:27.470 --> 03:30.430 So it's just choosing to to it's not choosing. 03:30.430 --> 03:31.430 It's like coded here. 03:31.430 --> 03:34.030 But but it is sending those messages off. 03:34.470 --> 03:41.110 And with what comes back uh, with the result, it then puts together a little piece of text called 03:41.110 --> 03:43.870 judgment, which says you are playing a game of rock, paper, scissors. 03:43.870 --> 03:44.000 The scissors. 03:44.000 --> 03:45.640 The players have made these choices. 03:45.680 --> 03:45.920 Sorry. 03:45.960 --> 03:46.800 You're judging again. 03:46.840 --> 03:47.720 Rock paper, scissors. 03:47.720 --> 03:49.240 Players have made these choices. 03:49.560 --> 03:49.920 Um. 03:49.920 --> 03:52.040 And then it ends with, uh. 03:52.520 --> 03:53.240 Who wins? 03:53.280 --> 03:54.080 Question mark. 03:54.200 --> 03:57.760 After slipping in the player one's choice and player two's choice. 03:57.920 --> 04:03.120 And then that gets dispatched off to my model, which is GPT four mini. 04:03.440 --> 04:05.880 And we return the response. 04:06.560 --> 04:11.680 So just to summarize, we have a total of three agents that we've defined here. 04:11.800 --> 04:16.840 Three of these like agent managers these agent wrappers, they're not they're not real agents. 04:16.840 --> 04:18.760 They delegate to real agents. 04:19.000 --> 04:23.120 One is called player one and it delegates to OpenAI to GPT four. 04:23.320 --> 04:27.920 Mini one is called player two and it delegates to llama 3.2. 04:28.080 --> 04:29.520 And then they don't have any prompts. 04:29.520 --> 04:31.600 They just do whatever instructions they're told. 04:32.040 --> 04:34.240 And then there's one called judge. 04:34.600 --> 04:41.360 And judge very clearly sends an instruction to each of player one and two saying pick rock or paper 04:41.400 --> 04:42.000 or scissors. 04:42.200 --> 04:47.240 And then with what comes back, it then judges them and returns the outcome. 04:47.520 --> 04:52.360 So there we have a simple setup for a game of rock, paper, scissors. 04:52.480 --> 04:55.200 And you can of course make this something more interesting. 04:55.200 --> 04:59.200 You can have it be be something that's that's entertaining or you can have it be something that's serious. 04:59.200 --> 05:00.080 It's commercial. 05:00.240 --> 05:06.360 Like when we're thinking next time in week six, we're going to be building like a financial trading 05:06.360 --> 05:07.000 setup. 05:07.000 --> 05:12.960 You could imagine this could be something where different agents can interact and can argue about something 05:12.960 --> 05:17.200 like whether or not a particular equity is a good investment or a poor investment. 05:17.200 --> 05:20.280 You could imagine that's the kind of interactions that could be going on. 05:20.280 --> 05:25.120 So this, this, whilst we're using this here for a superficial exercise, there's plenty of ways you 05:25.120 --> 05:31.280 can imagine anytime when you'd want a sort of autonomy and the ability for different agents to be interacting. 05:31.280 --> 05:35.520 You could build that kind of commercial logic into this kind of framework. 05:35.560 --> 05:37.450 And that is what Autogen core is for. 05:37.890 --> 05:39.730 Not for playing rock, paper, scissors. 05:39.730 --> 05:45.410 But anyways, rock paper scissors is the the cards you've been dealt and so you might as well play them. 05:45.450 --> 05:46.570 Let's give it a shot. 05:46.610 --> 05:53.250 And quite simply, we create a threaded agent a single threaded agent runtime, which is the the simple 05:53.290 --> 05:55.250 kind of local runtime. 05:55.410 --> 06:02.010 We register player one, we register player two and give it the ability to instantiate players one and 06:02.010 --> 06:02.370 two. 06:02.890 --> 06:08.970 And then we register the judge, the rock, paper, scissors, and and have everything ready to go. 06:08.970 --> 06:10.610 And we start our runtime. 06:10.650 --> 06:13.490 We've registered everything and this is where we do it. 06:13.530 --> 06:17.450 We are going to find our rock paper, scissors default agent. 06:17.930 --> 06:23.130 We are going to say go and set it going and let's see what happens. 06:23.130 --> 06:23.930 Off it runs. 06:23.970 --> 06:28.530 Agents are talking, stuff is happening and there we go. 06:29.010 --> 06:30.250 Player one said rock. 06:30.290 --> 06:31.650 Player two said scissors. 06:31.780 --> 06:34.060 Player one wins because rock beats scissors. 06:34.300 --> 06:34.780 Done. 06:35.420 --> 06:36.620 And there you have it. 06:36.620 --> 06:42.460 There you have agents being organized and interacting and agents being discovered. 06:42.460 --> 06:48.860 The rock, paper, scissors agent discovered these two agents by name by identifying them through the 06:48.860 --> 06:54.300 the the sending a message, uh, through the, uh, through the runtime. 06:54.540 --> 07:02.460 Um, and, uh, we saw agents interacting in a game of rock, paper, scissors and being judged managed 07:02.460 --> 07:03.740 by Autogen core. 07:03.780 --> 07:06.820 And so again, there are these two types standalone and distributed. 07:06.820 --> 07:08.100 Today we did standalone. 07:08.100 --> 07:09.700 Tomorrow we'll do distributed. 07:09.700 --> 07:12.500 And I just want to stress again that this is to give you a flavor. 07:12.660 --> 07:17.780 It's not necessarily as essential for for the building of agents that we're doing mostly on this course, 07:17.780 --> 07:22.620 but it's good for you to get this insight and to see it and compare it with the other offerings out 07:22.620 --> 07:23.020 there. 07:23.460 --> 07:25.540 So with that, that wraps up day three. 07:25.540 --> 07:27.500 And as I say, tomorrow we get to distributed. 07:27.500 --> 07:28.340 I'll see you then.