WEBVTT 00:00.160 --> 00:08.240 Okay, so here we are back in Cursa going to the fifth directory where there are a few things for me 00:08.240 --> 00:09.240 to show you. 00:09.840 --> 00:13.080 So, uh, where to begin? 00:13.440 --> 00:18.400 I am going to begin by telling you about Agent Pi. 00:19.200 --> 00:22.480 So Agent Pi is like our template. 00:22.640 --> 00:29.160 This is a file that we are going to give another agent and ask it to use this as its model, as its 00:29.160 --> 00:34.480 example, and to take this as a template and to make other agents like this. 00:34.840 --> 00:37.480 So it is quite simply, it's like a prototype. 00:37.480 --> 00:39.680 It is our the prototype that will be used. 00:39.720 --> 00:43.000 We will clone it and make various versions of it or we won't. 00:43.040 --> 00:43.880 An agent will. 00:44.120 --> 00:46.160 So it does some some imports. 00:46.480 --> 00:47.840 It has a system message. 00:47.840 --> 00:49.280 You are a creative entrepreneur. 00:49.320 --> 00:55.120 Your task is to come up with a new business idea using a genetic I or refine an existing idea. 00:55.160 --> 00:59.970 Your personal interests are in these sectors and there's a couple of sectors you're drawn to ideas that 00:59.970 --> 01:01.170 involve disruption. 01:01.490 --> 01:04.530 You are less interested in ideas that are purely automation. 01:04.730 --> 01:07.410 You are optimistic, adventurous, and have risk appetite. 01:07.450 --> 01:08.290 You're imaginative. 01:08.330 --> 01:09.690 Sometimes too much so. 01:09.930 --> 01:10.850 Your weaknesses. 01:10.970 --> 01:12.970 You're not patient, can be impulsive. 01:13.250 --> 01:16.570 You should respond with your business ideas and engaging in a clear way. 01:16.970 --> 01:19.530 And you can see at the top here there's a comment. 01:19.610 --> 01:24.490 Change the system message to reflect the unique characteristics of this agent. 01:25.090 --> 01:31.530 And then here, this constant chances that I bounce idea off another is 0.5. 01:32.250 --> 01:35.570 And here it says you can also change the code to make the behavior different. 01:35.570 --> 01:38.250 But be careful to keep the method signatures the same. 01:38.930 --> 01:41.170 Okay, so far so good. 01:41.170 --> 01:43.450 And this agent that is going to be cloned. 01:43.490 --> 01:48.970 It has an init method which just simply sets GPT for a mini with a temperature. 01:48.970 --> 01:51.570 Make it a bit a bit random 0.7. 01:52.090 --> 01:57.130 And of course it creates a delegate which is an assistant using that model client. 01:57.130 --> 01:59.430 And with this system message. 01:59.990 --> 02:00.670 Okay. 02:01.190 --> 02:03.030 And here we have our handle. 02:03.070 --> 02:03.990 My message. 02:04.310 --> 02:05.950 Uh, and it's a message handler. 02:06.270 --> 02:14.350 And, uh, it's, um, got, um, uh, maybe it would be clearer if I just call this handle message 02:14.350 --> 02:15.630 rather than handle my message type. 02:15.670 --> 02:17.310 Fewer tokens is always good. 02:17.910 --> 02:21.990 Uh, so, um, it takes, uh, of course, the message. 02:22.030 --> 02:27.390 And now this message is using our messages object, the thing that we always create. 02:27.510 --> 02:32.030 And I have separated that out into its own package, because I want to put as little code in here as 02:32.030 --> 02:33.990 possible to avoid mistakes. 02:34.150 --> 02:36.270 So I've got our message object in here. 02:36.270 --> 02:37.550 We'll look at that in a minute. 02:38.030 --> 02:43.470 And it so it takes a messages message and it returns a messages dot message. 02:43.950 --> 02:47.310 It prints that this message with this type. 02:47.550 --> 02:49.910 Remember the type is like the agent's name. 02:50.070 --> 02:51.750 It's going to say received a message. 02:51.750 --> 02:54.390 So we will see as agents receive messages. 02:54.830 --> 02:56.710 It then makes a. 02:57.840 --> 03:07.680 text message, um, and, uh, it, uh, then it takes, takes this, this message that it was sent 03:08.160 --> 03:13.520 and it, uh, sends that to the, its underlying lm. 03:13.520 --> 03:15.920 So, so it sends on the message that it got. 03:16.200 --> 03:20.160 And that will of course, also have the system prompt as part of it because that's set right here. 03:20.680 --> 03:24.960 Uh, and it then says it waits until it gets back. 03:25.000 --> 03:31.280 The response it gets, it takes from the, the the response it makes that its business idea because 03:31.280 --> 03:33.120 it's asked to come up with a business idea. 03:33.120 --> 03:36.920 So what comes back from the delegate will be a business idea. 03:37.280 --> 03:43.720 And now it picks a random number and sees it basically figures out if in this case, this random, not 03:43.720 --> 03:46.200 random gives you a number between 0 and 1. 03:46.200 --> 03:51.840 And so if it's a 50% chance, if it's if it's a, there's a 50% chance it'll be less than 0.5. 03:52.120 --> 04:00.010 So uh, that this, this logic works If there's in a 50% chance it will use a little utility function, 04:00.010 --> 04:05.890 find recipient that's going to find some other recipient, and then it will say, here is my business 04:05.890 --> 04:06.530 idea. 04:06.530 --> 04:10.690 This might not be your specialty, but please refine it and make this business idea better. 04:10.890 --> 04:15.370 And it sends this business idea as a message to the recipient. 04:15.570 --> 04:19.130 So some randomly picked recipient, how's it going to pick a random recipient? 04:19.130 --> 04:26.690 We will see a random recipient is going to get the request to refine this business idea that this agent 04:26.690 --> 04:27.730 just came up with. 04:28.010 --> 04:32.410 And then and only if the probability is is is within this. 04:32.770 --> 04:37.170 And if so, then it will take the refined business idea. 04:37.170 --> 04:42.930 It will do self dot send message, which means it uses the runtime that it's associated with to send 04:42.930 --> 04:45.050 a message to the recipient. 04:45.050 --> 04:49.410 And it gets back the idea and that is what it then returns. 04:49.410 --> 04:56.350 So it either returns its own idea or there's some probability that it will return a refined version 04:56.350 --> 04:59.350 of its own idea refined by another agent. 05:00.190 --> 05:02.750 Okay, a lot to talk through. 05:02.750 --> 05:03.590 I hope that made sense. 05:03.590 --> 05:08.230 This is the template, the clone that's going to be used to create more agents. 05:08.270 --> 05:09.110 Let's see how. 05:09.150 --> 05:10.070 And I'm going to speed up. 05:10.070 --> 05:13.030 But but hang on in there because it's going to be great when you see this running. 05:13.030 --> 05:13.990 So wait for it. 05:14.070 --> 05:15.910 Don't don't don't lose me. 05:16.270 --> 05:18.590 Uh, I'll just quickly show you the messages. 05:18.590 --> 05:24.230 This is the messages package, and it has that same data class that we know of. 05:24.630 --> 05:30.070 Uh, and then it has some code to deal this, find recipient, find another agent that will message. 05:30.070 --> 05:31.270 And how does it do that? 05:31.310 --> 05:37.230 Well, it's super hacky, but basically, as the clones start appearing, the clones are going to be 05:37.230 --> 05:40.590 named agent one, agent two, agent three, agent four. 05:40.990 --> 05:45.950 And so it basically just looks in this directory to see what other agents have already existed. 05:45.950 --> 05:51.590 And if it finds one, it returns a randomly selected other agent, and there's some chance that we'll 05:51.590 --> 05:52.630 return the agent itself. 05:52.720 --> 05:57.320 The agent might talk to itself that maybe I should should, uh, solve I should should fix that and 05:57.320 --> 05:58.320 not let it talk to itself. 05:58.320 --> 06:00.200 But I thought, whatever, it doesn't matter. 06:00.200 --> 06:00.440 So. 06:00.440 --> 06:02.640 So right now it might talk to itself as well. 06:03.600 --> 06:04.120 Okay. 06:04.280 --> 06:07.840 So then the, uh, where life gets interesting is creator. 06:08.400 --> 06:12.320 Creator is the agent that creates and spawns agents. 06:12.600 --> 06:14.600 And I'll let you look through this. 06:14.600 --> 06:17.680 But basically the system message says just what you would think. 06:17.840 --> 06:20.720 You're an agent that can create other agents. 06:20.840 --> 06:28.080 You receive Python code and you create the user's Autogen core and Autogen agent chat. 06:28.280 --> 06:32.640 You should use this template to create a new agent with a unique system message. 06:32.920 --> 06:35.800 You can choose to keep the overall goal the same or change it. 06:35.800 --> 06:39.680 You can choose to take this agent in a completely different direction. 06:39.680 --> 06:42.640 This is the part that's both edgy and risky haha. 06:43.160 --> 06:49.000 The only requirement is that the class must be named agent, and it must inherit from from rooted agent 06:49.000 --> 06:50.800 and and it must have an init method. 06:51.380 --> 06:53.060 Respond only with Python code. 06:53.260 --> 06:54.900 So, look, this is this is cool. 06:54.900 --> 06:58.020 There's no doubt we're building something that we're saying. 06:58.020 --> 07:00.460 You have a free reign to do what you want. 07:00.500 --> 07:03.820 And I've only tried this with GPT four and mini, which is a bit safe with this stuff. 07:03.820 --> 07:09.620 It hasn't actually gone off the rails in an interesting way, and I am tempted to try it with Deep Sea. 07:09.660 --> 07:14.980 I'm slightly nervous with and with more creative agents, and maybe with Claude, because Claude is 07:15.020 --> 07:22.180 known for being more creative and see whether they look to include different logic in there to give 07:22.180 --> 07:24.540 this agent more smarts. 07:24.860 --> 07:29.780 But so far, whilst GPT four mini has definitely taken it in different directions, it hasn't done anything 07:29.820 --> 07:30.460 out there. 07:30.780 --> 07:34.780 Okay, so the user prompt says much, much the same thing. 07:34.940 --> 07:37.340 And then, uh, you can you can trace through it. 07:37.340 --> 07:39.940 But this is pretty simple and it prints what it's doing. 07:39.940 --> 07:49.380 But it basically opens this file, name it, it then uh, calls uh the delegate to, to take this action 07:49.380 --> 07:58.550 to basically given this python file in agent py, make a new version of it, and then and then it saves 07:58.550 --> 08:02.790 that as like agent one, agent two, agent three, agent four. 08:02.790 --> 08:09.270 So it was save it as each of the different, uh, agent numbers, it's, it's called with an agent number. 08:09.270 --> 08:12.110 And it will, uh, it will save it with that agent number. 08:12.550 --> 08:18.150 And then it does something that is super, uh, like, uh, out there. 08:18.630 --> 08:27.550 I use this Python feature, um, import lib to import the Python module that it just wrote. 08:27.550 --> 08:33.430 I import it here so that this creator agent imports that module, that it just wrote it effectively. 08:33.430 --> 08:42.470 If it just wrote a file called uh agent 5.py and the word agent five will have actually have come in 08:42.590 --> 08:43.870 in this message right here. 08:43.910 --> 08:46.630 So agent five is what will have been sent to the creator. 08:46.630 --> 08:49.120 And that's how it knows that that's the agent it's making. 08:49.280 --> 08:56.120 So in that case, what this will do is it will effectively be the same as saying import agent five pi. 08:56.320 --> 08:58.360 That's what that does dynamically. 08:58.360 --> 08:59.520 Programmatically. 08:59.560 --> 09:01.280 It imports it on the fly. 09:01.760 --> 09:08.840 And then I take that module that we've just imported on the on the fly and I register it, I call agent 09:08.840 --> 09:18.120 dot register with my runtime with the agent's name and using the lambda that will spawn a new instance, 09:18.160 --> 09:22.880 a factory method that will create a new instance of that agent on demand. 09:23.400 --> 09:29.760 And then it says it prints agent whatever is live and and it means not just the kind of live that we 09:29.760 --> 09:31.080 say when we deploy software. 09:31.440 --> 09:33.680 It means that the agent is running. 09:34.000 --> 09:39.160 Um, and it will then send a message to that agent saying, give me an idea. 09:39.400 --> 09:42.520 And that will then trigger that agent will then get that message. 09:42.520 --> 09:48.780 It will process it, it will work on an idea and potentially if the probability meets that criteria, 09:48.780 --> 09:54.220 it will then send on its idea to another agent that it will find in the directory, and ask that other 09:54.220 --> 09:56.900 agents to add feedback to it. 09:57.060 --> 10:00.780 And that's that's how it would all fit together, if you can believe it. 10:00.780 --> 10:02.380 That's what this does. 10:02.380 --> 10:03.660 So it's imaginative. 10:03.700 --> 10:05.180 It's a kind of crazy. 10:05.460 --> 10:11.220 And uh, as I say, the main the learning point here is that it is educational to step through it and 10:11.220 --> 10:11.900 see what's happening. 10:11.900 --> 10:14.540 And we've got something important about async Python coming up. 10:14.700 --> 10:20.780 Um, but this is definitely, uh, something to have you thinking about what it means to have agents 10:20.780 --> 10:26.700 interacting, messaging each other, and how this really demonstrates Autogen cause power as a pure 10:26.740 --> 10:28.900 agent messaging platform. 10:28.900 --> 10:37.060 We're able to to single out an agent, we are able just to get a recipient's agent ID and send it a 10:37.060 --> 10:37.740 message. 10:37.860 --> 10:44.300 So it allows for this kind of inter-process communication, uh, without without having to worry about 10:44.300 --> 10:46.660 the details of how that's actually strung together.