WEBVTT 00:00.120 --> 00:01.200 All right, we're nearly there. 00:01.200 --> 00:02.040 We're nearly there. 00:02.040 --> 00:07.240 I've just got one more thing to show you, which is world pie, which is quite short, which is the 00:07.240 --> 00:08.240 overall thing. 00:08.280 --> 00:09.240 Creator pie. 00:09.280 --> 00:10.160 You have to send it. 00:10.200 --> 00:15.160 It's an agent, and you have to send creator a message like agent one, agent two, agent three, and 00:15.160 --> 00:16.200 it will create them. 00:16.200 --> 00:17.720 So this is where it all comes together. 00:17.720 --> 00:19.440 World pie is not an agent. 00:19.440 --> 00:21.280 This is this is just a Python script. 00:21.280 --> 00:23.040 Everything else you've seen has been agents. 00:23.080 --> 00:27.440 This is just a Python script, but it uses async Python in quite an interesting way. 00:27.960 --> 00:35.720 So there is this async function, this coroutine create and message which takes a worker uh, and a 00:35.720 --> 00:42.720 creator id and an I, which is going to be the agent number going 12345. 00:43.040 --> 00:46.880 Um, this how many agents up here is a pretty important one to look at. 00:46.880 --> 00:49.920 This is how many of these it's going to kick off at the same time. 00:49.960 --> 00:54.680 And the other thing that's unsafe about this, of course, is cost that this is going to hit, uh, 00:54.720 --> 00:56.080 the real API's. 00:56.320 --> 01:02.190 And for me it's cost like a couple of cents when I run this for 20 agents, which is pretty cheap, 01:02.190 --> 01:03.190 all things considered. 01:03.190 --> 01:05.630 And obviously if you do it for like three agents, it costs nothing. 01:05.630 --> 01:14.110 But there's a risk there that we are allowing in theory, the, the this, um, the agent creator could 01:14.110 --> 01:16.230 change the model requirements. 01:16.230 --> 01:21.150 It could, it could decide to use, uh, an expensive model instead of GPT four or many. 01:21.190 --> 01:22.830 And so that's that's just something to watch for. 01:22.870 --> 01:24.190 It's very unlikely. 01:24.390 --> 01:26.950 Uh, but it's not like it's not impossible. 01:27.230 --> 01:35.310 Anyways, we've got 20 agents here, so we've got this async method creator and and message it, uh, 01:35.590 --> 01:43.990 it says it takes this worker and it sends a message, uh, to the agent number, whatever dot dot pi. 01:44.510 --> 01:49.270 Um, and then it writes the results to idea whatever dot markdown. 01:49.670 --> 01:52.870 Um, that's, that's what all this, this thing here does. 01:52.990 --> 01:53.470 Okay. 01:53.630 --> 02:00.460 So now we go to our main async method and, our main async function, our coroutine. 02:00.660 --> 02:08.580 So it creates a host a gRPC worker agent runtime host as we did in the notebook before. 02:08.620 --> 02:13.700 It starts the host, it creates a worker and it starts the worker. 02:14.340 --> 02:17.180 It then registers the creator. 02:17.460 --> 02:21.860 So it registers the creator from creator dot pi that was imported here. 02:22.020 --> 02:24.780 So that will launch a creator. 02:25.420 --> 02:28.540 And this is its id creator default. 02:29.500 --> 02:31.900 Then this is the async thing. 02:32.100 --> 02:38.060 So I don't want to have a loop and call this creating worker after worker. 02:38.340 --> 02:43.780 Sorry creating agent after agent because then it will be serial or will happen one after another. 02:43.780 --> 02:48.900 If I do like a wait and then I call this this this coroutine, then we'll be waiting there. 02:48.900 --> 02:50.340 It will create one file. 02:50.340 --> 02:51.780 It will send it a message. 02:51.780 --> 02:52.940 Back will come the answer. 02:52.940 --> 02:55.660 Maybe it'll send it somewhere else, and then it will go on to the next. 02:55.660 --> 02:58.540 And we'll be sitting here waiting and it won't be exciting. 02:58.900 --> 03:03.420 But what you can do with coroutines, as you hopefully remember from we briefly talked about this, 03:03.620 --> 03:06.860 is that you can get a whole stash of coroutines together. 03:06.860 --> 03:10.900 So I haven't got the word await here, which you need if this is actually going to run. 03:10.900 --> 03:20.180 I'm just gathering a whole list of coroutines and then I await them using Asyncio gather and Asyncio 03:20.340 --> 03:23.300 gather runs them all in parallel. 03:23.740 --> 03:26.980 And as hopefully you remember, it's not like multithreading. 03:26.980 --> 03:31.820 They're not going to be actually running on different threads, like with the CPU chopping between them, 03:31.820 --> 03:34.660 but rather they're running in the event loop. 03:34.660 --> 03:37.020 That's how Asyncio Python works. 03:37.020 --> 03:41.300 They run an event loop so that every time it's waiting on OpenAI, which means it's waiting on a network 03:41.300 --> 03:45.700 connection, another one can be running, and that means they all get to run at the same time. 03:45.700 --> 03:52.100 And as long as I stay within my my rate limits with OpenAI, we'll be able to run plenty in parallel. 03:52.580 --> 03:53.740 And so it does all of that. 03:53.740 --> 03:59.410 And then at the end it stops and any exceptions it will, it will print them and that's it. 03:59.410 --> 04:04.930 This is, by the way, how you run async Python from the command line or from a from a Python module. 04:04.930 --> 04:12.450 Your main file needs to do this, which uh, brings which initiates an async IO event loop and then 04:12.450 --> 04:15.930 runs your async method, your coroutine. 04:15.930 --> 04:18.010 So that is how it hangs together. 04:18.210 --> 04:20.370 And I do believe this is it. 04:20.570 --> 04:22.250 I don't think I've got anything else to show you. 04:22.250 --> 04:23.250 I think you've seen everything. 04:23.250 --> 04:29.530 You've seen agent, which is the prototype, the thing that gets cloned creator which does the cloning 04:29.570 --> 04:30.610 agent by agent. 04:30.610 --> 04:32.010 It's called one by one. 04:32.370 --> 04:34.610 And, uh, world Pi is. 04:34.610 --> 04:36.130 This is not an agent. 04:36.130 --> 04:41.650 It's just Python code which will orchestrate, uh, which will launch the whole process running. 04:41.770 --> 04:42.130 Okay. 04:42.170 --> 04:44.650 All that remains to do is to actually show you this thing working. 04:44.690 --> 04:45.810 Let's see if it does work. 04:45.850 --> 04:46.650 See you in a sec. 04:46.690 --> 04:48.290 Okay, let's give this a whirl. 04:48.290 --> 04:50.970 So we'll bring up a command line. 04:50.970 --> 04:52.250 We'll make it nice and big. 04:52.250 --> 04:53.530 So we see what's going on. 04:53.880 --> 05:00.440 We will go into the fifth directory and now we don't say Python and then a module, but rather we say 05:00.480 --> 05:08.480 you've run and then world.py is our script that will call creator 20 times for 20 different creations. 05:08.480 --> 05:13.880 You're expecting to see Python modules appearing over time as it builds the Python agents. 05:14.000 --> 05:17.680 And then you should see ideas starting to appear as well. 05:17.960 --> 05:20.920 After the conversations have happened, it might take a bit of time. 05:20.920 --> 05:21.720 We'll see what happens. 05:21.720 --> 05:22.800 Do you think it's going to work? 05:23.320 --> 05:26.480 Some problems, some some messages about threads and gRPC. 05:26.480 --> 05:28.800 I don't know what that's about, but it doesn't seem to stop it. 05:28.840 --> 05:33.760 And then off it goes and you will see that it's that it says agent 16 is live. 05:33.760 --> 05:36.400 Agent 17 is live, agent nine is live. 05:36.600 --> 05:38.040 Lots of things going on. 05:38.040 --> 05:43.520 And now you'll see messages going between agents and lots, lots of stuff here. 05:43.680 --> 05:45.560 Let's take a look at some of these agents. 05:45.560 --> 05:48.120 You'll also start to see ideas are appearing here. 05:48.320 --> 05:49.720 We look at these agents. 05:50.240 --> 05:56.430 Here is a tech savvy innovator in the field of entertainment, gaming, film and virtual reality. 05:56.830 --> 06:02.350 0.3 chance that he speaks to someone else or she and a 0.8 temperature. 06:02.750 --> 06:04.590 Let's try this one here. 06:05.110 --> 06:06.550 It's also another. 06:06.830 --> 06:09.870 This one is futuristic technology solutions. 06:09.870 --> 06:15.510 Very short system message 0.8 0.4 16. 06:15.550 --> 06:17.230 Finance industry this one. 06:17.230 --> 06:19.350 Innovate within the finance industry. 06:19.390 --> 06:22.470 Fintech, e-commerce and digital transformations. 06:22.710 --> 06:23.710 I wonder what this one will be. 06:23.750 --> 06:26.390 Agent 16 we should go and look at idea 16. 06:26.390 --> 06:28.470 It's already finished the 20 ideas. 06:28.470 --> 06:29.510 So it's very fast. 06:29.510 --> 06:31.110 And that's the great thing about this. 06:31.150 --> 06:33.830 We just created 20 agents. 06:33.830 --> 06:35.230 They just came up with ideas. 06:35.230 --> 06:36.990 Many of them will have had conversations. 06:36.990 --> 06:39.630 It all just happened while you were watching. 06:39.630 --> 06:44.470 And I do hope that this was worth your time investment for for this and that. 06:44.470 --> 06:50.230 You see the the as I say, the entertainment value and the educational value of going through this. 06:50.430 --> 06:54.390 Let's take a look at agent 16 and see the financial services ideas. 06:54.750 --> 06:59.150 Uh, let's open a preview and let's get rid of this. 07:00.830 --> 07:02.230 Oh, and the fact that we've got this. 07:02.230 --> 07:04.230 I absolutely love the vision behind fin, buddy. 07:04.230 --> 07:10.150 This thing at the top here tells us that this is the, uh, that this was sent to another agent for 07:10.150 --> 07:10.590 feedback. 07:10.590 --> 07:13.190 At least one agent added feedback to this. 07:13.190 --> 07:15.590 So this this went through a couple of people. 07:15.790 --> 07:18.150 So I absolutely love the vision behind Fin Buddy. 07:18.150 --> 07:22.430 It's a powerful concept, the potential to make significant impact on financial literacy and accessibility 07:22.470 --> 07:24.470 for underserved communities. 07:24.470 --> 07:27.070 Let's refine it and elevate the idea even further. 07:27.230 --> 07:28.510 Refined concepts. 07:28.510 --> 07:29.310 So this is great. 07:29.310 --> 07:35.110 It's an idea for applying Agentic AI to, uh, to financial services that's been refined by another 07:35.110 --> 07:39.830 agent, Fin Buddy, your AI powered financial companion for empowerment and inclusivity. 07:40.270 --> 07:47.710 Uh, so it's, uh, it's basically an education platform for, uh, for people who are under underserved 07:47.710 --> 07:49.070 in financial services. 07:49.340 --> 07:51.140 Contextual financial advisor. 07:51.140 --> 07:53.220 Localised community support hubs. 07:53.500 --> 08:00.100 Micro investment pools users can join collaborative investment pools collectively decide on small scale 08:00.100 --> 08:02.980 investments that is cool and engagement driven. 08:02.980 --> 08:06.620 Gamification with real rewards partner with local businesses. 08:06.820 --> 08:10.020 Very interesting, very interesting. 08:10.220 --> 08:11.380 Well, I might give this a read. 08:11.460 --> 08:11.820 My own. 08:11.860 --> 08:13.220 You might not hear from me again. 08:13.500 --> 08:22.580 Uh, so anyway, the the goal of this is to is to give you a live agentic platform where agents are 08:22.580 --> 08:28.900 created and agents collaborate and interact in an autonomous way in such an autonomous way that they 08:28.900 --> 08:31.700 were even created by another agent. 08:31.940 --> 08:37.060 And I hope you find this as satisfying as I do, and I hope it was worth it. 08:37.060 --> 08:41.740 I hope it was worth the half an hour investment to get here and do take a look through the code. 08:41.740 --> 08:42.580 If nothing else. 08:42.580 --> 08:49.610 As I say, it's it's definitely got some interesting, uh, Ideas built in there about how to use things 08:49.610 --> 08:53.290 like autogen for this kind of messaging between agents. 08:53.450 --> 08:59.130 And if nothing else, if you give it a run, if you feel bold, you will find a bunch of ideas. 08:59.170 --> 09:02.970 And maybe, maybe you'll be off on your yacht before too long. 09:03.650 --> 09:05.130 All right, I'll see you for the wrap up. 09:05.170 --> 09:09.450 Well, thank you for indulging me for the last half an hour of going through that project. 09:09.450 --> 09:12.290 I do hope that you at least found it somewhat interesting. 09:12.330 --> 09:12.930 Educational. 09:12.930 --> 09:18.650 And, yeah, it's opened your eyes a bit into some of what's possible at the frontier of Agentic AI, 09:18.690 --> 09:20.930 which is very much where Autogen is. 09:20.930 --> 09:28.370 It's very much an experimental and futuristic platform as we've seen ourselves first hand. 09:28.610 --> 09:32.730 So the challenge for you, the end of week challenge is an optional one, because this week has been 09:32.730 --> 09:35.250 experimental and a researchy week. 09:35.490 --> 09:41.890 But should you wish, uh, invest some more time in this, in this idea, uh, see if you can work with 09:41.890 --> 09:47.000 me on on the making it a bit more robust and something that it's safer for people to run, perhaps by 09:47.000 --> 09:51.880 dockerizing it, but in particular, something which I think would be fascinating would be to have it 09:51.880 --> 09:59.720 so that the creator is not only able to write new agents, but it's also able to write a new version 09:59.720 --> 10:00.720 of itself. 10:00.840 --> 10:07.440 It can create a new creator, modeling it off its own template of its own file, which it can also read. 10:07.800 --> 10:11.240 It might be interesting to turn that into a tool, rather than just having it be there in the Python 10:11.240 --> 10:16.960 code, have it be able to run a tool that reads in its own code, and it can then rewrite itself, making 10:16.960 --> 10:21.080 a replica something which itself is able to create new agents. 10:21.080 --> 10:25.520 And perhaps it could change its own logic slightly in some interesting way. 10:25.680 --> 10:27.760 So I think that would be fascinating. 10:27.800 --> 10:28.680 Super meta. 10:28.920 --> 10:31.200 It's something that creates creators. 10:31.640 --> 10:36.320 And then, I mean, if it creates itself, then in theory it can create creators of creators. 10:36.560 --> 10:37.440 Uh, so. 10:37.480 --> 10:37.840 Yeah. 10:38.160 --> 10:39.040 Mind blows. 10:39.280 --> 10:40.240 Uh, interesting. 10:40.240 --> 10:41.400 Interesting project. 10:41.440 --> 10:48.800 A great way importantly to be experimenting with interactions between agents in this kind of idea of 10:48.920 --> 10:54.920 an environment where the where the messaging, the communication and creation of agents is a separate 10:54.920 --> 11:00.880 concern from the implementation of the agents, which is, of course, the fundamental point of Agincourt. 11:01.360 --> 11:01.960 All right. 11:02.000 --> 11:05.960 And with that, finally, I'll stop yammering away about this. 11:06.000 --> 11:11.200 We're with the agent creator is done, and we are on to week six. 11:11.360 --> 11:15.680 The the fantastic, the exciting conclusion. 11:15.680 --> 11:18.320 And it is going to be a legendary week. 11:18.600 --> 11:25.280 Uh, I can't wait to show you everything that MCP is about and can offer, and I can't wait to return 11:25.280 --> 11:27.760 to OpenAI agents SDK. 11:28.000 --> 11:34.560 Still my favorite, even after enjoying all of the others, enjoying Crew and Landgraf a lot, and being 11:34.560 --> 11:39.160 generally, uh, entertained by by autogen and by the its forward thinking ness. 11:39.160 --> 11:42.920 But OpenAI agents SDK is next with MCP. 11:43.120 --> 11:44.080 I'll see you then.