WEBVTT 1 00:00:00.840 --> 00:00:03.780 Hi, and welcome to this AI in C# video 2 00:00:03.780 --> 00:00:05.580 on the Microsoft Agent Framework. 3 00:00:06.180 --> 00:00:08.760 Today we're going to look into the A2A 4 00:00:08.760 --> 00:00:12.020 protocol, which is something Google have made, where 5 00:00:12.020 --> 00:00:14.880 we have multiple agents talking to each other. 6 00:00:15.880 --> 00:00:18.320 Before we jump into the code, let's quickly 7 00:00:18.320 --> 00:00:20.500 talk a little about what this is. 8 00:00:22.060 --> 00:00:25.740 And if you are familiar with MCPs, they 9 00:00:25.740 --> 00:00:27.840 are very similar to... 10 00:00:27.840 --> 00:00:33.180 A2A is very similar to MCPs, but there 11 00:00:33.180 --> 00:00:35.820 is a bit more of intelligence in here. 12 00:00:37.220 --> 00:00:40.120 Also, if you have seen my video about 13 00:00:40.120 --> 00:00:43.300 an agent calling another agent as a tool, 14 00:00:43.520 --> 00:00:45.920 that would be beneficial in order to understand 15 00:00:45.920 --> 00:00:51.540 this, because that is essentially what's happening, but 16 00:00:51.540 --> 00:00:55.000 where the other agent is just remote. 17 00:00:55.800 --> 00:00:57.520 But what happens is you have a local 18 00:00:57.520 --> 00:01:00.900 app, and you know a URL for another 19 00:01:00.900 --> 00:01:01.620 agent. 20 00:01:03.060 --> 00:01:05.440 That is something called an agent card. 21 00:01:06.280 --> 00:01:09.680 So the remote agent have some kind of 22 00:01:09.680 --> 00:01:10.180 skills. 23 00:01:10.660 --> 00:01:12.860 It's normally tools, but it could also be 24 00:01:12.860 --> 00:01:15.920 that it's very good at something, or use 25 00:01:15.920 --> 00:01:18.360 a special model, or something like that. 26 00:01:19.460 --> 00:01:23.440 So it's just a normal agent, and it 27 00:01:23.440 --> 00:01:26.460 has these tools, and it shows it to, 28 00:01:26.880 --> 00:01:29.120 or exposes that to the local app as 29 00:01:29.120 --> 00:01:31.480 what is called an agent card, where it 30 00:01:31.480 --> 00:01:34.500 tells, I'm this, I'm good at this, you 31 00:01:34.500 --> 00:01:36.240 can use me for this, and so on. 32 00:01:36.900 --> 00:01:38.580 It's sort of like a business card. 33 00:01:39.100 --> 00:01:41.600 If you met a person where you could 34 00:01:41.600 --> 00:01:43.760 find information about what they can do, and 35 00:01:43.760 --> 00:01:46.180 how they can help, and what to talk 36 00:01:46.180 --> 00:01:46.520 with them. 37 00:01:47.120 --> 00:01:49.700 So a little like MCPs list tools. 38 00:01:51.740 --> 00:01:55.080 And then the local app can use these 39 00:01:55.080 --> 00:01:59.620 tools over here through this connection, as if 40 00:01:59.620 --> 00:02:02.810 they were connected directly to it itself. 41 00:02:04.100 --> 00:02:06.000 And what we're about to see in code 42 00:02:06.000 --> 00:02:08.300 is of course happening all on my machine, 43 00:02:08.720 --> 00:02:11.560 but the important thing to remember is that 44 00:02:11.560 --> 00:02:14.400 these two machines can be in different parts 45 00:02:14.400 --> 00:02:17.480 of the world, because they're just talking via 46 00:02:17.480 --> 00:02:18.240 a URL. 47 00:02:18.240 --> 00:02:21.840 So let's jump in and see it in 48 00:02:21.840 --> 00:02:22.140 action. 49 00:02:27.160 --> 00:02:30.740 So in the sample repo, we have agent 50 00:02:30.740 --> 00:02:32.640 to agent here, and we have a server 51 00:02:32.640 --> 00:02:33.540 and a client. 52 00:02:35.500 --> 00:02:38.600 And first I will just show it run, 53 00:02:38.980 --> 00:02:40.580 and then we will go through the code. 54 00:02:41.220 --> 00:02:42.940 And in order to run it, I need 55 00:02:42.940 --> 00:02:45.560 to run both systems, and for that reason 56 00:02:45.560 --> 00:02:48.680 I went into the setup configuration here, and 57 00:02:48.680 --> 00:02:51.900 set up that these two start at the 58 00:02:51.900 --> 00:02:52.460 same time. 59 00:02:55.140 --> 00:02:58.540 So if I run it, and arrange a 60 00:02:58.540 --> 00:03:00.420 bunch of screens here so we can see 61 00:03:00.420 --> 00:03:07.960 it, then we have the server here, and 62 00:03:07.960 --> 00:03:09.340 here we have the client. 63 00:03:11.260 --> 00:03:14.360 And if I ask it, in this case 64 00:03:14.360 --> 00:03:16.820 it knows about some files on my machine. 65 00:03:17.660 --> 00:03:21.160 So if we find the files up here, 66 00:03:21.820 --> 00:03:23.580 you can see there's 10 files in this 67 00:03:23.580 --> 00:03:28.140 folder, so I can ask how many files. 68 00:03:29.560 --> 00:03:32.700 These files are the same as you saw 69 00:03:32.700 --> 00:03:35.580 in my function calling scenario, and it can 70 00:03:35.580 --> 00:03:38.260 see there's 10 files. 71 00:03:38.260 --> 00:03:40.260 And we can go in and say delete 72 00:03:40.260 --> 00:03:49.570 the orange.txt file. 73 00:03:53.040 --> 00:03:54.960 If we do that, it removes it over 74 00:03:54.960 --> 00:03:55.240 here. 75 00:03:55.560 --> 00:03:58.280 So we are about to get information, but 76 00:03:58.280 --> 00:03:59.260 also actions. 77 00:04:00.160 --> 00:04:03.740 And over here, things are happening on behalf 78 00:04:03.740 --> 00:04:05.820 of this, and had this just been one 79 00:04:05.820 --> 00:04:10.520 agent with the skills, meaning had it been 80 00:04:12.210 --> 00:04:16.010 down here in features, our normal function calling, 81 00:04:16.810 --> 00:04:19.450 it would just have been a tool that 82 00:04:19.450 --> 00:04:22.930 we gave to a system, and that would 83 00:04:22.930 --> 00:04:23.370 be it. 84 00:04:24.250 --> 00:04:27.810 But now it's happening over a URL, and 85 00:04:27.810 --> 00:04:30.670 again, the important part is to think the 86 00:04:30.670 --> 00:04:33.830 server here could be in a completely different 87 00:04:33.830 --> 00:04:35.750 part of the world while the client would 88 00:04:35.750 --> 00:04:36.570 be on my machine. 89 00:04:37.340 --> 00:04:39.330 But let's see how it's implemented. 90 00:04:42.850 --> 00:04:45.590 So let's first look at the server, because 91 00:04:45.590 --> 00:04:48.110 without the server, we can't call the client. 92 00:04:51.300 --> 00:04:53.960 So first up, we are using a new 93 00:04:53.960 --> 00:04:57.620 NuGet packet we haven't used before, and we're 94 00:04:57.620 --> 00:05:00.760 using AI hosting A2A ASP.NET Core. 95 00:05:01.960 --> 00:05:03.920 And since this is a console app, we're 96 00:05:03.920 --> 00:05:07.220 still making it work as ASP.NET Core, 97 00:05:07.340 --> 00:05:09.740 so we are actually exposing a URL. 98 00:05:10.460 --> 00:05:15.540 So when I run it and show you 99 00:05:15.540 --> 00:05:17.880 the server, you can see we get this 100 00:05:17.880 --> 00:05:20.320 URL up here, just like if it had 101 00:05:20.320 --> 00:05:23.120 been a web API or whatever. 102 00:05:23.920 --> 00:05:26.000 And as well, we also get something called 103 00:05:26.000 --> 00:05:29.100 well-known agent card JSON. 104 00:05:31.610 --> 00:05:35.670 And the way it works is we go 105 00:05:35.670 --> 00:05:38.710 to our server, and we do the normal 106 00:05:38.710 --> 00:05:42.250 business part as we do with any agents 107 00:05:42.250 --> 00:05:43.150 we have done so far. 108 00:05:43.150 --> 00:05:45.470 We take the Azure OpenAI client. 109 00:05:46.130 --> 00:05:48.850 In this case, we are grabbing the different 110 00:05:48.850 --> 00:05:49.330 tools. 111 00:05:49.550 --> 00:05:51.630 Again, exactly the same as we did down 112 00:05:51.630 --> 00:05:54.110 here in the tool calling advanced, which there's 113 00:05:54.110 --> 00:05:55.070 a separate video on. 114 00:05:55.790 --> 00:05:58.230 And then we're just making an agent just 115 00:05:58.230 --> 00:06:02.530 like we do, and we give them these 116 00:06:02.530 --> 00:06:04.890 tools, and we set up a little function 117 00:06:04.890 --> 00:06:07.170 calling middleware so we can see that the 118 00:06:07.170 --> 00:06:08.570 tools are being called. 119 00:06:09.310 --> 00:06:11.810 You don't technically need tools here. 120 00:06:11.810 --> 00:06:13.490 It could be that it was very good 121 00:06:13.490 --> 00:06:15.470 at understanding something. 122 00:06:15.710 --> 00:06:17.810 It was hooked up to a vector store. 123 00:06:18.010 --> 00:06:21.570 It could be something that you would like 124 00:06:21.570 --> 00:06:24.410 to expose to the world through something more 125 00:06:24.410 --> 00:06:27.930 intelligent than just bringing raw data back. 126 00:06:30.620 --> 00:06:32.140 But then we begin with the agent part, 127 00:06:32.540 --> 00:06:35.180 and the agent part not really here. 128 00:06:35.340 --> 00:06:38.460 We're just setting up that ASP.NET Core 129 00:06:38.460 --> 00:06:40.300 can have a URL. 130 00:06:40.980 --> 00:06:43.360 And then we're building something called an agent 131 00:06:43.360 --> 00:06:43.720 card. 132 00:06:44.040 --> 00:06:45.740 Again, the agent's business card. 133 00:06:46.200 --> 00:06:48.500 So this is what is being exchanged between 134 00:06:48.500 --> 00:06:50.180 the two, the client and the server. 135 00:06:50.980 --> 00:06:54.320 So we tell, hey, I'm an agent. 136 00:06:54.540 --> 00:06:55.680 I'm called files agent. 137 00:06:55.940 --> 00:06:58.080 I can handle requests related to files. 138 00:06:58.860 --> 00:07:03.720 I'm version 1.0. My input modes are 139 00:07:03.720 --> 00:07:05.820 text, and my output modes are text. 140 00:07:06.820 --> 00:07:09.940 Again, this protocol has lots and lots of 141 00:07:09.940 --> 00:07:12.740 other features, authentication and so on, but we're 142 00:07:12.740 --> 00:07:14.920 just diving into the basics here. 143 00:07:15.360 --> 00:07:16.320 We're not streaming. 144 00:07:16.680 --> 00:07:19.500 We're just giving data fullback, and we're not 145 00:07:19.500 --> 00:07:23.540 sending any notifications because something can also be 146 00:07:23.540 --> 00:07:25.560 long running and such things like that. 147 00:07:26.880 --> 00:07:29.500 But we have one skill, which we call 148 00:07:29.500 --> 00:07:32.940 an agent skill, which is just that we 149 00:07:32.940 --> 00:07:37.940 are file experts, and we can do the 150 00:07:37.940 --> 00:07:39.020 questions about the files. 151 00:07:40.060 --> 00:07:44.120 We can give some tags so different agent 152 00:07:44.120 --> 00:07:46.840 cards can be categorised and stuff, and we 153 00:07:46.840 --> 00:07:50.920 give a small example on how we work. 154 00:07:52.040 --> 00:07:53.840 And then we give the URL that we 155 00:07:53.840 --> 00:07:54.840 need to be. 156 00:07:55.860 --> 00:07:59.060 And the tools themselves are fairly simple. 157 00:07:59.200 --> 00:08:02.980 It's just create folder, delete folder, just normal 158 00:08:02.980 --> 00:08:04.260 C-sharp code. 159 00:08:04.520 --> 00:08:06.840 And again, if you want to know about 160 00:08:06.840 --> 00:08:09.940 this, see the tool calling advanced for full 161 00:08:09.940 --> 00:08:10.340 details. 162 00:08:11.460 --> 00:08:13.460 The final thing we do is just like 163 00:08:13.460 --> 00:08:16.260 a map get, map post in a minimal 164 00:08:16.260 --> 00:08:16.700 API. 165 00:08:16.900 --> 00:08:19.240 We do a map A to A, and 166 00:08:19.240 --> 00:08:21.640 we tell this is the agent we are 167 00:08:21.640 --> 00:08:22.140 exposing. 168 00:08:22.740 --> 00:08:26.280 Here's the part we are exposing it, meaning 169 00:08:26.280 --> 00:08:27.060 just the root. 170 00:08:27.920 --> 00:08:31.100 Here is the agent card, and here is 171 00:08:31.100 --> 00:08:35.600 the well-known agent card, so it knows 172 00:08:35.600 --> 00:08:39.400 this information when it needs to work. 173 00:08:40.320 --> 00:08:42.299 And then we are just running, meaning we 174 00:08:42.299 --> 00:08:44.560 are starting the little mini web server in 175 00:08:44.560 --> 00:08:45.940 our console app here. 176 00:08:48.420 --> 00:08:53.120 On the client side, it is fairly simple. 177 00:08:54.220 --> 00:08:56.800 In here, we instead use the A to 178 00:08:56.800 --> 00:08:58.740 A instead of hosting A2A, because 179 00:08:58.740 --> 00:08:59.740 we are the consumer. 180 00:09:01.050 --> 00:09:05.470 And we are going to start, because I'm 181 00:09:05.470 --> 00:09:08.210 starting at the same time, I'm just waiting 182 00:09:08.210 --> 00:09:10.190 one second just to be sure that the 183 00:09:10.190 --> 00:09:14.070 server is ready to begin to be consumed 184 00:09:14.070 --> 00:09:14.750 by the client. 185 00:09:16.610 --> 00:09:19.370 Then we do something called an A to 186 00:09:19.370 --> 00:09:22.090 A card resolver where we take the URL. 187 00:09:22.610 --> 00:09:24.090 Again, this could be a remote URL. 188 00:09:25.270 --> 00:09:29.690 And we get that, and the agent framework 189 00:09:29.690 --> 00:09:32.350 just gives us and give me that as 190 00:09:32.350 --> 00:09:34.210 an AI agent, and we're done. 191 00:09:34.790 --> 00:09:36.550 Because then we have an AI agent, we 192 00:09:36.550 --> 00:09:40.210 could begin to use that as such. 193 00:09:40.830 --> 00:09:45.570 But what we are doing is we're giving 194 00:09:45.570 --> 00:09:48.290 it as a tool to an agent that 195 00:09:48.290 --> 00:09:50.010 we control over on our side. 196 00:09:51.070 --> 00:09:53.530 That agent is just like normal. 197 00:09:53.530 --> 00:10:00.610 We give it instructions like normal, and we 198 00:10:00.610 --> 00:10:02.010 just make a chat loop. 199 00:10:02.330 --> 00:10:05.690 So nothing over here is different other than 200 00:10:05.690 --> 00:10:09.210 this, which we then give as a tool. 201 00:10:09.790 --> 00:10:13.310 Beyond that, it's just normal agent framework like 202 00:10:13.310 --> 00:10:14.150 we normally do. 203 00:10:16.090 --> 00:10:18.010 So there's not much to this. 204 00:10:18.250 --> 00:10:21.550 It's just this exchange between the two that 205 00:10:21.550 --> 00:10:26.690 this running in one process can get an 206 00:10:26.690 --> 00:10:28.970 agent that can be used as skills in 207 00:10:28.970 --> 00:10:29.710 another process. 208 00:10:33.060 --> 00:10:35.500 So when we again look at it here, 209 00:10:35.940 --> 00:10:41.800 and the server, let me try to get 210 00:10:41.800 --> 00:10:44.480 them a little more side-by-side here, 211 00:10:47.640 --> 00:10:56.200 this and the files, then we could, for 212 00:10:56.200 --> 00:11:04.540 example, say group the files into folders 213 00:11:06.430 --> 00:11:12.060 based on their colour. 214 00:11:16.520 --> 00:11:18.160 So now it begins to work over here. 215 00:11:18.540 --> 00:11:20.460 It found the root folder. 216 00:11:21.180 --> 00:11:25.910 It found the content of each file and 217 00:11:25.910 --> 00:11:30.210 so on, and it wants us to confirm. 218 00:11:33.040 --> 00:11:34.020 We want to proceed. 219 00:11:34.020 --> 00:11:35.640 But that is just AI. 220 00:11:36.220 --> 00:11:38.280 Had it been in the same root, then 221 00:11:38.280 --> 00:11:39.940 the same thing would have happened. 222 00:11:40.780 --> 00:11:43.260 But right now, it's doing its work and 223 00:11:43.260 --> 00:11:47.420 putting in banana and lemon and pear into 224 00:11:47.420 --> 00:11:49.060 the yellow and so on. 225 00:11:49.400 --> 00:11:52.140 So it's just AI, and again, it's just 226 00:11:52.140 --> 00:11:56.320 as if we had one agent without two, 227 00:11:57.020 --> 00:11:59.140 but the cool thing is, of course, that 228 00:12:00.450 --> 00:12:03.990 these are in two different locations in the 229 00:12:03.990 --> 00:12:04.230 world. 230 00:12:04.230 --> 00:12:06.690 In my case, it's two different processes that 231 00:12:06.690 --> 00:12:08.930 talk with each other, but it could have 232 00:12:08.930 --> 00:12:12.190 been two different servers, one sitting in US 233 00:12:12.190 --> 00:12:14.410 and one sitting in Europe, talking to each 234 00:12:14.410 --> 00:12:14.690 other. 235 00:12:16.330 --> 00:12:19.630 And again, this sounds very much like MCP, 236 00:12:20.330 --> 00:12:22.430 but you have more. 237 00:12:23.690 --> 00:12:26.250 You could technically just make a tool called 238 00:12:26.250 --> 00:12:31.150 files tool in MCP and have an AI 239 00:12:31.150 --> 00:12:34.290 work behind the scenes, but it's not really 240 00:12:34.290 --> 00:12:35.150 meant for that. 241 00:12:35.350 --> 00:12:36.770 This is more meant for that. 242 00:12:37.030 --> 00:12:38.030 It's bigger things. 243 00:12:38.230 --> 00:12:41.230 It's more advanced things that are happening. 244 00:12:42.110 --> 00:12:43.950 So if you're using one or the other, 245 00:12:44.210 --> 00:12:49.110 I will probably use MCP more than A2A, 246 00:12:49.230 --> 00:12:51.730 but I can't see a world where A2A 247 00:12:51.730 --> 00:12:52.950 makes sense. 248 00:12:54.750 --> 00:12:56.390 So that is everything. 249 00:12:56.590 --> 00:12:57.770 Very simple to do. 250 00:12:58.650 --> 00:13:02.170 A little complex topic, but behind the scenes 251 00:13:02.170 --> 00:13:06.110 it's actually very simple that it's just these 252 00:13:06.110 --> 00:13:07.790 two lines of code that give us an 253 00:13:07.790 --> 00:13:09.210 agent that we can use as a tool. 254 00:13:10.270 --> 00:13:10.850 That's it. 255 00:13:11.230 --> 00:13:12.050 Thank you for attending.