WEBVTT 1 00:00:00.450 --> 00:00:03.050 As we saw in Hello World, we can 2 00:00:03.050 --> 00:00:07.109 use OpenAI or Azure OpenAI with Microsoft Agent 3 00:00:07.109 --> 00:00:07.730 Framework. 4 00:00:08.130 --> 00:00:10.210 But we are in no way limited to 5 00:00:10.210 --> 00:00:10.630 those. 6 00:00:11.170 --> 00:00:13.750 The reason why I chose those and spent 7 00:00:13.750 --> 00:00:17.450 more time than other models is that this 8 00:00:17.450 --> 00:00:19.370 is what I'm going to use throughout this 9 00:00:19.370 --> 00:00:21.390 course. 10 00:00:22.090 --> 00:00:24.510 And I'm especially going to use Azure OpenAI 11 00:00:24.510 --> 00:00:28.070 because that's my preferred provider. 12 00:00:29.850 --> 00:00:32.850 Azure OpenAI is pay-as-you-go, so 13 00:00:32.850 --> 00:00:34.690 it's cheaper, but it requires a little more 14 00:00:34.690 --> 00:00:37.550 setup, while OpenAI, as I mentioned in the 15 00:00:37.550 --> 00:00:41.390 video, needs five US dollars up front before 16 00:00:41.390 --> 00:00:42.650 you can begin to use it. 17 00:00:43.930 --> 00:00:48.290 But again, there's so many more models, Google's 18 00:00:48.290 --> 00:00:50.930 models, Anthropic models, Mistral, and so on. 19 00:00:50.930 --> 00:00:54.470 And I'm going to go through how each 20 00:00:54.470 --> 00:00:59.770 of these are being connected using Microsoft Agent 21 00:00:59.770 --> 00:01:02.390 Framework, so you have a reference if you 22 00:01:02.390 --> 00:01:04.069 want to use some of the others. 23 00:01:05.190 --> 00:01:06.890 So let's get through this. 24 00:01:07.840 --> 00:01:12.550 So OpenAI, we saw, very simple, OpenAI Client, 25 00:01:13.090 --> 00:01:15.550 Chat Client, and so on. 26 00:01:15.930 --> 00:01:18.750 And we only used this one specific NuGet 27 00:01:18.750 --> 00:01:19.230 package. 28 00:01:20.790 --> 00:01:23.850 Azure OpenAI, a little more complicated because we 29 00:01:23.850 --> 00:01:26.290 needed two NuGet packages, and we needed to 30 00:01:26.290 --> 00:01:30.170 put in an endpoint as well. 31 00:01:33.610 --> 00:01:36.910 Google is probably the one that are the 32 00:01:36.910 --> 00:01:40.330 least supported at the moment, as of this 33 00:01:40.330 --> 00:01:43.470 recording being 6th of January, 2026. 34 00:01:44.270 --> 00:01:49.390 There is a new official Google NuGet package 35 00:01:49.390 --> 00:01:52.310 underway, but so far we have only been 36 00:01:52.310 --> 00:01:54.450 to an unofficial package. 37 00:01:54.450 --> 00:01:57.350 As soon as the official package goes live, 38 00:01:57.510 --> 00:02:00.730 I will believe that it's just as capable 39 00:02:00.730 --> 00:02:01.550 as the rest. 40 00:02:02.370 --> 00:02:04.790 But for now, we have this generative AI 41 00:02:04.790 --> 00:02:08.410 Chat Client package that we need to NuUp. 42 00:02:08.990 --> 00:02:10.950 And while it looks like it can do 43 00:02:10.950 --> 00:02:14.370 the same here, it has certain limitations when 44 00:02:14.370 --> 00:02:16.230 you come to more advanced stuff. 45 00:02:17.390 --> 00:02:20.970 In terms of getting your API keys, you 46 00:02:20.970 --> 00:02:24.630 do that in aistudio.google.com or you 47 00:02:24.630 --> 00:02:26.650 can also use the Google Cloud Platform. 48 00:02:29.050 --> 00:02:30.450 Then we have Anthropic. 49 00:02:30.950 --> 00:02:36.710 Anthropic have an official Microsoft Agent Framework package, 50 00:02:36.890 --> 00:02:38.150 just like OpenAI have. 51 00:02:39.710 --> 00:02:41.990 And in that, we just need to new up 52 00:02:41.990 --> 00:02:44.530 an Anthropic Client, and in that we need 53 00:02:44.530 --> 00:02:45.650 to give the API key. 54 00:02:46.930 --> 00:02:50.130 And when we then use the agent, we 55 00:02:50.130 --> 00:02:54.050 need to give some extra information, because we 56 00:02:54.050 --> 00:02:56.910 can't give the model in the client itself. 57 00:02:57.090 --> 00:02:58.450 We need to give it in something called 58 00:02:58.450 --> 00:02:59.610 Chat Options here. 59 00:03:00.990 --> 00:03:03.910 And a very important thing for Anthropic is 60 00:03:03.910 --> 00:03:06.250 you also always need to give how many 61 00:03:06.250 --> 00:03:08.910 tokens per call you can use. 62 00:03:09.670 --> 00:03:12.590 So we do almost the same, but we 63 00:03:12.590 --> 00:03:16.010 just need to give some additional Chat Client 64 00:03:16.010 --> 00:03:20.090 agent options here with some model ID and 65 00:03:20.090 --> 00:03:21.690 options. 66 00:03:22.190 --> 00:03:23.990 But beyond that, it's exactly the same. 67 00:03:26.390 --> 00:03:32.270 Mistral from France have also only an unofficial 68 00:03:32.270 --> 00:03:35.690 package, but you can use it together with 69 00:03:35.690 --> 00:03:36.810 the agent's AI. 70 00:03:36.810 --> 00:03:41.410 And you just new up a Mistral client, give 71 00:03:41.410 --> 00:03:47.550 your API key, and write mistral-client.completions 72 00:03:47.550 --> 00:03:50.870 .create-ai-agent, because this completions is an 73 00:03:50.870 --> 00:03:51.690 iChat client. 74 00:03:52.730 --> 00:03:55.650 You go to admin.mistral.ai in order 75 00:03:55.650 --> 00:03:58.590 to get your models and API keys. 76 00:04:01.450 --> 00:04:06.430 XAI or Grok is possible to use as 77 00:04:06.430 --> 00:04:07.290 well. 78 00:04:07.430 --> 00:04:10.330 And Grok is one of those models that 79 00:04:10.330 --> 00:04:14.730 have just adhered to the OpenAI spec instead 80 00:04:14.730 --> 00:04:16.290 of making their own API. 81 00:04:17.149 --> 00:04:19.430 So what you do is you just use 82 00:04:19.430 --> 00:04:23.450 the OpenAI package, and instead of just giving 83 00:04:23.450 --> 00:04:25.870 the API key, you also need to give 84 00:04:25.870 --> 00:04:28.630 some options where you point to an endpoint, 85 00:04:28.930 --> 00:04:32.190 meaning you point to XAI's endpoint instead of 86 00:04:32.190 --> 00:04:33.230 OpenAI's endpoint. 87 00:04:34.170 --> 00:04:37.070 But beyond that, it works exactly the same 88 00:04:37.070 --> 00:04:40.190 way as OpenAI. 89 00:04:40.850 --> 00:04:43.150 And you will see this throughout some of 90 00:04:43.150 --> 00:04:46.890 these models that they have just taken the 91 00:04:46.890 --> 00:04:49.230 same approach as OpenAI. 92 00:04:51.670 --> 00:04:56.370 Microsoft Foundry is a bigger thing. 93 00:04:57.150 --> 00:05:00.490 You still use two NuGet packages, but instead 94 00:05:00.490 --> 00:05:02.550 of the OpenAI package, you use the Azure 95 00:05:02.550 --> 00:05:03.490 AI package. 96 00:05:04.250 --> 00:05:08.370 So this is something called persistent agents, meaning 97 00:05:08.370 --> 00:05:11.590 that in the other cases, we have in 98 00:05:11.590 --> 00:05:13.250 memory made our agent. 99 00:05:13.250 --> 00:05:16.590 In Microsoft Foundry, you actually make an agent, 100 00:05:16.730 --> 00:05:19.430 and it becomes a permanent thing up in 101 00:05:19.430 --> 00:05:20.070 that portal. 102 00:05:22.570 --> 00:05:26.330 You use something called AI project client in 103 00:05:26.330 --> 00:05:29.210 order to do this, and you can't get 104 00:05:29.210 --> 00:05:30.670 an API key for it. 105 00:05:30.830 --> 00:05:33.990 You need to use a role-based access 106 00:05:33.990 --> 00:05:36.290 system from Azure to work with them. 107 00:05:37.630 --> 00:05:42.270 Once you do that, you get your client, 108 00:05:42.530 --> 00:05:45.950 and it can create agents, and there are 109 00:05:45.950 --> 00:05:48.090 even version histories, so you can go back 110 00:05:48.090 --> 00:05:52.690 and forward between agents, and you can set 111 00:05:52.690 --> 00:05:55.510 up your models on what agents should be 112 00:05:55.510 --> 00:05:55.830 used. 113 00:05:56.810 --> 00:05:59.830 Once you have done all that, you can 114 00:05:59.830 --> 00:06:03.530 just get your AI agent like normal, and 115 00:06:03.530 --> 00:06:05.790 you don't need to create it every single 116 00:06:05.790 --> 00:06:06.090 time. 117 00:06:06.370 --> 00:06:07.930 If you have the name and it already 118 00:06:07.930 --> 00:06:10.710 exists, you can just get the agent in 119 00:06:10.710 --> 00:06:11.290 that manner. 120 00:06:12.670 --> 00:06:16.370 There will be separate lectures on Microsoft Foundry 121 00:06:16.370 --> 00:06:19.790 because it's a much bigger thing, but for 122 00:06:19.790 --> 00:06:21.870 now, we can just see that it is 123 00:06:21.870 --> 00:06:24.710 possible to work with them, but they work 124 00:06:24.710 --> 00:06:26.450 slightly different than the rest. 125 00:06:28.490 --> 00:06:31.850 Then we have Amazon Bedrock, which is part 126 00:06:31.850 --> 00:06:35.550 of the Amazon AWS system. 127 00:06:37.550 --> 00:06:41.710 They have their own official NuGet package for 128 00:06:41.710 --> 00:06:45.690 ME-AI, meaning Microsoft Extensions AI, and we 129 00:06:45.690 --> 00:06:49.610 use them together with Agents AI, and once 130 00:06:49.610 --> 00:06:52.470 you do that, you can make a new 131 00:06:52.470 --> 00:06:52.870 client. 132 00:06:53.710 --> 00:06:56.370 In order to do that, you need to 133 00:06:56.370 --> 00:06:59.490 make an environment variable, so it needs to 134 00:06:59.490 --> 00:07:01.130 be in environment values. 135 00:07:01.310 --> 00:07:03.030 You can't really set it on the client 136 00:07:03.030 --> 00:07:06.530 itself, but no big deal. 137 00:07:07.070 --> 00:07:09.770 Then you need to know what region you 138 00:07:09.770 --> 00:07:14.170 are in and get this runtime client. 139 00:07:15.050 --> 00:07:17.090 Once you have one of those, you can 140 00:07:17.090 --> 00:07:20.970 turn it into these iChat clients, and once 141 00:07:20.970 --> 00:07:23.630 you have that, you get the normal Microsoft 142 00:07:23.630 --> 00:07:24.310 agent. 143 00:07:27.230 --> 00:07:30.490 Then we have Ollama, which is an offline 144 00:07:30.490 --> 00:07:33.810 model, so you go and download the Olama 145 00:07:33.810 --> 00:07:35.050 service. 146 00:07:35.510 --> 00:07:39.190 You instal it in your machine, and you 147 00:07:39.190 --> 00:07:42.130 run some commands to download the offline models. 148 00:07:43.510 --> 00:07:47.170 Once you have done that, you make an 149 00:07:47.170 --> 00:07:50.150 Olama client using something called OllamaSharp. 150 00:07:50.350 --> 00:07:53.570 This is technically unofficial, but as it's offline, 151 00:07:53.710 --> 00:07:55.290 it doesn't really matter too much. 152 00:07:57.390 --> 00:08:00.370 This client is by default also an iChat 153 00:08:00.370 --> 00:08:04.190 client, and instead of some endpoint up in 154 00:08:04.190 --> 00:08:06.330 the cloud, you hook it up to a 155 00:08:06.330 --> 00:08:09.410 local host or wherever you host the Olama 156 00:08:09.410 --> 00:08:10.850 service. 157 00:08:12.130 --> 00:08:15.210 You tell what model you want, and then 158 00:08:15.210 --> 00:08:17.430 you just take this client and turn it 159 00:08:17.430 --> 00:08:20.910 into an iChat client agent, and from there 160 00:08:20.910 --> 00:08:22.270 on, it's exactly the same. 161 00:08:25.250 --> 00:08:29.090 Another offline thing you can do is Microsoft 162 00:08:29.090 --> 00:08:31.390 version called Foundry Local. 163 00:08:32.390 --> 00:08:37.130 In Foundry Local, you have Winget, or you 164 00:08:37.130 --> 00:08:39.610 can download it from a website, and you 165 00:08:39.610 --> 00:08:43.350 can instal the Microsoft Foundry Local as a 166 00:08:43.350 --> 00:08:44.390 Windows service. 167 00:08:45.230 --> 00:08:48.790 Once you have that, you have a NuGet 168 00:08:48.790 --> 00:08:53.170 package dedicated for that from Microsoft, and behind 169 00:08:53.170 --> 00:08:56.130 the scenes, it's using the OpenAI spec, so 170 00:08:56.130 --> 00:08:59.310 we just use the Microsoft Agent Framework OpenAI 171 00:08:59.310 --> 00:09:01.110 NuGet package. 172 00:09:02.670 --> 00:09:04.870 In order to do this, we then use 173 00:09:04.870 --> 00:09:07.990 something called a Foundry Local Manager that says 174 00:09:07.990 --> 00:09:09.630 it wants to start the model. 175 00:09:11.210 --> 00:09:13.590 So we have our model name, we start 176 00:09:13.590 --> 00:09:17.010 that model, and we get a manager back 177 00:09:17.010 --> 00:09:19.310 that it has now started the model, where 178 00:09:19.310 --> 00:09:21.810 we will get the information about the model, 179 00:09:22.050 --> 00:09:24.730 like, for example, its real ID. 180 00:09:24.730 --> 00:09:27.850 This is just the alias of the model, 181 00:09:28.330 --> 00:09:31.890 like the open source chat GPT. 182 00:09:33.590 --> 00:09:35.790 Once you have that, you need to make 183 00:09:35.790 --> 00:09:39.710 an OpenAI client, and all OpenAI clients need 184 00:09:39.710 --> 00:09:41.570 an API key, but you can just paste 185 00:09:41.570 --> 00:09:42.570 anything in here. 186 00:09:43.110 --> 00:09:46.350 In this case, no API key, and the 187 00:09:46.350 --> 00:09:48.790 endpoint will be given on the fly by 188 00:09:48.790 --> 00:09:52.010 the manager, so it will be some local 189 00:09:52.010 --> 00:09:55.370 host and a port that it assigns on 190 00:09:55.370 --> 00:09:55.750 the fly. 191 00:09:57.270 --> 00:09:59.570 And once we have that, it is exactly 192 00:09:59.570 --> 00:10:01.390 the same as we normally do. 193 00:10:01.910 --> 00:10:04.110 We get our chat client with our model, 194 00:10:04.630 --> 00:10:07.550 and we get our agent, so we can 195 00:10:07.550 --> 00:10:09.130 do our run async. 196 00:10:12.610 --> 00:10:14.270 Then we have OpenRouter. 197 00:10:14.270 --> 00:10:17.990 OpenRouter is one of these companies that have 198 00:10:17.990 --> 00:10:21.450 made deals with all the different providers, so 199 00:10:21.450 --> 00:10:25.410 you can use Anthropic, you can use OpenAI, 200 00:10:25.710 --> 00:10:27.790 you can use Grok, you can use all 201 00:10:27.790 --> 00:10:33.410 the different models from almost any provider out 202 00:10:33.410 --> 00:10:33.730 there. 203 00:10:34.550 --> 00:10:39.110 And they have, as well, taken the OpenAI 204 00:10:39.110 --> 00:10:45.850 as a NuGet package, so you just new 205 00:10:45.850 --> 00:10:49.590 up with OpenAI, and you put in your 206 00:10:49.590 --> 00:10:56.130 endpoint being OpenRouter.ai.api.v1. And once 207 00:10:56.130 --> 00:10:58.050 you have that, it works exactly the same. 208 00:10:59.310 --> 00:11:05.910 Same goes for another of these multi-LLM 209 00:11:05.910 --> 00:11:09.990 providers, together.ai. It's exactly the same, but 210 00:11:09.990 --> 00:11:11.210 just with a different URL. 211 00:11:13.630 --> 00:11:16.870 We also have Cohere, works exactly the same 212 00:11:16.870 --> 00:11:17.630 way as well. 213 00:11:18.650 --> 00:11:22.710 They have their own models only, and they 214 00:11:22.710 --> 00:11:25.530 are fairly limited, but they are good for 215 00:11:25.530 --> 00:11:28.410 small education scenarios. 216 00:11:31.030 --> 00:11:34.230 We have HuggingFace, which is one of these 217 00:11:34.230 --> 00:11:40.030 free offerings, so you can technically work on 218 00:11:40.030 --> 00:11:40.930 all this for free. 219 00:11:41.350 --> 00:11:43.770 They don't have some of the big models. 220 00:11:43.910 --> 00:11:46.730 It's more the more open source, the more 221 00:11:46.730 --> 00:11:51.030 small models that exist out there, and they 222 00:11:51.030 --> 00:11:53.570 are by far the biggest provider of models. 223 00:11:54.170 --> 00:11:57.550 You can use them using OpenAI, so it's 224 00:11:57.550 --> 00:12:00.890 exactly the same setup, just with a different 225 00:12:00.890 --> 00:12:01.430 endpoint. 226 00:12:03.290 --> 00:12:06.230 The final one we have is GitHub Models, 227 00:12:06.370 --> 00:12:08.870 which is probably the best way of getting 228 00:12:08.870 --> 00:12:14.110 hold of OpenAI and all the bigger models 229 00:12:14.110 --> 00:12:18.610 for free, because as part of GitHub, you 230 00:12:18.610 --> 00:12:23.450 can get with some fairly heavy rate limits, 231 00:12:23.690 --> 00:12:26.150 but I guess you could manage throughout this 232 00:12:26.150 --> 00:12:30.110 course to stay on GitHub Models if you 233 00:12:30.110 --> 00:12:30.790 wish to. 234 00:12:31.730 --> 00:12:34.070 You go up and make a personal access 235 00:12:34.070 --> 00:12:37.330 token up in GitHub, and then you use 236 00:12:37.330 --> 00:12:43.290 Azure AI Inference as a connector, so you 237 00:12:43.290 --> 00:12:46.670 make one of these chat completion clients that 238 00:12:46.670 --> 00:12:50.870 needs a URL, which is the inference URL 239 00:12:50.870 --> 00:12:55.930 for GitHub, the Azure key credentials, which is 240 00:12:55.930 --> 00:13:00.630 your GitHub personal access token, and then you 241 00:13:00.630 --> 00:13:05.090 can say, as AI client, in order to 242 00:13:05.090 --> 00:13:10.170 get your AI client here, and then it's 243 00:13:10.170 --> 00:13:12.290 run like normal, but it requires a bit 244 00:13:12.290 --> 00:13:15.570 more and more different NuGet packages. 245 00:13:15.890 --> 00:13:21.090 That's the negative part about this, plus the 246 00:13:21.090 --> 00:13:23.870 quite heavy rate limits. 247 00:13:24.130 --> 00:13:25.950 Of course, if you pay, you can use 248 00:13:25.950 --> 00:13:32.690 it without rate limits, but if you want 249 00:13:32.690 --> 00:13:34.430 it free, it is one of the options 250 00:13:34.430 --> 00:13:37.650 to actually get OpenAI for free. 251 00:13:38.510 --> 00:13:43.050 And that's all the different providers I have 252 00:13:43.050 --> 00:13:45.570 tried over the times. 253 00:13:46.010 --> 00:13:48.310 There are probably many more, and they are 254 00:13:48.310 --> 00:13:55.170 probably mostly just some OpenAI spec version, so 255 00:13:55.170 --> 00:13:59.710 as you can see, you can easily hook 256 00:13:59.710 --> 00:14:03.510 up to almost anything in this agent framework, 257 00:14:03.510 --> 00:14:07.550 so it's no way a vendor login to 258 00:14:07.550 --> 00:14:11.990 Microsoft or to OpenAI in any way. 259 00:14:13.670 --> 00:14:15.450 So that was the bonus. 260 00:14:15.830 --> 00:14:19.890 Let's go back to some of the more 261 00:14:19.890 --> 00:14:24.430 educational parts of the agent framework before we 262 00:14:24.430 --> 00:14:27.550 dive into some of the deeper topics.