WEBVTT 1 00:00:01.120 --> 00:00:03.300 Hi and welcome to this AI in C# 2 00:00:03.300 --> 00:00:05.620 video where we'll be covering the different 3 00:00:05.620 --> 00:00:08.900 options you have to use LLMs in 4 00:00:08.900 --> 00:00:10.300 C# in 2025. 5 00:00:11.260 --> 00:00:15.400 So we'll be covering Semantic Kernel, Microsoft Extensions 6 00:00:15.400 --> 00:00:21.160 AI and the newest addition, Microsoft Agent Framework. 7 00:00:23.140 --> 00:00:26.360 So let's quickly have a look at the 8 00:00:26.360 --> 00:00:26.680 three. 9 00:00:27.900 --> 00:00:30.540 Semantic Kernel is the oldest, releasing back in 10 00:00:30.540 --> 00:00:35.140 March 2023 and then came along in 2024 11 00:00:35.140 --> 00:00:40.680 October, Microsoft Extensions AI and now in October 12 00:00:40.680 --> 00:00:45.200 2025 we now have the Microsoft Agent Framework. 13 00:00:47.010 --> 00:00:49.510 All of them else are exactly the same, 14 00:00:49.610 --> 00:00:53.090 it's the Microsoft that are maintainer of all 15 00:00:53.090 --> 00:00:53.650 of the three. 16 00:00:55.090 --> 00:00:58.370 Two of them, Semantic Kernel and Microsoft Agent 17 00:00:58.370 --> 00:01:01.210 Framework being the same team while Microsoft Extensions 18 00:01:01.210 --> 00:01:04.230 AI is the .NET team. 19 00:01:05.250 --> 00:01:08.730 Else they're all free, they're all MIT licences 20 00:01:09.210 --> 00:01:11.370 and they're all released as NuGet packages. 21 00:01:13.730 --> 00:01:16.430 So let's do some side-by-side comparison 22 00:01:16.430 --> 00:01:17.090 of them. 23 00:01:18.090 --> 00:01:19.970 First up we just have chat. 24 00:01:21.010 --> 00:01:22.790 In Semantic Kernel you needed to make something 25 00:01:22.790 --> 00:01:27.350 called a kernel builder and add a kind 26 00:01:27.350 --> 00:01:30.290 of mini dependency injection which was a little 27 00:01:30.290 --> 00:01:32.370 funky but it worked. 28 00:01:33.090 --> 00:01:36.350 Then you made a chat completion agent where 29 00:01:36.350 --> 00:01:38.470 you could give your instructions to the agent, 30 00:01:38.610 --> 00:01:39.650 the system instructions. 31 00:01:40.630 --> 00:01:44.110 And then you did an invoke async on 32 00:01:44.110 --> 00:01:46.890 your question and for some reason it was 33 00:01:46.890 --> 00:01:51.370 for each back, not needed really but that 34 00:01:51.370 --> 00:01:52.390 was the way it was done. 35 00:01:54.310 --> 00:01:57.290 Extensions AI is a more low-level system 36 00:01:57.290 --> 00:02:00.030 so it doesn't have the concept of agents 37 00:02:00.770 --> 00:02:05.250 and for that reason it needs to just 38 00:02:05.250 --> 00:02:09.169 make a raw client to whatever you use, 39 00:02:09.190 --> 00:02:10.590 in this case Azure OpenAI. 40 00:02:11.250 --> 00:02:14.150 And then it turns it into an iChat 41 00:02:14.150 --> 00:02:17.830 client that can be used to do this. 42 00:02:18.830 --> 00:02:24.670 And the response is fairly concise but if 43 00:02:24.670 --> 00:02:28.950 you need system messages or instructions it's always 44 00:02:28.950 --> 00:02:31.350 something you need to put into the individual 45 00:02:31.350 --> 00:02:34.810 response which is kind of low-level and 46 00:02:34.810 --> 00:02:38.030 to me one of the most annoying things 47 00:02:38.030 --> 00:02:39.090 about Extensions AI. 48 00:02:40.110 --> 00:02:42.010 But else it's a very concise system. 49 00:02:44.140 --> 00:02:47.080 Microsoft's agent framework takes the best of both 50 00:02:47.080 --> 00:02:47.480 worlds. 51 00:02:48.300 --> 00:02:50.820 It still does the same as Extensions AI 52 00:02:50.820 --> 00:02:54.900 by newing up the raw client but instead 53 00:02:54.900 --> 00:02:57.480 of as an iChat client it says create 54 00:02:57.480 --> 00:02:58.400 AI agents. 55 00:02:59.120 --> 00:03:01.180 And here you can directly give the instructions 56 00:03:01.180 --> 00:03:04.360 and they are kept so they will be 57 00:03:04.360 --> 00:03:08.100 sent with every call to the agent. 58 00:03:08.900 --> 00:03:12.560 So you just make a run async and 59 00:03:12.560 --> 00:03:15.760 get your response back which is more or 60 00:03:15.760 --> 00:03:17.520 less the same as the chat response. 61 00:03:17.940 --> 00:03:20.180 In general these two are very much alike 62 00:03:20.180 --> 00:03:24.420 but with the agent on top it makes 63 00:03:24.420 --> 00:03:27.860 agent framework more easy in my opinion. 64 00:03:30.440 --> 00:03:32.100 When it comes to tool calling they can 65 00:03:32.100 --> 00:03:32.800 all do that. 66 00:03:33.700 --> 00:03:38.460 Semantic kernel need to have special classes that 67 00:03:38.460 --> 00:03:40.340 are decorated with a kernel function. 68 00:03:43.700 --> 00:03:46.440 And then we do the same stuff again 69 00:03:46.440 --> 00:03:49.380 but on the kernel we import the plugin 70 00:03:49.380 --> 00:03:53.040 as they call them into the kernel. 71 00:03:53.600 --> 00:03:56.740 And then when we do the agent we 72 00:03:56.740 --> 00:03:59.160 need to tell it with a bunch of 73 00:03:59.160 --> 00:04:02.420 kernel arguments and Azure AI prompts here that 74 00:04:02.420 --> 00:04:04.180 it's allowed to call them else it won't 75 00:04:04.180 --> 00:04:04.640 call them. 76 00:04:06.140 --> 00:04:10.600 And then you just do the normal loop 77 00:04:10.600 --> 00:04:13.300 for chat and you use the tool. 78 00:04:16.579 --> 00:04:22.480 Extensions AI uses delicate methods for tools so 79 00:04:22.480 --> 00:04:25.580 you can either just assign them using an 80 00:04:25.580 --> 00:04:28.540 AI function create here or you can do 81 00:04:28.540 --> 00:04:31.360 use some reflection to find all your functions 82 00:04:31.360 --> 00:04:32.160 in a class. 83 00:04:33.820 --> 00:04:37.020 Again we need to have the system message 84 00:04:37.020 --> 00:04:39.480 out by itself so but beyond that it 85 00:04:39.480 --> 00:04:42.640 is very simple and we don't need to 86 00:04:42.640 --> 00:04:44.380 tell that it's allowed to call them they 87 00:04:44.380 --> 00:04:47.760 are auto tool calling by default. 88 00:04:50.830 --> 00:04:54.810 Agent framework is exactly the same but whenever 89 00:04:54.810 --> 00:04:58.450 you create your agent you can tell which 90 00:04:58.450 --> 00:05:00.070 tools that agent have available. 91 00:05:00.370 --> 00:05:02.890 Down here you needed to tell what tools 92 00:05:02.890 --> 00:05:05.570 you wanted to use for each individual call. 93 00:05:06.030 --> 00:05:07.890 Here you can define all the calls on 94 00:05:07.890 --> 00:05:10.870 the agent and just make your different calls. 95 00:05:12.450 --> 00:05:14.810 But beyond that they are very much the 96 00:05:14.810 --> 00:05:17.150 same they're using the same way of defining 97 00:05:17.150 --> 00:05:22.470 tools so a little more easy because of 98 00:05:22.470 --> 00:05:24.590 the agents but nothing more than that. 99 00:05:26.900 --> 00:05:29.780 When it comes to unstructured outputs they all 100 00:05:29.780 --> 00:05:34.680 use some kind of output object in this 101 00:05:34.680 --> 00:05:35.400 case a person. 102 00:05:36.960 --> 00:05:41.900 Semantic kernel needs to on the agent say 103 00:05:41.900 --> 00:05:44.580 that that agent should always give a response 104 00:05:44.580 --> 00:05:47.400 format of the type of the person in 105 00:05:47.400 --> 00:05:51.720 this case and then normally call and the 106 00:05:51.720 --> 00:05:55.740 message back here would be a JSON string 107 00:05:55.740 --> 00:05:58.220 that you then need to deserialize. 108 00:05:59.720 --> 00:06:03.940 Extension AI make it much easier in that 109 00:06:03.940 --> 00:06:08.180 you have a generic method of the call 110 00:06:08.180 --> 00:06:11.460 to the LLM where you just give the 111 00:06:11.460 --> 00:06:14.020 object that you want back and then you 112 00:06:14.020 --> 00:06:16.900 get that back as a response so you 113 00:06:16.900 --> 00:06:19.940 get your object so you don't need to 114 00:06:19.940 --> 00:06:23.080 tell anything what type it is that happens 115 00:06:23.080 --> 00:06:25.840 in the call and you don't need to 116 00:06:25.840 --> 00:06:26.900 deserialize yourself. 117 00:06:29.280 --> 00:06:33.380 Agent framework does the same but of course 118 00:06:33.380 --> 00:06:36.980 again have the agent part and it needs 119 00:06:36.980 --> 00:06:40.060 to do a run async instead of a 120 00:06:40.060 --> 00:06:42.440 get response async but else they are exactly 121 00:06:42.440 --> 00:06:42.940 the same. 122 00:06:45.160 --> 00:06:48.240 There is certain type of agents that can't 123 00:06:48.240 --> 00:06:53.900 do this meaning the chat client agents are 124 00:06:53.900 --> 00:06:56.000 the only ones that can do it but 125 00:06:56.000 --> 00:06:58.300 they have some more advanced agents that are 126 00:06:58.300 --> 00:07:01.220 not allowed to use structured output. 127 00:07:04.810 --> 00:07:08.010 Other concepts well if we do a little 128 00:07:08.010 --> 00:07:10.730 side-by-side comparison the number of new 129 00:07:10.730 --> 00:07:12.330 gets you need in order to do hello 130 00:07:12.330 --> 00:07:14.430 world is two in each one of them. 131 00:07:15.790 --> 00:07:17.530 You can see the names here. 132 00:07:19.130 --> 00:07:19.770 Embeddings. 133 00:07:22.010 --> 00:07:24.010 Microsoft Extension AI is the only one who 134 00:07:24.010 --> 00:07:24.770 have embeddings. 135 00:07:25.170 --> 00:07:26.690 The two others leverage them. 136 00:07:26.950 --> 00:07:29.190 Semantic kernel had an old one but I 137 00:07:29.190 --> 00:07:32.210 haven't listed it here because nowadays if you 138 00:07:32.210 --> 00:07:35.190 use semantic kernel you just leverage Microsoft Extension 139 00:07:35.190 --> 00:07:35.490 AI. 140 00:07:38.250 --> 00:07:41.190 Agents semantic kernel and agent framework have that 141 00:07:41.190 --> 00:07:45.330 concept while extensions AI being a more low 142 00:07:45.330 --> 00:07:48.650 -level system you need to more control your 143 00:07:48.650 --> 00:07:51.810 own system messages and stuff like that. 144 00:07:53.050 --> 00:07:56.130 So everything happens on each individual call instead 145 00:07:56.130 --> 00:07:59.610 of defining everything on an agent and then 146 00:08:00.310 --> 00:08:01.250 working from there. 147 00:08:02.850 --> 00:08:03.530 Workflows. 148 00:08:03.870 --> 00:08:05.990 Semantic kernel have a workflow they call it 149 00:08:05.990 --> 00:08:10.190 process framework and agent framework of course have 150 00:08:10.190 --> 00:08:11.470 the workflows. 151 00:08:12.910 --> 00:08:17.030 Extensions AI don't have that concept so you 152 00:08:17.030 --> 00:08:18.350 can't leverage that. 153 00:08:20.150 --> 00:08:23.950 Agentic workflows is something agent framework have picked 154 00:08:23.950 --> 00:08:30.630 up from their other team called Autogen and 155 00:08:30.630 --> 00:08:32.250 that's now in the agent framework. 156 00:08:32.530 --> 00:08:35.510 It wasn't really in the semantic kernel and 157 00:08:35.510 --> 00:08:38.970 again Microsoft Extensions AI don't have any concept 158 00:08:38.970 --> 00:08:39.549 of this. 159 00:08:43.300 --> 00:08:48.620 Vector stores is something neither extensions AI or 160 00:08:48.620 --> 00:08:50.100 agent framework have. 161 00:08:50.540 --> 00:08:55.620 They have the embeddings but they don't have 162 00:08:55.620 --> 00:08:58.100 anything to actually put it up into a 163 00:08:58.100 --> 00:09:01.720 vector store like Azure AI search or into 164 00:09:01.720 --> 00:09:03.600 a SQL server or something like that. 165 00:09:04.060 --> 00:09:07.940 For that something called Microsoft Extensions vector data 166 00:09:07.940 --> 00:09:09.020 exists. 167 00:09:09.180 --> 00:09:11.700 It's a NuGet package and that NuGet package 168 00:09:11.700 --> 00:09:14.120 actually lives inside semantic kernel. 169 00:09:15.340 --> 00:09:18.960 At some point semantic kernel need to pull 170 00:09:18.960 --> 00:09:21.340 it out because it's not really semantic kernel 171 00:09:21.340 --> 00:09:23.560 that have the support for vector stores. 172 00:09:24.120 --> 00:09:28.060 It's just a NuGet package that the semantic 173 00:09:28.060 --> 00:09:31.360 kernel repo owns that have that support. 174 00:09:32.460 --> 00:09:36.300 So it's only in namespace only that they 175 00:09:36.300 --> 00:09:37.860 support it while the others don't. 176 00:09:38.960 --> 00:09:41.480 But in real life you will still use 177 00:09:41.480 --> 00:09:44.980 this part over in agent framework or in 178 00:09:44.980 --> 00:09:47.720 extensions AI if you use that for it. 179 00:09:50.290 --> 00:09:53.010 If we need to give some props to 180 00:09:53.010 --> 00:09:55.490 each one of them what is they better 181 00:09:55.490 --> 00:09:56.990 at than any of the others? 182 00:09:57.670 --> 00:10:00.190 Well semantic kernel is better at setting reasoning 183 00:10:00.190 --> 00:10:03.430 effort on LLMs. So which if you for 184 00:10:03.430 --> 00:10:06.390 example have chat GPT-5 you need to 185 00:10:06.390 --> 00:10:10.630 set between minimal, low, medium or high in 186 00:10:10.630 --> 00:10:14.370 reasoning effort and it has a nice little 187 00:10:14.370 --> 00:10:17.430 thing on your agent that you can just 188 00:10:17.430 --> 00:10:18.350 set that value. 189 00:10:19.750 --> 00:10:22.990 Neither extensions AI or agent framework have this. 190 00:10:24.050 --> 00:10:27.110 So for that reason technically semantic kernel is 191 00:10:27.110 --> 00:10:28.330 better at setting them. 192 00:10:29.070 --> 00:10:30.790 All of the three can do it but 193 00:10:30.790 --> 00:10:31.470 it's just better. 194 00:10:31.790 --> 00:10:33.530 It's more cumbersome over here. 195 00:10:35.530 --> 00:10:38.030 Extensions AI is better in embeddings because it's 196 00:10:38.030 --> 00:10:40.230 the only one who can do it as 197 00:10:40.230 --> 00:10:40.510 such. 198 00:10:40.710 --> 00:10:42.790 So I don't really know if you can 199 00:10:42.790 --> 00:10:45.290 make it better by being the only one 200 00:10:45.290 --> 00:10:45.710 who can. 201 00:10:46.670 --> 00:10:50.690 While agent framework is better at chats it's 202 00:10:50.690 --> 00:10:54.890 more concise than agent framework and because of 203 00:10:54.890 --> 00:10:57.150 the system information you can put directly on 204 00:10:57.150 --> 00:10:58.410 agents it's easier. 205 00:10:59.130 --> 00:11:00.090 It's better at agents. 206 00:11:00.390 --> 00:11:04.850 The agents are much more evolved than semantic 207 00:11:04.850 --> 00:11:05.350 kernels. 208 00:11:06.030 --> 00:11:07.310 It's better at workflows. 209 00:11:07.690 --> 00:11:10.390 It's a new version of the workflow and 210 00:11:10.390 --> 00:11:12.270 the identic workflow is the only one that 211 00:11:13.120 --> 00:11:14.250 that is over here. 212 00:11:16.270 --> 00:11:21.390 When it comes to LLMs extensions AI and 213 00:11:21.390 --> 00:11:24.010 agent framework are exactly the same because they 214 00:11:24.010 --> 00:11:27.270 both leverage that anything that supports an iChat 215 00:11:27.270 --> 00:11:31.230 client it can support and that's virtually everything 216 00:11:31.230 --> 00:11:31.970 out there. 217 00:11:34.260 --> 00:11:37.420 Semantic kernel has its own connectors which is 218 00:11:37.420 --> 00:11:41.280 a problem of course because if new features 219 00:11:41.280 --> 00:11:43.820 come along they need to implement it instead 220 00:11:43.820 --> 00:11:47.020 of just leveraging the raw clients like the 221 00:11:47.020 --> 00:11:47.740 two others do. 222 00:11:48.180 --> 00:11:52.720 But they do support OpenAI, Azure OpenAI, AI 223 00:11:52.720 --> 00:11:57.240 Foundry, Google, offline modes which is also of 224 00:11:57.240 --> 00:11:59.860 course supported by the two others and they 225 00:11:59.860 --> 00:12:02.460 have a sort of connection to Cloud. 226 00:12:03.060 --> 00:12:07.780 So unless you're using the OpenAI, Azure OpenAI, 227 00:12:08.320 --> 00:12:11.900 I would be more worried sticking with semantic 228 00:12:11.900 --> 00:12:13.660 kernel in having those. 229 00:12:15.580 --> 00:12:18.020 So that's everything I have for you in 230 00:12:18.020 --> 00:12:19.040 terms of comparison. 231 00:12:19.760 --> 00:12:22.720 I'm definitely gonna stick with agent framework. 232 00:12:22.940 --> 00:12:28.760 I find extensions AI too low level, too 233 00:12:28.760 --> 00:12:33.520 much extra work here and there compared to 234 00:12:33.520 --> 00:12:34.940 what agent framework can do. 235 00:12:35.140 --> 00:12:37.280 And on top of that it gives all 236 00:12:37.280 --> 00:12:40.260 the nice workflows as well. 237 00:12:40.760 --> 00:12:42.540 And semantic kernel I see it as a 238 00:12:42.540 --> 00:12:48.420 legacy product now that agent framework will supersede 239 00:12:48.420 --> 00:12:50.560 in every way. 240 00:12:51.080 --> 00:12:56.380 I think there's one thing with Amazon Bedrock 241 00:12:56.380 --> 00:12:59.040 agents that are not supported yet in agent 242 00:12:59.040 --> 00:13:01.460 framework but beyond that it can do everything 243 00:13:01.460 --> 00:13:03.820 that semantic kernel could do. 244 00:13:05.800 --> 00:13:08.520 Say for the vector stores of course which 245 00:13:08.520 --> 00:13:13.200 again is just a namespace only that is 246 00:13:13.200 --> 00:13:13.420 there. 247 00:13:13.880 --> 00:13:16.400 So that is the only place I will 248 00:13:16.400 --> 00:13:19.240 ever use a NuGet package going forward and 249 00:13:19.240 --> 00:13:22.080 I will be migrating all my semantic kernel 250 00:13:22.080 --> 00:13:25.680 over to agent framework which in turn then 251 00:13:25.680 --> 00:13:28.860 just use the extensions AI behind the scenes 252 00:13:28.860 --> 00:13:30.260 for the low level stuff. 253 00:13:31.220 --> 00:13:32.180 So that's everything. 254 00:13:32.940 --> 00:13:34.020 See you in the next one.