WEBVTT 1 00:00:00.260 --> 00:00:02.820 Hi and welcome to this video series where 2 00:00:02.820 --> 00:00:06.120 we're going to talk about Microsoft Foundry. 3 00:00:06.880 --> 00:00:09.820 In this part one we will start with 4 00:00:09.820 --> 00:00:12.720 a high-level overview what it is and 5 00:00:12.720 --> 00:00:15.900 then we'll dive into the various topics in 6 00:00:15.900 --> 00:00:17.080 subsequent videos. 7 00:00:17.440 --> 00:00:18.360 So let's get started. 8 00:00:19.040 --> 00:00:20.680 So if I could only give you a 9 00:00:20.680 --> 00:00:23.280 single sentence of what this is, it would 10 00:00:23.280 --> 00:00:23.900 be this. 11 00:00:24.420 --> 00:00:28.060 Microsoft's one-stop shop for all things AI. 12 00:00:29.020 --> 00:00:32.340 As in a company is able to manage, 13 00:00:32.580 --> 00:00:35.600 develop all their AI solution in a single 14 00:00:35.600 --> 00:00:41.240 portal with Microsoft's known expertise in enterprise stuff 15 00:00:41.240 --> 00:00:43.460 like compliance and governance. 16 00:00:44.880 --> 00:00:47.700 Before we go into the offerings, let's quickly 17 00:00:47.700 --> 00:00:50.320 have a look at the history and rebranding 18 00:00:50.320 --> 00:00:51.940 that led to Microsoft Foundry. 19 00:00:54.350 --> 00:00:58.010 What Foundry is today started out as something 20 00:00:58.010 --> 00:01:00.110 called Microsoft Cognitive Services. 21 00:01:00.690 --> 00:01:03.489 This was back in the early days of 22 00:01:03.489 --> 00:01:08.510 AI where vision, speech and language capabilities were 23 00:01:08.510 --> 00:01:09.050 the thing. 24 00:01:09.610 --> 00:01:10.970 It was before GEN-AI. 25 00:01:12.550 --> 00:01:15.170 You will still see the Cognitive Service name 26 00:01:15.170 --> 00:01:18.890 popping up in various technical things, permissions and 27 00:01:18.890 --> 00:01:20.130 so on in the day. 28 00:01:22.570 --> 00:01:26.470 But after the moment happened in November 2022, 29 00:01:27.190 --> 00:01:28.830 everything was suddenly AI. 30 00:01:29.050 --> 00:01:32.230 So Cognitive Service was rebranded to Azure AI 31 00:01:32.230 --> 00:01:32.910 Services. 32 00:01:33.510 --> 00:01:35.430 And this is where we got the Azure 33 00:01:35.430 --> 00:01:37.930 OpenAI Service, which we also have as a 34 00:01:37.930 --> 00:01:39.050 NuGet package today. 35 00:01:39.630 --> 00:01:41.830 And we still use that if we want 36 00:01:41.830 --> 00:01:46.050 to work with individual modules in the Microsoft 37 00:01:46.050 --> 00:01:46.670 Foundry. 38 00:01:48.270 --> 00:01:52.290 The next years was the maturity phase of 39 00:01:52.290 --> 00:01:52.690 AI. 40 00:01:53.570 --> 00:01:58.250 And back then, everything just happened in the 41 00:01:58.250 --> 00:01:59.930 Azure portal as services. 42 00:02:00.370 --> 00:02:06.530 But in Build 2025, the Azure AI Foundry 43 00:02:06.530 --> 00:02:09.850 was introduced, which was for a very brief 44 00:02:09.850 --> 00:02:12.630 time also called the Azure OpenAI Foundry. 45 00:02:13.610 --> 00:02:18.890 But this shifted to this one-stop mentality, 46 00:02:18.890 --> 00:02:23.090 more models than just OpenAI, and it became 47 00:02:23.090 --> 00:02:27.630 a dedicated portal on ai.azure.com. 48 00:02:28.670 --> 00:02:31.870 So here we don't just have model services, 49 00:02:32.070 --> 00:02:34.770 but we also have persistent agents that was 50 00:02:34.770 --> 00:02:37.670 hosted directly in the portal and could be 51 00:02:37.670 --> 00:02:42.610 controlled and managed like most enterprise companies demanded. 52 00:02:45.790 --> 00:02:49.210 And finally, in November 2025, at the Microsoft 53 00:02:49.210 --> 00:02:53.650 Ignite Conference, we saw the latest rebranding with 54 00:02:53.650 --> 00:02:56.570 the name switching from Azure AI Foundry to 55 00:02:56.570 --> 00:02:58.090 Microsoft Foundry. 56 00:02:58.410 --> 00:03:01.910 Together with this name change, there was also 57 00:03:01.910 --> 00:03:06.670 a kind of V2 of the portal where 58 00:03:06.670 --> 00:03:10.910 the services got more evolved and workflows was 59 00:03:10.910 --> 00:03:11.570 introduced. 60 00:03:12.830 --> 00:03:17.110 As of this recording in December 2025, the 61 00:03:17.110 --> 00:03:20.350 new V2 portal of the Microsoft Foundry is 62 00:03:20.350 --> 00:03:21.310 still in preview. 63 00:03:22.210 --> 00:03:24.790 So you can still switch back and forward, 64 00:03:25.690 --> 00:03:28.270 but the new portal is the future. 65 00:03:29.070 --> 00:03:30.550 And it looks like this. 66 00:03:31.050 --> 00:03:33.430 And this is what we're going to go 67 00:03:33.430 --> 00:03:36.350 walk through in the rest of this video. 68 00:03:39.270 --> 00:03:41.130 In the new portal, there are three main 69 00:03:41.130 --> 00:03:46.850 sections, Discovery, Build, and Operate, which you access 70 00:03:46.850 --> 00:03:47.970 from the top right. 71 00:03:49.870 --> 00:03:53.470 Discovery is a gallery of what models and 72 00:03:53.470 --> 00:03:55.850 tools the Foundry portal offers. 73 00:03:56.510 --> 00:04:01.250 As of this recording, there are 11,251 74 00:04:01.250 --> 00:04:03.670 models, which you can choose. 75 00:04:04.310 --> 00:04:06.470 And this is an impressive number, but in 76 00:04:06.470 --> 00:04:09.450 reality, most of them are actually from Huggingface, 77 00:04:09.590 --> 00:04:11.830 like 10,800 of them. 78 00:04:13.110 --> 00:04:17.010 Azure themselves only host 77 models, which is 79 00:04:17.010 --> 00:04:18.130 still a lot. 80 00:04:19.230 --> 00:04:23.490 And in the agent capabilities, only 43 of 81 00:04:23.490 --> 00:04:30.210 those work together with the agent capabilities. 82 00:04:31.190 --> 00:04:34.130 And that is mostly the OpenAI models and 83 00:04:34.130 --> 00:04:36.930 the ones that use the same OpenAI spec. 84 00:04:38.850 --> 00:04:42.210 There are various info on each model, and 85 00:04:42.210 --> 00:04:45.770 there's comparison features in the gallery that helps 86 00:04:45.770 --> 00:04:48.710 you pick and choose the one model that 87 00:04:48.710 --> 00:04:49.870 fits you the best. 88 00:04:53.000 --> 00:04:58.160 Discovery also have various tools that they can 89 00:04:58.160 --> 00:04:59.760 use together with agents. 90 00:05:00.440 --> 00:05:04.720 Again, they have an impressive 1,451 different 91 00:05:04.720 --> 00:05:05.200 tools. 92 00:05:05.480 --> 00:05:09.960 Most of them are, however, classical logic app 93 00:05:09.960 --> 00:05:10.480 connectors. 94 00:05:11.020 --> 00:05:14.400 But more interesting, there's various MCP tools, and 95 00:05:14.400 --> 00:05:16.680 there's also the ability to make your own 96 00:05:16.680 --> 00:05:17.180 tools. 97 00:05:19.810 --> 00:05:24.870 Finally, in the Discovery section are 14 solution 98 00:05:24.870 --> 00:05:28.430 templates, which is our various GitHub repo samples. 99 00:05:29.310 --> 00:05:31.190 Most of them are in Python, but there's 100 00:05:31.190 --> 00:05:32.930 a few of them in C-sharp. 101 00:05:35.550 --> 00:05:38.590 Next, we have the build section, which is 102 00:05:38.590 --> 00:05:40.250 the most interesting for developers. 103 00:05:41.690 --> 00:05:45.210 In here, we have agents, which is persistent 104 00:05:45.210 --> 00:05:47.830 agents where we can create one or more 105 00:05:47.830 --> 00:05:52.010 of them using either the GUI or create 106 00:05:52.010 --> 00:05:52.850 it in code. 107 00:05:53.590 --> 00:05:56.830 You can then add skills, knowledge, and memory 108 00:05:56.830 --> 00:05:59.730 to them, and you can leverage them using 109 00:05:59.730 --> 00:06:03.470 code, or you can publish them to Microsoft 110 00:06:03.470 --> 00:06:05.750 Copilot or Microsoft Teams. 111 00:06:06.350 --> 00:06:09.190 We will cover agents in much more details 112 00:06:09.190 --> 00:06:12.010 in subsequent parts of this series. 113 00:06:13.450 --> 00:06:17.870 You also have in the build section the 114 00:06:17.870 --> 00:06:21.310 newest addition to Foundry, which is workflows. 115 00:06:22.270 --> 00:06:27.290 Workflows is simply just multiple agents stringed together 116 00:06:27.290 --> 00:06:32.790 in a set of steps in order to 117 00:06:32.790 --> 00:06:34.890 solve more complex tasks. 118 00:06:34.890 --> 00:06:38.190 We'll again have details in a later part 119 00:06:38.190 --> 00:06:39.050 of this series. 120 00:06:41.680 --> 00:06:44.940 The build section also has the model section, 121 00:06:46.000 --> 00:06:49.000 but not as a gallery, but more what 122 00:06:49.000 --> 00:06:51.800 models you have actually deployed in order to 123 00:06:51.800 --> 00:06:52.600 use the agents. 124 00:06:54.360 --> 00:06:56.740 So you can either use these models you 125 00:06:56.740 --> 00:07:00.760 deploy in the agents or directly in your 126 00:07:00.760 --> 00:07:02.160 code without agents. 127 00:07:03.280 --> 00:07:06.880 It is also here you control how the 128 00:07:06.880 --> 00:07:09.640 model is hosted, what throughput it has, what 129 00:07:09.640 --> 00:07:11.200 version it has, and so on. 130 00:07:13.760 --> 00:07:17.600 Next up is fine tuning, and this is 131 00:07:17.600 --> 00:07:21.300 if you need that, you can perform a 132 00:07:21.300 --> 00:07:25.800 subset of fine tuning on certain models. 133 00:07:27.200 --> 00:07:30.300 You will need to provide some training data 134 00:07:30.300 --> 00:07:34.140 for the fine tuning, and then deploy that 135 00:07:34.140 --> 00:07:36.060 fine-tuned model yourself. 136 00:07:37.850 --> 00:07:40.290 We will not cover fine-tuning in this 137 00:07:40.290 --> 00:07:43.730 series, as it's very rarely used and is 138 00:07:43.730 --> 00:07:47.290 only a sort of last resort for very, 139 00:07:47.350 --> 00:07:50.250 very niche and specialised scenarios. 140 00:07:53.570 --> 00:07:57.130 Next up we have tools, knowledge, and data. 141 00:07:57.830 --> 00:08:01.910 These are essentially just where you create and 142 00:08:01.910 --> 00:08:05.470 define the capabilities of agents. 143 00:08:06.510 --> 00:08:09.210 So for example, you define an MCP server 144 00:08:09.210 --> 00:08:12.070 or you hook up to a vector store 145 00:08:12.070 --> 00:08:16.050 as a knowledge, and that is the configurations 146 00:08:16.050 --> 00:08:18.350 that you then hook up to your agents. 147 00:08:21.250 --> 00:08:23.650 The final two entries in the build sections 148 00:08:23.650 --> 00:08:25.810 are evaluation and guardrails. 149 00:08:26.950 --> 00:08:29.510 Evaluations being that you can add some data 150 00:08:29.510 --> 00:08:33.510 sets to your models that you can train 151 00:08:34.190 --> 00:08:35.010 them against. 152 00:08:35.350 --> 00:08:38.510 So you can make a model or an 153 00:08:38.510 --> 00:08:42.010 agent, you give it the training set, you 154 00:08:42.010 --> 00:08:45.270 tweak your code and your agents, and you 155 00:08:45.270 --> 00:08:48.690 see if it becomes better or worse, and 156 00:08:48.690 --> 00:08:52.910 if there's handle of the quality and security 157 00:08:52.910 --> 00:08:54.350 at the time. 158 00:08:55.690 --> 00:08:58.870 Finally, guardrails allows you to guard against certain 159 00:08:58.870 --> 00:09:02.670 uses of an agent like sexual content, self 160 00:09:02.670 --> 00:09:04.590 -harm, hate speech, and so on. 161 00:09:05.710 --> 00:09:08.550 So if you have certain use cases that 162 00:09:08.550 --> 00:09:11.270 deviate from the out-of-the-box default 163 00:09:11.270 --> 00:09:14.930 guardrails, it could for example be that you 164 00:09:14.930 --> 00:09:17.310 had a medical AI that need to have 165 00:09:17.310 --> 00:09:20.210 certain words be allowed in certain areas that 166 00:09:20.210 --> 00:09:24.490 would perhaps hit sexual content or hate content 167 00:09:24.490 --> 00:09:25.530 and so on. 168 00:09:25.810 --> 00:09:27.690 While it could also be that you want 169 00:09:27.690 --> 00:09:32.070 to have them even more strict, like for 170 00:09:32.070 --> 00:09:33.890 example, in a kid's app where you want 171 00:09:33.890 --> 00:09:38.330 to block certain things in a higher level 172 00:09:38.330 --> 00:09:40.010 of detail than the defaults. 173 00:09:42.320 --> 00:09:45.320 The final section in Microsoft Foundry is the 174 00:09:45.320 --> 00:09:46.300 operate part. 175 00:09:46.900 --> 00:09:50.220 This is where you control all the individual 176 00:09:50.220 --> 00:09:53.800 parts of agents, tool, knowledge, and so on. 177 00:09:53.940 --> 00:09:56.740 Because behind the scenes, they are still Azure 178 00:09:56.740 --> 00:10:03.160 services, blob storage, databases, and so on. 179 00:10:04.340 --> 00:10:07.120 You also can see a lot of stats 180 00:10:07.120 --> 00:10:12.660 in here and control quotas and user access 181 00:10:12.660 --> 00:10:18.200 control in all the control parts and control 182 00:10:18.200 --> 00:10:21.380 planes of the Foundry portal. 183 00:10:24.840 --> 00:10:26.660 So this is everything in this video. 184 00:10:26.980 --> 00:10:29.840 In the next videos, we will primarily dive 185 00:10:29.840 --> 00:10:32.060 into the build section, where we will go 186 00:10:32.060 --> 00:10:35.020 from just knowing the portal now to actually 187 00:10:35.020 --> 00:10:36.760 building something on top of it. 188 00:10:37.180 --> 00:10:37.920 So see you now.