WEBVTT 1 00:00:00.000 --> 00:00:06.000 Hi, and welcome to this short video on what the Agent Framework Toolkit is. 2 00:00:06.000 --> 00:00:09.000 Agent Framework Toolkit is something I, from time to time, 3 00:00:09.000 --> 00:00:13.000 mention in my various other videos, 4 00:00:13.000 --> 00:00:16.000 and I have made some dedicated videos about it before, 5 00:00:16.000 --> 00:00:22.000 but many people don't know what it is, so a small introduction here. 6 00:00:22.000 --> 00:00:25.000 So, Agent Framework Toolkit is, as you see on the screen, 7 00:00:25.000 --> 00:00:29.500 a set of three NuGet packages that sits on top of Agent Framework. 8 00:00:29.500 --> 00:00:32.500 and adds various quality-of-life features. 9 00:00:32.500 --> 00:00:38.500 So, the three main purposes is easier agent creation, easier tool creation, 10 00:00:38.500 --> 00:00:41.500 and out-of-the-box tools. 11 00:00:41.500 --> 00:00:46.500 This is for C Sharp only, so not for Python. 12 00:00:46.500 --> 00:00:51.500 And the best way to explain what it is, is by giving a few examples. 13 00:00:51.500 --> 00:00:55.500 So the first one is these easier agents, 14 00:00:55.500 --> 00:00:58.500 which is done using opinionated factories 15 00:00:58.500 --> 00:01:03.500 against the various providers, OpenAI, Google, Anthropic, and so on. 16 00:01:03.500 --> 00:01:08.500 And, as an example, if we want to make an agent 17 00:01:08.500 --> 00:01:12.500 for plain vanilla Agent Framework 18 00:01:12.500 --> 00:01:16.500 using Azure OpenAI with a GPT-5 Mini, 19 00:01:16.500 --> 00:01:19.500 with low reasoning effort and tool calling middleware, 20 00:01:19.500 --> 00:01:22.500 we would need to write something like this. 21 00:01:23.000 --> 00:01:27.000 We would need to make our client, we would need to make our agent, 22 00:01:27.000 --> 00:01:30.500 and inside our agent we need to set some extra things 23 00:01:30.500 --> 00:01:33.000 in order to set the low reasoning effort. 24 00:01:33.000 --> 00:01:35.500 This is what's called breaking glass, 25 00:01:35.500 --> 00:01:42.500 because it's kind of a very, very low-level overriding object here. 26 00:01:45.500 --> 00:01:47.500 On top of that, in order to set our things, 27 00:01:47.500 --> 00:01:50.500 we need to put our tools inside the chat options 28 00:01:50.500 --> 00:01:55.500 when we want to do it here, so we can't place it like normal. 29 00:01:55.500 --> 00:01:59.500 And we need to use, as builder, use tool calling middleware.build 30 00:01:59.500 --> 00:02:03.500 in order to actually get the tool calling in there as well, 31 00:02:03.500 --> 00:02:05.500 because that's not set on here. 32 00:02:06.500 --> 00:02:11.500 And the tool calling down here is the regular tool call, 33 00:02:11.500 --> 00:02:13.500 in which case, in this case, 34 00:02:13.500 --> 00:02:16.500 we're just printing out what tools we're calling. 35 00:02:17.500 --> 00:02:23.500 So this is the experience using the vanilla system. 36 00:02:23.500 --> 00:02:27.500 If you add the Agent Framework Toolkit on top, 37 00:02:27.500 --> 00:02:30.500 you would be able to do what you see on the screen 38 00:02:30.500 --> 00:02:32.500 with this code instead. 39 00:02:33.500 --> 00:02:39.500 So you would end up making an Azure OpenAI factory for agents, 40 00:02:39.500 --> 00:02:43.500 and you would make an Azure OpenAI agent. 41 00:02:43.500 --> 00:02:45.500 Had it been Google, it would be a Google factory 42 00:02:45.500 --> 00:02:47.500 with a Google agent and so on. 43 00:02:48.500 --> 00:02:54.500 And what I have done is I have made the configuration 44 00:02:54.500 --> 00:02:56.500 of each agent a flat structure. 45 00:02:56.500 --> 00:03:00.500 There's nothing multi-level in any way. 46 00:03:00.500 --> 00:03:03.500 Everything has been put down, so it's very, very easy to do. 47 00:03:03.500 --> 00:03:08.500 So, for example, reasoning effort is directly on some agent options. 48 00:03:08.500 --> 00:03:10.500 Tools is directly there, 49 00:03:10.500 --> 00:03:13.500 and a special raw tool called Details is there 50 00:03:13.500 --> 00:03:18.500 to take care of middleware in just a read manner. 51 00:03:20.500 --> 00:03:27.500 So much less code and much easier code to understand 52 00:03:27.500 --> 00:03:29.500 because it's just a flat model. 53 00:03:33.500 --> 00:03:37.500 And this is supported by all the major providers. 54 00:03:37.500 --> 00:03:40.500 So, for example, here is Antropics Cloud. 55 00:03:40.500 --> 00:03:42.500 We can do exactly the same. 56 00:03:42.500 --> 00:03:45.500 It's just called an Antropic Agent Factory 57 00:03:45.500 --> 00:03:48.500 with an Antropic Agent coming back. 58 00:03:48.500 --> 00:03:52.500 And in here, we have specific Antropic Agent options 59 00:03:52.500 --> 00:03:58.500 because in Antropic, the max output tokens is a mandatory field. 60 00:03:58.500 --> 00:04:01.500 So this one ensures that you remember to set it. 61 00:04:03.500 --> 00:04:06.500 And it, of course, supports their breaking glass features 62 00:04:06.500 --> 00:04:09.500 like budget tokens and so on. 63 00:04:10.500 --> 00:04:14.500 In the same manner, Google, you just make a Google agent 64 00:04:14.500 --> 00:04:16.500 or an AI agent, as you can see here. 65 00:04:16.500 --> 00:04:18.500 More about that later. 66 00:04:18.500 --> 00:04:22.500 And you can set, for example, their thinking level. 67 00:04:22.500 --> 00:04:26.500 And down here, you can see all the different providers 68 00:04:26.500 --> 00:04:28.500 that I support at the moment. 69 00:04:28.500 --> 00:04:33.500 So it should be fairly simple, and if you need one more provider, 70 00:04:33.500 --> 00:04:36.500 it's very easy to extend these. 71 00:04:38.500 --> 00:04:41.500 But you might think, okay, if we're going opinionated, 72 00:04:41.500 --> 00:04:46.500 it might not be able to do everything that Agent Framework can do. 73 00:04:46.500 --> 00:04:48.500 And that's not the case. 74 00:04:48.500 --> 00:04:51.500 There's no limitations despite being opinionated. 75 00:04:51.500 --> 00:04:55.500 So on the connection side, we can still use credentials 76 00:04:55.500 --> 00:04:57.500 instead of API keys. 77 00:04:57.500 --> 00:05:00.500 We can still go to responses API. 78 00:05:00.500 --> 00:05:02.500 We can still set network timeouts. 79 00:05:02.500 --> 00:05:05.500 Or if anything special you do, 80 00:05:05.500 --> 00:05:11.500 there's always a hook-in to the additional underlying options. 81 00:05:12.500 --> 00:05:15.500 And on the agent side, you don't need to make it a Google agent 82 00:05:15.500 --> 00:05:17.500 or an OpenAI agent. 83 00:05:18.500 --> 00:05:21.500 It's just a wrapper around AI agents. 84 00:05:21.500 --> 00:05:24.500 So if you need all the agents to be in AI agents 85 00:05:24.500 --> 00:05:27.500 to pass around to the rest of Agent Framework, 86 00:05:27.500 --> 00:05:29.500 you can do that. 87 00:05:29.500 --> 00:05:31.500 And it supports all the options. 88 00:05:31.500 --> 00:05:34.500 All the common options like instructions, models, tools, 89 00:05:34.500 --> 00:05:38.500 less options like ID, max description, temperature, 90 00:05:38.500 --> 00:05:41.500 all the different providers and factories. 91 00:05:41.500 --> 00:05:44.500 So AI context providers can be set, 92 00:05:44.500 --> 00:05:46.500 the chat history can be set, 93 00:05:46.500 --> 00:05:50.500 client factory and logging factory, 94 00:05:50.500 --> 00:05:52.500 and the services as well. 95 00:05:53.500 --> 00:05:55.500 We can support all the middleware, 96 00:05:55.500 --> 00:06:00.500 but it's been put into being properties on the options 97 00:06:00.500 --> 00:06:03.500 instead of down as builder, 98 00:06:03.500 --> 00:06:05.500 so it's easier to work with. 99 00:06:06.500 --> 00:06:09.500 And then there's, of course, the OpenAI-specific features 100 00:06:09.500 --> 00:06:13.500 like client type, reasoning effort, reasoning summary, 101 00:06:13.500 --> 00:06:15.500 service tier, and so on. 102 00:06:16.500 --> 00:06:19.500 And finally, I put in a bunch of debugging features, 103 00:06:19.500 --> 00:06:22.500 so you can, with one line here, 104 00:06:22.500 --> 00:06:29.500 add the raw HTTP tool client, 105 00:06:29.500 --> 00:06:32.500 so you can begin to see the raw JSONs being sent 106 00:06:32.500 --> 00:06:35.500 to the LLM and from the LLM. 107 00:06:36.500 --> 00:06:39.500 And if, again, there's anything I'm missing, 108 00:06:40.500 --> 00:06:43.500 there's an option to provide extra things 109 00:06:43.500 --> 00:06:45.500 for chat client agent options. 110 00:06:46.500 --> 00:06:49.500 So, again, no limitations at all. 111 00:06:52.500 --> 00:06:56.500 Similarly, we also have our tool factory, 112 00:06:56.500 --> 00:06:59.500 which makes tools easier to create. 113 00:06:59.500 --> 00:07:03.500 So, this reintroduces something that Semantic Kernel had 114 00:07:03.500 --> 00:07:07.500 in that we can add attributes onto our classes 115 00:07:07.500 --> 00:07:13.500 and the various methods we want with an AI tool attribute, 116 00:07:13.500 --> 00:07:15.500 where we give the tool name and the description. 117 00:07:16.500 --> 00:07:20.500 And if we do that, we can simply use this AI tools factory 118 00:07:20.500 --> 00:07:24.500 to, say, get tools from the raw class 119 00:07:25.500 --> 00:07:27.500 or the instance if you want to. 120 00:07:28.500 --> 00:07:30.500 You can, of course, do this with normal reflection. 121 00:07:30.500 --> 00:07:33.500 I've just done the reflection part for you. 122 00:07:36.500 --> 00:07:39.500 Similar, it has hooks into MCP servers, 123 00:07:39.500 --> 00:07:43.500 so the factory can just turn 124 00:07:43.500 --> 00:07:50.500 an MCP remote from just a URL into tools. 125 00:07:50.500 --> 00:07:55.500 You can, of course, specify header properties and so on if need be. 126 00:07:56.500 --> 00:08:00.500 And then there's local MCP servers that is also supported. 127 00:08:01.500 --> 00:08:05.500 Finally, I have my own implementation of agent skills 128 00:08:05.500 --> 00:08:09.500 that came a little before agent framework put it in. 129 00:08:10.500 --> 00:08:12.500 And it has its variety. 130 00:08:12.500 --> 00:08:17.500 It turns it into tools instead of using the AI context provider. 131 00:08:20.500 --> 00:08:25.500 Besides being able to take your own tools and define them easily, 132 00:08:26.500 --> 00:08:30.500 I also put in a set of common out-of-the-box tools. 133 00:08:30.500 --> 00:08:34.500 So, there's time tools, so you can tell your agent 134 00:08:34.500 --> 00:08:37.500 exactly what the time of day it is, 135 00:08:37.500 --> 00:08:40.500 either in UTC or in local time. 136 00:08:41.500 --> 00:08:44.500 There's integration into some weather tools 137 00:08:44.500 --> 00:08:47.500 for real open weather map. 138 00:08:47.500 --> 00:08:50.500 You just need a read API key from them, 139 00:08:50.500 --> 00:08:54.500 and then you have full access to weather around the world. 140 00:08:55.500 --> 00:08:58.500 There's file system tools, so you can create 141 00:08:58.500 --> 00:09:02.500 and get files and folders on the system. 142 00:09:02.500 --> 00:09:06.500 And there's guardrails built into them, 143 00:09:06.500 --> 00:09:09.500 so you can set what exactly they are allowed to use, 144 00:09:09.500 --> 00:09:11.500 of which folders and so on. 145 00:09:11.500 --> 00:09:14.500 And by the way, these tools are completely out of the box, 146 00:09:15.500 --> 00:09:18.500 but you need to define if you want to use them. 147 00:09:18.500 --> 00:09:21.500 It's not like if you use agent framework tool kit 148 00:09:21.500 --> 00:09:24.500 that you suddenly get full access to your file systems. 149 00:09:24.500 --> 00:09:29.500 That's on your definition of if you want to use them or not. 150 00:09:30.500 --> 00:09:34.500 There's website tools for getting content off a page, 151 00:09:34.500 --> 00:09:36.500 screen-scraping. 152 00:09:36.500 --> 00:09:40.500 There's HTTP client tools, so you can actually 153 00:09:40.500 --> 00:09:43.500 call APIs and stuff like that, 154 00:09:43.500 --> 00:09:47.500 if you desire to do so, and it can, again, 155 00:09:47.500 --> 00:09:52.500 have guardrails in on what different endpoints 156 00:09:52.500 --> 00:09:54.500 you are allowed to use. 157 00:09:54.500 --> 00:09:57.500 And finally, there's a little email tool 158 00:09:57.500 --> 00:10:00.500 where you can send SMTP emails. 159 00:10:00.500 --> 00:10:04.500 So just tools that you sometimes need 160 00:10:04.500 --> 00:10:06.500 and don't want to write yourself. 161 00:10:06.500 --> 00:10:08.500 So they have been written for you 162 00:10:08.500 --> 00:10:12.500 with lots of configuration options into every single tool. 163 00:10:15.500 --> 00:10:19.500 So that's everything. There is more. 164 00:10:19.500 --> 00:10:22.500 There's various quality-of-life extension methods 165 00:10:22.500 --> 00:10:28.500 and bedding factories, but we don't want to go into this. 166 00:10:28.500 --> 00:10:32.500 Instead, I want you to go in and check out the link. 167 00:10:32.500 --> 00:10:36.500 It goes to the repo, which has a bunch of readmes 168 00:10:36.500 --> 00:10:40.500 about what every feature works and a wiki about it, 169 00:10:40.500 --> 00:10:43.500 and, of course, access to the NuGet packages 170 00:10:43.500 --> 00:10:46.500 that you can get for free and just put on top 171 00:10:46.500 --> 00:10:48.500 with no limitations. 172 00:10:48.500 --> 00:10:50.500 So that's everything. 173 00:10:50.500 --> 00:10:52.500 See you in the next one.