Microsoft Agent Framework releasing version 1.0

Key Takeaways

I can now send a request to my workflow, right? Just start triggering my workflow. It's going to the blog post grafter agent is now working. You see the history >> the nodes spinning up. Hello everyone. We are back at Microsoft I hope you are at Microsoft show and we're back with the Microsoft agent framework team. Hello everyone. So we are here at Redmond and we have the pleasure to have an announcement here. Sean, tell me what's going on. What are you going to announce today on the agent framework? I had the first And what's going on? Tell me. Yeah, well, so you know, by popular demand we're going to have release agent framework to version 1.0. We feel ready. We really thank everyone for having the feedback they've given working with the community over the last kind of 5 months yet feedback we finished up all sorts of features on agent framework and so we now have version 1.0 Microsoft agent framework ready for you to use and our goal really is making something that is ready for your production and scale applications for agent systems. That's exciting time because you know, we we had a lot of feedback since we announced here at the show and um and I'm going to see now not just you know what's new from the first video to this one you know on that feedback period from public preview and Bon GA but also maybe what's a little bit from what's what's coming you know, what's next what's on the road map as well. Uh Ruby want to show us something? You want to Yeah. Go straight to the demo? Yeah. I just want to mention that now with agent framework GA we are also GAing uh the VS code extension support for us. So it's already for your production use as well if you happen to work in VS code. I think what you're going to show is like how how you can use the extension really closely with agent framework. We're looking at speed your agent development but then really good really excited to see it. Oh to get from a scratch the extension that's it. All right, here I am in VS code. So I have the VS code extension inside here. Um here on the left side we have this code toolbar where we're going to use a part of the the features from here. But the most exciting thing I want to show you here is actually our value right inside in the code chat window. Um so the app will get not only install a bunch of UI tools, you know, into VS code that you can work with but also install Microsoft foundry skill. I'll show you. We can we can check. It's actually here by typing in draft skill. So the extension installs two skills into your environment. So the first one is called agent framework code gen. So this skill has all the knowledge about agent framework and how to work with it best practices and different patterns. It knows how to uh create agents in the right. The second one we install is called Microsoft foundry and by name you can tell this is for Microsoft foundry service. Once you build your agents you can deploy to Microsoft foundry for management or evaluation or a bunch of cool stuff afterwards. So this one knows how to work with the service. So one of them is about like your deployment to the service into foundry and one of them is about building agent skills later. And then the nice thing of course is I think those are using like MCP servers and stuff on the back end. And so that's always up to date with like our latest guidance our latest SDKs all all the stuff for >> Absolutely. Yeah, and these skills get updated automatically. You know, you don't need to install a newer version of the extension. So automatically happens for us. This sounds great. So that's you start from scratch and halfway done. Mm. Yeah. Now when these skills get installed we now can turn co-pilot into an expert in agent framework. So now what I can just prompt it to say, you know, create me a multi-agent workflow. So in this case this is a a session I did earlier so it's just for time's sake I'm going to show you this one. Um that can draft blog posts and review the blog post content using a foundry model that I have deployed in the cloud. Um and then the the scorer is going to you know, read the content that it generated and then it score it. If it's good it will score it over 75 it's going to publish the blog blog post otherwise do not. So it's kind of a multi-agent workflow that we want to orchestrate here. Um And then co-pilot would figure out the right skill to use, right? So that's part of the benefits of having skills laid out like she so it knows, okay, I'm going to read the foundry skill because I mentioned I want to use a foundry model. And then it said, oh I know I want to you know, have to read this agent framework code gen skill so I know how to craft my code. So then it did a bunch of thinking um and generated full project for me. >> And all this normally takes a few minutes like you've used this before. It's kind of working through the skills. It's generating code. It's trying to Yeah. It's a full project. It's and they are on my machine you know, with these files created. Um it even knows that oh I have a foundry project fit in my VS code extension right here. This is a foundry project. It actually looked into that and knows I have two models deployed so it put the right end point into my environment. >> That's my favorite feature about foundry because I'm always trying to guess what part of the URL I need which region I need all that stuff and all that info. So it's pretty much ready to go. It even included these VS code launch files and and tasks. So it's already to go. And the launch files are about debugging and >> Oh debugging. Yeah. And for me I love the docker file there. Ready to deploy. Yeah, it's all all good to go. So So if you look at the code that this is the main code generated for my workflow, right? We're using the agent framework here because I'm building a multi-agent workflow. It's using a workflow builder and it knows how to orchestrate my multi-agents together. Um I can you know, come in here and separate to say what we would normally do with any code. Um and then we can just hit and I don't I don't want to show this is the debug configuration we generated by default. That's what the launch.json file has already set up for us. So all I have to do is to press the button. And just like how we would debug any applications in VS code, right? So that's going to launch both sessions. Um and then I hit a breakpoint at some point. But what's even cooler than you know, just debugging a file is we put this kind of UI on top of our debugging experience. space here. What we call it agent inspector. So it's the lot of things all in one. So we you know, came up with this name that hoping that it reflects what it actually does. It's really about inspecting the running agents or workflows. So on the left hand side is um agent because we composed our workflow using code. There's nothing visual. I don't know if I did it right. Well and and this is a pretty simple workflow but like very quickly you can get into workflows with with an app as an agent and loops and it can sometimes be very hard to see what's visualize is going on in code. So it's really nice to be able to kind of see see this kind of graph flow. So So that part is something new from from the first session from the last video. So was here on VS code before? I did I mean this this is new. So what I think that we showed last time is that is we have an internal tool with agent framework called D Y which is doing something very similar. Probably when you last saw it probably when we did single agents but we've added multi-agent. We share a lot of the capabilities and code between the two teams so that you can use it you know, within VS code within your debugging environment or you can use the same kind of capabilities kind of stand alone as part of agent framework for for testing and validating your agents as well. But at the same time what you saw is a playground. It it shows you the different executors the different nodes within your graph allows you to talk to your agent allows you to see the importantly the events of the tool calls which is really important when you're kind of debugging your your agent systems. Yeah, you can see everything in one place. I can see the really UI amazing. Yeah, and this is not just a static right? This updates dynamically as you run through this workflow. >> There's even a deploy button but we don't really talk about this today maybe next time. >> [clears throat] >> Like I can now send a request to my workflow, right? Just start triggering my workflow. It's going to the blog post grafter agent is now working. You see the >> And you can see the nodes spinning on the left. >> can see it will still this flow. Now it's trying to look for the latest release so VS code release and draft blog post and then then um as as we move along here we'll see more information popping up not yet but we'll be able to see it the input and output of each agent. So I click on it. What agent node I see the information here individual events uh are captured here as well and any tool calling that happened uh during this workflow. So it's really all in one place. It's a playground. It's a visualizer. It's a way a tracer, trace viewer all in here. And this runs in kind of fully automated machine. So I see here there's many of my co-workers not yet deployed to any instance yet. So you can iterate as many times as you want. So you still can open these on the browser? As a dev UI as before open open the UI over the browser. >> Yeah, you can use the dev UI tool as part of the major framework or you can use this part of the extension. >> On the extension. That that's in here, yeah. Yeah, so now we are moving to the second agent. So let's run through this workflow from our one now. We get its information here and more events getting collected. So you can drain any of these events. I like that trace as well. I love the trace. >> [laughter] >> You're like a DJ after the traces a little overwhelming. Were you trying to find that bug? Yeah, the traces, yeah. Yeah, now it's moved on to pass the draft to the reviewer to score for. Yeah. So yeah, it's pretty cool. And you know, I know you know, this is a fairly simple workflow, but I've definitely thrown some pretty complicated workflow at the tools and it's done a pretty good job. And we can see here, you know, one of the nice things I love about this kind of visual tool is that it's still running on the code. It's not like a pure low code experience where you're kind of track in that experience. It's literally running my pro code here and so the breakpoint is hit and now I'm like right back into the debugger and I can see each of the state of my machine. I can see all the all the watch points are Like how you would debug any other applications. I can continue this workflow that's back here. Let's try to finish up the last step in my workflow. So yeah, and then you can see um it's it's get a tool is called as well in my workflow cuz I asked to save my bar codes to word document at the end. So there's this tool call happened. So you can kind of see everything with yeah, related to your workflow and then fixing issues if if there's any. And then love that everything runs locally. Now I can bring up co-pilots, you know, ask it to try to fix issues and so on. So What kind of you know, improvement we did on the workflow itself? What is possible now more use cases available? Yeah, well, I mean, I think um you know, over the last few months we've been spending a lot of engineering effort on on the workflow engine. A lot of it has been around bug fixes and performance and and those kind of things. But we you know, we've added some more patterns. We've added some more capabilities where you can do things like like fan in and fan out, map reduce. So you can kind of parallelize your agent systems. And um what we're also starting to look at is like how you could use these workflows in in these kind of longer running more intensive Um so one of the things we're adding into the framework is the ability to support these kind of long running workflows that can be resumed and checkpointed uh especially as we see a lot of these kind of claw type scenarios, open claw becoming a lot more powerful. We want people to be able to build those on agent framework um but really have the power to you know, deploy those at scale, observe them at scale, be able to run them at scale um in you kind of a same development environment uh allows those features. Yeah, and um one thing we're going to bring here another video we like discuss a little bit about memory as well. So for enterprise companies that you know, looking to use these in from working production they want to make sure that all the the the box are ticked you know, ticked. So um assistance of the workflow if something goes wrong um and you told me something about something coming in the future that you work on will be looping those workflows. Tell me. Is something we can talk about or just Yeah, well, I mean, that's what I was getting at with kind of these claw loops and these agent loops which is you know, a lot of those feedback that we getting from the That's right. Yeah. Yeah, I think I think a lot of them you know, especially the newer models over the last few months we've seen have I've been able to kind of pay attention a lot longer and and so kind of running those models in these longer loops is something that's really interesting and um you know, a lot of challenges that come from that are places where we think agent framework can help. One is is like these long running tasks they work a lot better when you kind of give your agent a computer. So like giving it access to a file system, the ability to execute code, be able to do shell execute. That's all like really important to do in a way where you can handle permissions and it's not just kind of running wild on your system. The other you know, you were mentioning memory about kind of keeping that context um constraint for your agent. So we've added things around chat history, compaction scenarios where you know, you can write strategies where we can pull out some of those tool calls from the history, maybe you don't need them anymore. We can summarize some of the history, we can truncate it. Kind of keeping your agent on task as it as it runs longer and longer. Yeah, it's great. And someone watching you know, we get feedbacks a lot from the community from the Microsoft piece. When one that's like really digging on those you know, on different locals the last few months. What do you see on what What do you see improving more on the last few months or on the latest The leadership what's the big one? Yeah. You know, big C# and .NET community out there with the ability to walk. Make sure that C# is a first class language in agent framework. So developers have the choice of and picking their partners Python only? Python. What so what language does it support? I remember that was a big question on the first video when we announced the project was is which language is Yeah, so I mean, we we we we take it very seriously especially for .NET and Python. I think those are the two languages for agent framework right now. We'd love to do some more languages in the future but we're we're just starting to think about it. But those are kind of the two and we do take very seriously trying to have parity between the two languages. As we're developing features like as they're preview, you know, depending on which engineer, which engineering team is working on them, they may start on Python first. They may start on .NET first. So you'll see stuff kind of go moving at different rates. But once we kind of finalize a feature, we do want to make sure that has parity for both of those languages cuz we we do see a lot a lot of folks using .NET and C# to build these agent systems especially at scale as you move into these enterprise environments which is a real focus for Yeah, I know I have a lot Python we going to bring you back me. Don't worry. We can to talk about. So someone want to start tomorrow. I'm happy that C# now is supported. We are committed to keep this growing. So how they get started? Just get the extension? Yeah, I think the easiest way to get started. I think so. I think so. And the other thing I'll mention, you know, getting started if you're starting from scratch, definitely get the extension. But the other place where the extension is really really useful is is migrating from other frameworks. So in a lot of ways the agent framework is the successor to semantic kernel, it is the successor to the autogen. And so if you were building solutions like in those frameworks already, just open them up in VS code, use the extension, ask it to migrate you to agent framework. I've seen really really good results doing that. Folks have been able to migrate very quickly and easily to the new agent framework which is our consolidated path forward. Like with both the autogen team and the semantic kernel team are working together on agent framework. It's it's our our path forward for for agent building at Microsoft. And so we always looking help also because this show we always promoting the open source open source project and we ask now people to help. So if you want to join, join the project, you know, become a contributor, maybe documentation, get the first issue there for the GitHub and then and help. How is how is that community engagement? >> Yeah, for sure. So obviously go to our GitHub repo agent framework GitHub github.com/microsoft/agentframework. Issues, discussions, we're all very active there. We have an active discord. You can find a link to that on our on our GitHub. And then also I'd encourage folks, we do weekly office hours. You can meet the engineering team. Most of the engineering team is there every week um and and you can join us and talk about stuff. And then yes, certainly we we we accept contributions um and and integrations with agent framework. So you know, we'd love to get engagement from the community for that. That's all. Any final thoughts on Well, I'm excited for sure. Great journey and really excited that we are here finally at the G max. Yeah. Can't wait for you guys to try out and see what you're building when agent framework and building and and let us know what you want to see more where we can you know, do better what's missing and to better meet your needs. Yeah, let us know. Make sure that you follow the Microsoft developer channel and the open Microsoft show. We going to bring more videos. So we there. Let's summarize. So don't forget the deploy button I told you there. Don't forget the memory. We don't forget about you know, Python for Python developer and for C#, let's bring a few videos later. So watch out. See you all soon. Thank you.