Decode AI

AI Innovations Unveiled - From MCP, LLama Index, Copilot and Agents

Michael & Ralf Episode 12

Send us a text

Summary

In this episode of Decode AI, Michael Plettner and Ralf Richter discuss the latest advancements in AI technologies, including the Model Context Protocol (MCP), enhancements to M365 Copilot, and the new features of GitHub Copilot. They explore the implications of autonomous software agents, the capabilities of Llama Index, and the automation platform N8n. The conversation highlights the importance of these tools in streamlining workflows and enhancing productivity in software development. The episode concludes with a preview of upcoming events related to AI.

Takeaways

  • MCP protocol is a collaborative standard for AI agents.
  • M365 Copilot has improved search and content generation features.
  • GitHub Copilot's agent mode allows for autonomous debugging.
  • Project Paravan aims to create autonomous software agents.
  • Llama 3.1 offers competitive performance at lower costs.
  • N8n is a powerful automation platform for AI workflows.
  • AI tools are evolving to assist in software development.
  • The importance of creativity in coding remains essential.
  • AI is improving but still requires human oversight.
  • Upcoming events will focus on AI and agent technologies.

Links to the different topics

MCP 

Microsoft 365 Copilot 

GitHub Copilot 

Llama Index 

Automation Framework n8n 

AI, Microsoft Build, OpenAI, language models, AI development tools, hardware advancements, Google Gemini, technology development


Hello and welcome to the brand new episode of Decode AI. My name is Michael. I'm Ralf and today's episode is packed with updates. We're diving into a Gentic AI, the Model Context Protocol, M365 Copilot, GitHub Copilot, Llama Index, and at least a tool, N8n. That's crazy, full of different topics. So come with us, dive into the different specialties of everything. That was a little bit Italian here. Have you heard that? Anyways, so let's kick off with the first topic. Let's talk about agents and the MCP, the Microsoft Certified Profite. No, the... the Model Context Protocol. Exactly. That's another acronym we have in the AI world you may think you're familiar with. But that's something pretty new, which is excellent because it's not something coming from a single company like OpenAI or Meta or something like that. It's combined from different companies. kind of standard, I would say. There's no standard in the name, so I just assume it's a recommendation to work with agents, to bring more capabilities and standard procedures to AI-driven agents. And now the point where you can correct me what I understood wrong. Well, you just kicked it off, right? So the Agenda AI topic is already all over. And yeah, so what is the pursue goals for autonomously and making decisions based on evolving context? And that's where the MCP jumps in and it's like, and it's not like it's an open standard and it's only developed by OpenAI also, as you've mentioned already, So Anthrophic is in, Microsoft is in, we have OpenAI there. And this MCP allows agents to maintain a state, use tools, and also access memory in standardized and interoperable ways. So that means that an agent becomes more efficient, has more capabilities, and also has like a memory of the things it is doing, which is pretty important. as you don't have to take care of it yourself, for instance. So really, really cool stuff. when we talk about that stuff and the architecture of it is really cool. So you have like a client server model in it and you have to involve a host like cloud desktops or other clients. And you have also to run a server. to manage the secure and efficient data access. So based upon that, it is pretty powerful and brings up few topics to me in mind. So taking care of security for MCP agent is very important. You will have clients and server. So it's really nothing new in the IT world as we have client server applications for what? Three, four decades already. And so it's cool and it's standardized. So that's really very important. So privacy and security is something we have to think of. But on the other hand, we have the benefits like it is simplified with integration. So you quickly can connect AR models to various systems without building custom solutions. You have the flexibility. like easily switching between different AI providers or data sources with minimal changes. And this is crucial. So it brings scalability. So the modular designer supports easy expansion and adaption to growing needs. So that's really a cool thing for MCP protocols here. So MC protocols to be correctly. It's difficult. It's always difficult. Yeah, that sounds fantastic. And even when I think about the development, I would say, of agents in general over the last six months, we have seen so many different ways how agents evolve in the Microsoft, but also in other companies. So it feels like all the big companies realize they could try to sell something like a model or a closed box using a specific model, like Microsoft does with this MG65 Copilot But there's another opportunity to have a more specific task-oriented tool, like an agent who can help the companies who sell something, but also the companies who want to use the agents. to address more specific topics. And that's something I would like to bring up. you see the Microsoft world where we have seen the first agents with reasoning models behind that. So this is just a development and evolvement, I would say, how critical and great agents has been become over the last six months. And now we have Yeah, a standard, as you said, to orchestrate more details to those agents. Yeah. So an agent is now something which can be enabled to be a long lift agent. So it can handle tasks over extended periods and it has memory and also supports dynamic reasoning, for instance. So it's really cool. You definitely need to check that out. So maybe you have a hint where to. Absolutely. And as always, you can find a link in the show notes to find more details on public websites, you don't have to trust the two dudes from the internet or in a podcast talking about some stuff. So there's official articles, documentation, we give you some entries to follow up the topic. Have a look into the show notes. Yeah. Summarizing up, MCP protocol is as well as AI there to stay. And it's already implemented, for instance, at Google DeepMind or at GitHub Copilot. So take a look at it. It's really worth it. And now Michael, switching to your more comfort zone, when we talk about M365 Copilots and with April, there were a... A lot of updates, some powerful improvements to M365 Copilot, especially around search and content generation, am I right? Yeah, since a long time, we have the search connectors already. So when you work in the field of SharePoint or if you work with Microsoft 365 in general, there are some search connectors since, I don't know, two years, three years. And those search connectors could help to connect to different data sources. and bring the knowledge of different data sources into your general search you have usually within SharePoint, for example, or the online services. And you can use those connections already for specific agents. That's the general use case for, I would say, for four, five, six months already. So you can use it for As I mentioned, specific use cases, want to address one of my customers, for example, is using it to bring some Confluence data into the general agent capabilities and combined SharePoint and Confluence data, bring everything together. that's itself, this itself is not the new point, it's coming to... M365 Copilot, which is kind of closed box, I would say, where you can address so many different aspects. You can add some add-ins, some... I'm not sure if plugins is the right word. So you can program some stuff and it can work with that together. But now we have the improvement. that the M365 Copilot can use the search connectors coming from enterprise search where the search connectors feed everything in, can work with the data right away without any additional configuration if it's not necessary. That's something I like. Yeah. me automatically? Or is there some extra effort I have to put in to get that feature running? Well, the feature itself is available and we have got a lot of improvements. Now you can index. I think it's about 500 million items per search connector. But you have to configure it. Also how it handles the different permissions you may have in a different system. For example, you don't want to... spoil everything from your confluence or fire drive, fire drive for your local drive, your local storage or Jira where you don't have permissions to. So that's one of the fundamentals, but it has to be configured. So you have to use an account, a specific setting. actually in the search connector to ensure you just provide the right data instead of everything on the file share, for example, or I don't know, Google Drive, Confluence, Jira, there are about 20 different services you can connect and it's still growing. So if it's not there, what you're currently looking for, or if it's currently not there, what you're looking for, then you just wait, I wouldn't say days. but months maybe, so it's growing. The opportunity to use this service is huge, and you can definitely improve your data with external sources. That's an awesome feature. Enterprise Search sounds like a must-have for the most businesses. So cool. We're a little bit Microsoft-centric at the moment, but there's also a new feature like generating images, tables, and presentations from a prompt by the latest AI models. Is that feature in preview? Is it live or do I need to build an extra... subscription for it or how does that work and what is coming there Well, Microsoft always improves the models behind the curtain. usually what happens, they have a close work relationship with OpenAI, as you may have heard about. And every time OpenAI develops a new model, Microsoft is looking for how can we use this model? How does it fit into our strategy? It's not just copy paste and provide this model. It's also about the different boundaries, the different topics coming from responsible AI and how Microsoft improves the model to get the right answers and gives you the work feedback you're looking for and not something coming from the web. So it's more specific. It's not just use this model because it's new and it's better and it's cheaper or faster or So it makes sense to, from Microsoft perspective, to ensure the quality is there and not only to provide the latest models. That being said, of course, they improve the models and they put in new models into M365 Copilot. So yeah, that's available. It is as... as much as I know, but GA is a different term or it's a complicated term in the Microsoft world. It means global availability, that means it's all general availability. That means in general, it should be available, but... Coming soon to your town. it takes some time to roll out to all the different multiple hundreds of thousands different tenants. So yeah, but the improvement is we got better quality or we get better quality. That's actually something many people pointed out that they got from the designer. which is the product name of creating pictures from Microsoft. The quality is not good, not on the same level as we have with other tools. I don't want to talk about AI tools built to create images. That's another story completely. But if you compare it with even with OpenAI and JetJPT, capabilities to create pictures there. That's something you have missed in the M365 Copilot. And now it's got improvements and it's better. I wouldn't say it's perfect, but it's on a good way for a one-stop solution. It's really good. So summarizing that up means there is an update on the way towards all the tenants out there in the world, depending on the data center you're connected to. may take some time when it is also available for you. And then you say the update improves a lot what the abilities of the Copilot in that region can do. Is that right? Awesome. So cool. I'd love it. Okay, the next topic is something I barely use. I use Copilot in GitHub. just to go back to the history, that was the first Copilot we have ever seen from Microsoft. So it's the oldest one. But the GitHub Copilot got also some improvements, especially with Workspace. I would say that the product name is Workspace. know, it was Workspace. So yeah, Workspace has got some improvements, but that's not the important thing. The important thing is that GitHub Copilot got an agent mode and that brings autonomous debugging and process optimization, which is currently in preview. And the agent mode elevates AI assisted coding by enabling GitHub Copilot to independently iterate on code, identify and fix errors. and analyze runtime issues without requiring manual copy pasting. So you may have heard about Vibe coding. This brings it to another stage. So it also suggests terminal commands and prompts users for execution. So to try out agent mode, developers need to use VS Code insiders and activate the feature in Copilot chat settings. So the future plans are to include it in other IDEs as well and expand that functionality to them. So that's really cool. Another cool feature is that we have now the copilot edits officially available in VS Code. So that's something which was tested a long time ago. And it's now really a function like version one out there and is available in VS code first and will also expand to all other IDs. You can switch your models and you can combine powerful language models like Jimini 3.5 Summit, Claude Summit, mean 3.5, holy shit, Claude 3.5 Summit. or Gemini 2.0 Flash, and there are others as well. So developers can adjust suggested changes at any time and even revert to previous versions. So future enhancements aim to improve performance and integration with Copilot chat, as well as enable automatic selection of relevant files for editing. So that's pretty cool. Then we have the project Paravan coming up. So that's something which is going to make me a little bit. So I got to get goosebumps when I read this because next step towards autonomous software agents. Holy shit. Project Paravan offers a glimpse into the future of autonomous software development agents. These agents are envisioned to handle, routine tasks independently, such as addressing issues, creating pull requests and processing feedback. So. It's really something where you can delegate the task towards copilot. It sets up a secure sandbox environment, clones the repository for you, analyzes the code and makes necessary adjustments to it. That's spooky for me. I'm keen to test it. So that's really something overall, I would say these updates position GitHub Copilot as an even more powerful tool for developers. and transforming into a true partner in software development. it is really cool. GitHub is bringing out there with their GitHub Copilot. Yeah, that's really something. I'm still having goosebumps here. I totally understand that it sounds like witchcraft actually. I said I use it barely, but I use it sometimes. It was also kind of magical when I just wrote some code and I got some suggestions and I could interact with Copilot to review something I wrote before the Copilot, GitHub Copilot era. and then double check where my mistakes are and then work with that. But to realize that the agent is not only suggesting something, but actually doing something like an autonomous agent. And this is, I would say, also kind of future steps we will see in different areas. sorry, agents is a huge topic right now for me. I don't want to go back every time, but that's fantastic. That's exactly what you think about when you maybe go back 20 years ago and someone was thinking about the future, some real unrealistic future parts where you just thought, yeah, and I don't have to care about all the... the boring, repetitive stuff like you said it, check in the project, you get an error message, coming back, review everything. And if it's coming back with a solution and it's resolved already, that's... I totally understand your goosebumps. I'm not a developer. That's absolutely cool. Yeah, it's amazing what's happening there and I can't believe it. So this moves us from simple code suggestions, as you've mentioned a moment ago, to actual goal oriented development powered by AI. So it is really now your co-programmer when you use it right. I'm still not saying that GitHub Copilot will write its own code. So you have to put in structure ideas and on so on and so forth, but it's getting close. What I assume and I know if you have already some experience, I hope at least, if it's coming to the point where it checks with your existing data what you have maybe done already and suggest and put in some fixes based on the experience it already has from code you programmed already, that would be fantastic. Yeah, and it can help. mean, we're looking on the shortage of developers in the not so far future. well... Yeah, was the interesting point. Sorry, I want to go to a side quest. Can I? All right. I've been on an event earlier this month and it was on the University of Stuttgart. And there was an event talking about AI M365. There was a professor talking responsible for train, no not train, teach programming to students. And the entry was, here is a video how long it takes to solve our problems we've created 10 years ago. And it started a 45 second video. running through everything because the student just used GitHub Copilot and everything was there. No issues, everything went through, but the student had no idea what's going on there. And then they explained how they fixed it by adding AI, giving AI the opportunity to support someone like a... beginner. so that the student can write something and the AI is giving not the right answer but pointing to some issues. So that's something I was I really liked the idea about. And then we had a discussion is development coding actually something we need in a future state? And the answer is maybe. Maybe not, because you can use something like GitHub Copilot and it writes code right away. But you have to understand what it's written there and you have to understand the need it complete, it should fulfill and not only have some code. you still need someone who is able to bring everything together and to... Yeah, to double check. Sorry for saying that, but going through the code and saying that's important. the only the writing, the code writing itself is much faster, much better with AI combined. And I like the idea. And I think with those agents in GitHub Copilot, we are really close to a state where it's It's there. Yeah, it's close to it. So two things came up while you were talking. First of all, AI is not fail safe. Absolutely important, you have to know it's guessing. Second one is, it's getting better by guessing. So trying to ensure that its answer is correct. So better guesses than we can ever make. And second thing is, AI doesn't have yet the ability to think creative. And to be a good developer, you don't need only to understand coding and all the semantics and everything around. You also have to have understanding of credit, creativity, or creativity for on how to solve issues with several things and or to bring up a solution for special demands. And this is something which is till now still missing in AI. So I'm fortunate within my lifetime, I won't see that we're... I hope in another decade I may use my fingers and still write some code assisted by AI instead of double checking what AI put out. Yeah, I I like the discussion and to see also the challenges of someone who is just starting with the topic or the easier, the easiness of someone who can start with the topic, but where it fails to bring actually some learning aspects into, it's not just now write me a code to do this or that. It's more. That's interesting to see also the development in our school system. That's interesting. Our next topic is... Lama index to be precise. You're right. That's cool. So first of all, Llama has brought out its new model, the version 3.1, capable to three sizes or available in three sizes. So 8 billion, 70 billion, and 405 billion per meters. The largest model, Llama 3.1.405B is designed to compete with leading proprietary models like OpenAI's GPT-4.0, offering compatible performance at a prox half of the costs. So these models support multiple languages, including German, and feature a context window of up to 128,000 tokens, which is huge. So... Meta envisions Llama 3.1 as a foundation for developing AI agents capable of automating tasks such as code generation, data analysis, customer support, and for all that stuff, what you can think of. Meta has plans. We've talked about MCPs. So Meta has plans for something similar to it, but the other way around. requests for comments, RFC, for the Lama Stack API to encourage community engagement and development. So that's pretty cool. 3.1 is already available in Germany. so you can start using it. In summary, Lama 3.1 represents Meta's strategic move to democratize access to advanced AI models promoting an open source ecosystem that balances performance, cost efficiency, and developer autonomy, which is quite cool. But there is another thing which Llama Index came out. So MCP is one thing, agentic AI is another thing. And then we have now agent workflows from Llama Index, a robust system designed to simplify creation and orchestration of AI agent systems. So imagine you have a bunch of agents, how do you orchestrate them? So building upon their existing workflow obstructions, Agent Workflow offers a structured general approach to manage complex multi-agent interactions efficiently. So what are the key features of Agent Workflow? We have stateful coordination. multi-agent orchestration, real-time monitoring, which is also bringing up transparency, mandatory need for EU AI Act. So, because that provides capabilities to stream and monitor agent activities, enhancing transparency and debuggability. Nice word, right? Debuggability. And then the last point is that it has the flexibility of... its agent type. So it supports various agent architectures, including function agents for LLMs with function calling capabilities and React agents for general LLMs. So that's pretty cool. It's particularly useful for scenarios where you have the requirement of complex task executions, where you want to build a research assistant which is capable to search and analyze information from multiple sources, synthesize finding into a coherent report, putting up it into reviews and we find the content accuracy. maintain the content across multiple interactions, and then handles complex back and forth between different specialized components so that you're abstracting that boilerplate code and complexity towards agent workflow, which enables developers to focus on building about building robust stateful agent applications with ease. Cool. So Llama index, that's a cool thing. And then, Michael, is something. I don't know if you know it. There's a cool tool. want to give it a shout out today. To be honest, no. I don't know that. I've seen what you gave me to realize there's more out there I currently know. yeah, it sounds like... Shall we spoil the name or...? Yeah, yeah, spoil it if you can. No, I'm not sure. Nathan? I don't know. N-A-N? N8n, that's correct. So, it sounds like a security instance to ensure you can work better with agents and ensure they are protected. That's what I've understood from the parts I've seen. Am I wrong? You smile like I am. So NNN is something you can automate stuff with. And it offers you two ways to do that. So first of all, you can go by low code. So you have a graphical user interface where you can just drag and drop your stuffs and items around to bring up like summarization of your, what you want to do. Like you're searching for documents on your Azure storage, then you can do that. And you can just click your flow together and you can have a trigger whenever a new file receives, put something there, add tags, whatever you want. And no wonder that we are talking about it, but N8n is open source and is a workflow automation platform that combines the flexibility of code with the ease of no-core tools. So I like it a lot. I also use it a lot these times and days. It enables users to automate complex tasks by connecting various applications. And there are a lot connectors out already and you can then, or connect other services. And you can do that through the visual interface I've said already. And then you can make it accessible to both developers and non-technical users. Cause also the result of what is happening in the low code platform. So when you're using the graphical user interface to drag and drop your stuff together, it'll end up in code. And that's so cool. So that you can hand over that to developers to maintain it and so on and so forth. That's a little bit better than I have expected it at the end of the day. So the results were pretty cool. So you have. the key features within N8n. You have that visual workflow builder. You have the extensive integrations with, what is it, 400 pre-built integrations already. So just to name some of them, Slack, Google Sheets, GitHub, Microsoft, M365. So there are a lot, also Azure, including, then you have the possibility to add custom code to it. So whenever there you're stuck with something and you cannot find the right tooling or item you want to drag on it. You can engage in developer or develop yourself by Python, something which you need in your workflow. And then, this is what we talking about it, there are AI capabilities. So N8n integrates with most popular AI frameworks like long chain, or other automation stuff. And that means building an agent becomes more easy with an 8N than ever before. And this is so cool because it enables you to creation of AI part workflows and agents that can process and analyze data intelligently. I love it a lot. And then you have that deployment options where you can decide to run it on premises. a private cloud, private data center, public cloud. It runs in a container. I'm running it in a Docker instance. That's so cool and doesn't blow up my machine. And I also built some agents with it already. So what are the use cases? Business process automation, data integration, AI driven workflows. And for sure you can run custom applications with it. So I love it because it is open source. and has the flexible architecture, which makes it a very powerful tool for organizations who are seeking to streamline operations, enhance productivity, or maintain control over their automation processes. And that's really cool. So for instance, you can bring in Azure and OpenAI to make your own N8n workflow agent and AI-powered. Okay, that sounds a little bit different as I understood that. was, I wouldn't say I was fooled, but I was checking out the website and I've got an entry and I've seen an overview about bringing multiple services together and I was directly into Azure, Android. And so I said, yeah, that's, that's a way how you can bring up some Entry ID instances and more security into your agents. And I absolutely ignored the part, are multiple scenarios. That was just a single one. I'm the only one with Entry ID. And I also ignored there's a huge headline stating it's an automation platform. Thank you for clarifying that. I've learned something again. That's also good. So that's one of the reasons why we do this. yeah, that's my way to say we would like to add something now. I'd love to see, try it out, give it a chance and have a look. So it's free to check out and you can also run it on your own. Like I do in the Docker setup. It's well explained. Give it a try. Shoot out. It's also something I just realized and thought, whoa, that's impressive. Running into a Docker, that means it's really flexible. It is. Crazy. And then before we... One more thing. You can also build your own MCP server with an 8N. So now I'm done. As always, you can find the link in the show notes. So have a look. There are some real good use cases if you read more than just one page. But yeah, I've seen that you can host it on Docker and I was really impressed about that. Before we end our episode for today, I would like to highlight one... Actually, I would like to highlight many events, but one we are working close together to bring it in a very unique location, something really unique in Germany. Yes. So yeah, you're right. So it's, it's wrap up for this month AI rundown and yeah, we have to highlight something, Michael. I give you the name and you maybe give it to the rest. So we're talking about Agent Con. Happened the first time in the Netherlands on 22nd of April. Was a huge success. Nearly 300 visitors there and we had a really cool talks and workshops out there, free to visit event. But Michael, you can add anything you want. As I said, agents is a huge thing and realizing this is something which is, well, we have a specific conference for is definitely something you should be interested in. And the biggest and better, the best, my gosh, I should end with English. The best part is it's not only Microsoft driven. The idea is to bring up agents in general, talking about AI and agents in general, not only Microsoft driven. And we bring up some different companies together and it's right in front of the Cloudland conference at the same location at Cloudland. So if you're interested in this topic, I'm pretty sure you're also interested in Cloudland conference. And now we are coming to the dates. It's... in a real good season where you can enjoy the weather when all the sessions have ended at the evening. And it's in the upper north, I would say. I mentioned a very unique location. The very unique location is, and I don't know how to say it in English, Heide Park Solltau. Amusement park. I like Heide Park more. More my gosh. Yeah, but it's an entertaining park. One of the biggest ones in Germany. And it's something special. It's a location, right, where you can combine before or after the session, stay in, go going to the to the park itself. And then you also have the opportunity to have some hotels directly at the location and the dates. I missed the part with the dates. It's the first week of July, the end of June, the week of the 13th. Bring it to the point. It's calendar week. No, just kidding. It's the week starting with June 13th, going up to the 4th of July, I think. It's the 30th of June, yes, Monday. We kick off at 3 p.m. So it is in a good condition that you can either go first to the park or you have to work or whatever and then visit this free-to-visit event at the Abenteuer Hotel in the Heide Park. Yeah, that's crazy, right? It's pretty hard to talk about this with a specific unique name in English because I feel at least from my point of view, everyone in Germany knows the Heide Park Soltau. But if you say amusement park, that's something with a red light district, right? No, that's yeah, it's pretty tough. Again, there's a link in the show notes. Feel free and have a look into. Well, well wait wait wait wait wait wait, it's the link to ensure you find the right name and don't have to search for it. So AgentCon Soltau at Hyder Park. And it's brought to you by the global AI community, which is really cool. So all chapters within Germany are working close together to bring this event live to you and enabling you to utilize agent AI in future to bring you ideas, to train you, to give a mind or thoughtful ideas. on use cases and so on and so forth. So really worth to go there. And as Michael said, maybe you want to check out the Cloudland too. It's a festival event. So that means if you want to, it's almost 24 hours a day where you can consume cool talks, where you can chat and network with high tech professionals and so on and so forth. So really cool event at the Hyder Park. One more thing to add. As it is a free event, the Hyde Park Soltau ticket is itself is not included and will never be included. So you have to spend that money on your own. If you plan to stay a night there, please visit also cloudland.org, go to the shop and scroll down to the accommodation options and book your hotel room over there. All the way around, you do not have the opportunity to book any rooms during that time at the Ardenteuer Hotel Hyder Park. So I had to add that. I'm sorry. Absolutely. That's good you're pointing that out. Just to ensure no one is telling his company or her company, I'm leaving for a conference. then, yeah, it's a little bit surprised it's not included directly. Yeah. All right. That's all wrap up for this month AR Rundown. Thanks for tuning in. It's always a pleasure talking with you. If you like the episode, give us a follow. share with your network and we will catch up on the next one. the closing part. Are you doing that or I am doing that? It is up on you if you want to. All right, all right. try, I try. Lucky me, it's written there. Stay tuned, stay in, stay. I messed it up even if it's written there. Stay tuned, stay interested, sign up, listen up. Here we go. Bye bye. Take care. Thanks for listening. See you next time.

People on this episode

Podcasts we love

Check out these other fine podcasts recommended by us, not an algorithm.

MVP Voices Artwork

MVP Voices

Ralf Richter
Talk Microsoft 365 Artwork

Talk Microsoft 365

Talk Microsoft 365