chapters transcript notes
click any line to jump to that moment in the video
0:16 hello and welcome to a brand new show. This is a new podcast is if one wasn't enough, if you didn't the explicit measures, maybe this podcast will fit your need. I'm here with Matthias Turpak. We are starting a brand 0:26 new podcast. Hello Matthias, welcome to the show. Hello. Very, very excited about this. And how are you? I'm doing . Thank you very much. 0:36 We have just recently come off of the high of the Microsoft Fabric conference. We're very excited about all the agentic experiences that we're seeing. I'm seeing a lot of people on social media 0:47 and YouTube recently talking about we're building this thing. We have a CLI thing. What's all this stuff mean? And we did a recent session at the Fabric conference. 0:57 We did a full day pre-con, 8 hours. Really good. Just trying to get people's head around what is an agent. How does this stuff work? What does this mean to you? And we 1:07 felt this is a conversation that needs to go beyond the scope of just the Fabric conference. , what do you think? Where are we going to take this? let's see. , we've had 1:18 we spent loads of time preparing for that workshop, which was fascinating and really enjoyable. And every time we talk agentic topics, you and I, it it feels infinite, ? It's 1:30 it's never ending in terms of sort of new aspects of it, new developments. It feels it constantly keeps you on your toes if you're following the agentic space. . 1:42 we are just using that a bit as an outlet for that. And hopefully, [clears throat] for folks watching and listening, 1:53 providing a bit of guidance and you know, pointing out some some helpful things. There will be opinions for sure. much true. Yeah, and we'll see 2:05 if we if we do any predictions at any point, how how or wrong they were. No one knows, ? That's the 2:15 one thing that we can know for sure at this point. I think you make a really good point there. And we've also struggled a lot getting prepared for the Fabric conference because we started talking about this as soon as we got 2:25 accepted. And this was months out and we were trying to unpack what does this mean? What does this look ? How do we teach things? And as we're building this, month 2:35 after month, things were adjusting and changing. And we're , how do we build how do we build a session to teach people about things where the foundation of what agents are 2:45 doing inside of organizations or what we're using to build them with is fundamentally shifting. We started with a lot of heavy use around MCP and we're all on skills. And there's this agentic looping stuff that's going 2:56 on. Open claw comes out and that changes the game yet again. , this is a very fluid space and we're just going to be open and candid about what we're doing and what we're learning here. Yeah. 3:07 The name of the episode here is entering the AI mind palace. , I've watched I shows that make you think. One of the shows that I to watch was a BBC, 3:18 which I think you would it's that's in your area. The BBC was a Sherlock Holmes series. Really interesting. I Cumberbatch 3:28 Oh, what's the guy's name who does the Patrick Cumberbatch? I don't remember his name. But he's he does the he does the the show. He is 3:38 the investigator, ? He he is Sherlock Holmes. And every often he would enter into his mind palace. And it's this really weird infographic and he would go but that's how he 3:48 visualizes the world. If you look up the definition of a mind palace, it's your mental model of framing a room and putting things in a room. And those are memories and ideas and it helps 3:59 you visualize or materialize concepts about things. , this very first episode is just us talking about what does agentic mean to us? What does 4:10 this look ? What is our mental picture? What is our mental model of what an AI agent does and how it works? And , with that concept for this first episode, let's just start 4:21 unpacking what is an agent? , Matthias, I'll I'll kick it over to you first. Give you the first ring here at the at the first belt in the ring around what is 4:32 what is an agentic experience? What is an agent to you? . I think that's it sounds an easy question, but really it's not because all sorts of things are called agentic 4:42 nowadays, ? . , let me come at this from a very technical point of view and then I'm sure you'll have a very different 4:52 approach to it, ? , everyone who's played around with AI experiences with most certainly have started with 5:02 some chat experience, ? Chat GPT obviously in was the one thing I would say that brought AI into the consumer space big time a couple years ago. And I'm I'm 5:13 sure lots of people would have started there. , you have this thing that you're chatting with, , you put a question or a prompt in and you're getting a reply back, ? 5:26 let me differentiate that from an agent, which is ultimately the same thing, ? You put a prompt in, a request, a task using natural language and 5:39 ultimately you get some reply back. But what makes an agent different is everything that happens in between. . , unlike a simple chat system 5:50 where it's just text in, text out, an agent knows how to do stuff. An agent has -called tools, ? And can create things for you, can call 6:02 APIs, can make things, can change things, can update things. And , in a way, those are side effects, ? Ultimately, an agent experience is 6:13 prompt in, response back. But what really matters to you is all the stuff that happens in between. And , that would be the technical definition. I your definition. 6:24 I was going to go way more technical than you were on the definition of things. I was thinking way deeper around, , what is an agent? An agent is the harness and it's the memory and it's the tool that 6:36 you use. And and , there's a whole new layer of language that I'm using that people aren't familiar with. And I'm starting to see more of the language starting to appear 6:46 across social media and people are picking up on it. , I agree with everything you said. I I love the idea of the agent is the doer of things. It's it's supplementing 6:56 action on top of a creator's experience. One thing I feel and I'm going to maybe portray this a bit of a lens of what how I see Fabric and how Fabric is 7:06 using agents a little bit. Maybe I'll taint the lens here a little bit from that perspective. On one hand, I look at Fabric and a lot of it has been agents or large language models that have been 7:17 applied to semantic models ask a question, get an answer out, ? Q&A, I have a question, agent you go figure some things out about the model and then return results back to me. 7:28 I feel that adds minimal value to the user, to the end user. You're spending a lot of money on tokens. This agent thing is not cheap. Just to get what is my sales this month? 7:38 it doesn't seem an immense amount of value add. Where I start really finding incredible value is when I use the agents to build with me. And I think this is what you're talking about, which is the agent doing 7:49 things, ? When the agent creates the report, when the agent creates and modifies my semantic model, when the agent builds me a web application and helps me to design and 8:00 build a back end and APIs and functions. there's a lot of things that I would normally have coded myself and it would have taken me a lot longer to do. But the time compression from building a 8:11 website with a full back end and front end to create a website with back end and front end is almost hours compared to days or weeks. , again, you'll have to finesse the the output 8:21 product. I'm not going to say it's going to get it the first time away. But where I resonate is agents are there to help you with the creator's experience. Whatever the creator's 8:32 experience is, I'm whatever you're building, they're there to assist with that. It's not taking away your reasoning. It doesn't take it away the business process. It doesn't take away 8:42 the the the problem you're trying to solve. It's a supplemental tool that I'm using to do this. And I look at everything in the lens of , what's the harness? , the the harness for the 8:54 large language model and I think I I think I I my mental model is a lot of the large language model Opus or Sonnet or Chat GPT or any other model that you're 9:05 going to throw at things, that is what you said earlier. Text in, text out. That's all we're getting. And then I move over to this area of , what does the harness do? 9:15 And I think maybe we could spend a little time on identifying what as an outsider, someone who's looking at agentic things, how can we pick up on 9:26 what is a harness? And maybe some maybe talk to us Matthias about some harnesses that you're using in your workflow. Yes. I think 9:36 harness is probably what a lot of people would confuse with agent. And and 9:47 not necessarily because it's their limitation, but because it's really blurry, ? , let's give one specific example. Obviously, Claude Code is is a system 9:58 that is very popular and and obviously has is owning its space in a phenomenal way, ? , the question Let's be really clear. Claude code, you say 10:09 Claude code, I think Claude code software. Are you saying physically the desktop application? is it there's a downloadable EXE that is Claude code, yes? 10:19 what I'm thinking the CLI, ? , the the the coding [laughter] tool, ? At this point. Slightly different than what I'm thinking. I'm thinking of When you say Claude code, I think there's an EXE, an executable, that 10:31 is a software piece that you download directly to your machine and you open it up and it looks a window feels a little bit VS code slightly, but it's an executable. you're talking about Claude code the 10:41 command line interface a interface, the CLI, and that's what you're just saying of that's the the harness that you're describing. exactly, ? , one may say a tool 10:52 that is an agent, ? And and you're Yeah. this is your coding agent. And in fact, , it's it's frequently called that, ? and that doesn't 11:02 just apply to Claude code, it applies to Copilot, Codex, you name it, ? technically, 11:12 it's it's an agent which is the wrapper within which your agents run. , what I described earlier, you know, in a in a technical way is Sure. 11:22 the agent is effectively just the thing that receives a prompt, talks to an LLM, has access to tools, has 11:33 permissions, and pretty much then runs through a circle of agentic loop. And everything around it, everything that sets up the agent, everything that 11:44 provides a UI for you to put prompts in and get feedback responses out of everything that makes it live within an application, 11:55 on your desktop, on a website, in in in the console, that's technically the harness. and I think moving forward, 12:07 this will be the real differentiator in in how good and how efficient and and and how popular, frankly, agentic 12:18 experiences will be. It's choosing the harness that does all the setup around the actual LLM interaction. 12:28 I'm going to I'm going to rattle off just a couple identified harnesses that I see. And give me your read on am I on am I on point with these or am I I'm off base? I would argue that VS code is a 12:40 harness, ? , the chat experience on the hand side, it it's harnessing the large language model. , that's that's one of the harnesses. Yeah. There's a lot of really neat things about that harness. You can select which model, there's an auto 12:51 model selector, you can have YOLO mode, which is there's a lot of approvals that you need to have the the agent to adjust files or not adjust things or run command scripts or or not 13:01 run them. The harness in in the VS code experience is letting you identify what level of security you want on or off while you're working with the agent while you're having it do things. 13:12 Another harness I think I'd to identify is when I go look at Fabric. . I think it inside powerbi.com in Fabric, whenever you click those 13:23 chat it experiences. . Those are also, I would argue, a somewhat of a harness as . , very limited, ? And , this is where I I have a little bit of 13:33 gripes here around some of this is where the harnesses that I observe inside Fabric are much feat- less feature rich than what I see inside VS code. , that's another 13:45 harness I would look at. I'm building custom applications. And in my custom application, I have something that has you enter in a prompt and then I can do something with that and I 13:55 send an API call over to Foundry and Foundry returns a result as . I'm going to argue you can build your own custom harness for things 14:06 with web applications. And I'm even going to say I've used a lot more of AI Foundry or Azure Foundry to help use the agent use the AI, the 14:16 large language model in concert with tools, data agents, other systems that I can bring those executable 14:27 or or action-based things directly to the agent, ? , I want my agent to show up. I'm going to send you some text, write me an email, write me a draft email, and then go put it in my draft inbox in my 14:38 Outlook, ? That that's something that AI Foundry can do. . And , that's where I would argue that Foundry is also another harness. Probably more developer-centric, but another harness as . Any other 14:49 harnesses that you see in the marketplace that are interest- Oh, I forgot the biggest one. I have an open claw . next to me, that's another version of a harness. 14:59 They've got software you put on a computer and you give full control to that machine through open claw. Anything else that I missed? , even tools Lovable, for instance, 15:10 which I know you're a big fan of, ? of that, yes. , , I would say an agent is not something you can directly interact with. An 15:20 agent, by definition, is always a piece of software that needs to run somewhere and that need that need and you need some interface . to 15:30 talk to the agent and get something back from it. , that that bit is is the is the harness, ? as you mentioned, VS code chat 15:40 is is incredible. I know we both love it. It's it's it's it's fantastic how how it's progressing every single 15:51 month, , they're really really putting much effort and and and good engineering into it. I love it. I can't wait to get out 16:01 monthly new VS code releases. they generally don't contain Yeah. They're they're doing weekly releases, I I heard. . . And I'm not sure whether 16:12 you followed it, but this , you know, there is the mainstream VS code release, but there's also VS code Insiders, which is always a month ahead. Yes. 16:22 but it's it's publicly available. , it's public preview, ? if you want to get a preview of what's coming next month or or two months, you download VS code Insiders 16:32 and you get early access to new features. , this month, they launched as part of VS code Insiders a 16:45 side-by-side app, a dedicated new app that's effectively the the chat experience as as a standalone app. , this is big big deal. and 16:58 again, if you want to experience it, just go and download VS code Insiders and I'm sure it'll hit regular VS code next month or . . I will put that here in the chat window 17:08 as and we'll make sure we ask you include this as in the description of the video. , down below in the video, we'll make sure we add the Insiders link here as in case you want to get a hold of that. The link 17:18 will also be in the chat window here as . , in case you're interested in getting engaged with that, learning about what the new agentic experiences that are coming out for VS code, you can see those as . , that's also in 17:28 the chat window as . . , where do you Can I Sorry, can can ahead. Can I can I add when you were talking about your take on agents earlier, ? 17:39 Sure. I was missing, I think, one one very important aspect here, ? I think you were mostly talking about, let's say, the engineering 17:50 perspective, , what what do engineers, what do coders Yeah. in terms of agents. Sure. I think something that is getting bigger and 18:00 bigger nowadays is the other side of the coin, which is dedicated agents and and agentic experiences and agentic workflow 18:12 portals for non-coders. And I think Cohere and and everything that's in the same space, Claude Cohere stuff. A Cohere type type of assistance for 18:23 things, yes. And and this is what we're going to see I believe lots of innovation and lots of competing products. And this is what's 18:33 what will bring agentic experiences really deep into enterprises and and businesses of all sizes. And and , what we're talking about is 18:48 graphical interfaces that allow knowledge workers, non-engineers, to make connections, , to their email, to their notes system, to their 19:01 you name it. and create an automation just by providing a prompt. 19:12 just by writing some some some specs, if you will. and and effectively, I would say this is creating automation 19:22 software without a software developer. this is from my point of view, this this is this is the big revolution. That's, , you're taking the developer out of the equation, which 19:33 historically has always been the expensive and and and and difficult part it 19:43 it would have brought in either an inability to make an idea happen or it would have 19:54 massively delayed it. And that's all gone . And this is I think where the real potential is. Everyone is empowered to create their 20:05 own tools and workflows. I would I'm going to 100% agree with your statement here. . And I'm going to double down on this statement as . I feel I was just reading an 20:15 article and I'm I'm blanking on the company. There was some recent company that is in the agentic space, but they're they're they're moving much forward much more forward into this. . And they've been watching what code 20:27 or features have been developed by their teams. . And they're seeing a rapid increase. Oh gosh, I can't remember the name of the company. But it's it's going to bother me . I'm 20:37 going to go find the article. I'm going to go pull it out. But they are they're they're trying to gamify their AI a little bit, ? there's this there's two worlds of this. There's the developer side, which again 20:47 I think we as developers are going through a an internal crisis a little bit. Some developers are very resistant to this. why would I give the agent all this power? I my whole career has been writing code. Why do I 20:58 want to relinquish this? Why do I want to let the agent do all of that? and then on the other side of this we have a lot of business users I don't want to write any code. Look, I can go to lovable. I can go to replet 21:08 and I can make these incredible applications that are very robust experiences with minimal effort and I just talk to it. And 21:18 as we move towards this business side of things of what that solution's going to look . I really one thing I really co-work. This co-work kind of experience. Claude co-work 21:28 came out first. Microsoft has adopted or absorbed the Claude co-work by licensing it for their own uses for co-pilot as . I really feel 21:38 in general co-pilot has missed the mark agentically. A lot of these other companies are doing a much better job. Loop is a document collection tool that I've been using for years since it 21:48 came out. I really it. But it doesn't have the same features as what a notion does, which is a more agentically built more tightly coupled between a co-pilot and the loop experience, but 21:59 all built inside notion. I think this is the way of the way people are going to start thinking. And they were going back to this article that I was reading. The article was 22:09 saying that more than 10 or 15% of their PMs, the people that were just directing what features need to be built, they're building their own features, creating their own pull 22:19 request, making code happen and developing and pulling things out because the cycle time between idea and chew out something with a developer 22:30 has been reduced to to almost nothing. I can hand it over to this agent and 30 minutes to an hour I can get something built and created. And it's simple things I find as one who's a 22:40 bit higher level in the organization, there's always things you want to adjust about an app. This doesn't look . There's a bug here. This this formatting is off or this button is wrong or inaccurate or 22:51 whatever the whatever the things are. Those simple changes are just describe it, boom, off to the races and it fixes itself. I really this. I think 23:02 let's mention one tool here. N A 10 is is what started that whole space, . co-work is something 23:14 that came much later. Yes. and that I believe has 23:24 we're going to see much more of of those kinds of tools moving forward. I really N A 10. Are you using N A 10? no, 23:34 not for anything I'm personally doing. just for exploration. I've been playing around with N A 10. I have a an instance running in multiple environments. I've installed a 23:44 couple N A 10 on containers in Azure for customers. . Really the tool. And and to me I get frustrated at Power Automate because this feels Power 23:56 Automate. There's connectors and there's triggers and there's things that start stuff and things that but the difference is the way N A 10 did agentic 24:06 experiences is just really good. I really it. You drop in this node cuz again in for those who are familiar with Power Automate, you have these nodes or actions that are 24:16 occurring. And the the flow inside Power Automate is very linear. It's either horizontal or vertical. It's all you get. And it's very difficult to kind 24:27 of compress a workflow down to a smaller screen size. Where N A 10 doesn't really have that. It's just wires between different noted elements nodal objects. And then there's this one object called the agentic node. 24:40 You can give it instructions. You can give it skills. You can apply tools to it. And you apply what model runs inside that agentic node. Which going back to it's almost it's it's almost 24:51 helping you build the harness in real time. Pick the model, pick the tools, pick instructions, and you stitch it all together to make this actionable thing that you can use and repeatedly 25:01 run a process on top of it. . This is what I want. someone take note of this cuz this is what I want to be built in fabric and in Power BI and in Power Automate. This is 25:12 the solution in my opinion. the analogy that comes to mind here is when Open Claw rocked the world earlier this year 25:22 and suddenly, , had this incredible initial public release. what they said was this is the 25:33 personal assistant that we've always been promised in the past, but it's it was never delivered. And this is the first time where people get the 25:45 the assistant experience that people really really wanted and need. And I think we're at 25:56 a similar point with respect to workflow automation, , the what people would have wanted to get out of Power Automate, we're at a 26:09 point, , with other kinds of tools where we're getting, you know, those kinds of capabilities. And can't wait to see what's happening. 26:19 I'm going to wrap here. I think this Open Claw thing is very very interesting to me. I don't know if you're if you're following Microsoft stocks at all, but yesterday there was a pretty pretty good sizable bump for 26:30 Microsoft. Microsoft just recently announced they're trying to build a Open Claw part of their product feature. this is on TechCrunch. 26:40 this is not me making this. This is from TechCrunch. There's a TechCrunch credit here. And and they're saying they just announced this and whether it's Claude or the model that's, , 26:51 Opus or Sonnet or whatever that the item is here. Microsoft is looking to build more of the Open Claw type experience for an agentic item. Which I think is the way to go. 27:02 It focuses more on the creator experience. And to be honest, I get frustrated when I can't develop code talking to it through my phone or chatting to something. And it's powerful . this is fundamentally changing 27:14 how I work here. I'm going to wrap with some final thoughts, ? Yeah. The agentic space is here. W- w- it's it's happening. It's already available. We have to pivot our mind thinking. 27:26 We have to change our mental picture. We have to change our mind palace to start incorporating more things around what does agentic look in your space? And we're hoping that this show we're 27:37 going to do this regularly. There's going to be two episodes a week. We're going to do it on Tuesdays and Fridays. And we're going to go as long as we can until we run out of topics. we'll just see what happens here. but as we 27:47 unpack what agentic means for us, we're very much invested in this is the future. It's going to happen. We're going to try to continually unpack on this show both with a mix of just chatting and 27:58 discussing articles and things that we find, but also going further and doing real demos. this Friday, Matthias, you're going to prepare some content around what does an agentic pattern look 28:08 for you? And we're going to live demo real agentic things that we're building and using and developing today that will likely help you as a developer or a business user in the future. Any 28:19 final thoughts from you? yeah, I've been spending pretty much the last year or building agentic experiences to help myself and my my own daily 28:30 workflows. And I I I don't want to go back, ? I'm completely sold. [laughter] Exactly. this can only get better . 28:40 definitely keen, , to share a a few snippets here. but I think I mean I also have some concerns. 28:51 which hopefully we we get to share moving forward. we want to make sure, , we've we've got a balanced view here and True. Yes. 29:01 but let's keep that for another day. Awesome. Thank you all much. We appreciate you listening to the show. This is our introduction to agentic thinking. We're very excited to bring this show to you. We hope you enjoy it. 29:12 Many more conversations to come. We'll see you on Friday. Thank you all. Absolutely. Bye. Agentic thinking. [music] 29:29 Agentic thinking. [music]