Agentic Dev of PBI Reports – Ep.566

Kurt Buhler is back for the last day of a four-episode run, and the topic he asked for is agentic development of Power BI reports. Tommy Puglia and Mike Carlo spend the hour on why reports have lagged semantic models, what Kurt says changed in the week before the show, and the planning that still has to exist before an agent touches a visual.

News & Announcements

• Fabric’s Quest — Tommy’s text adventure, which he and Mike played the day before the show, is a Fabric app with a SQL database supplying the prompts. It plays like the command games of the 80s and 90s, with the snark of Homestar Runner’s Peasant’s Quest: the site boots at 320×200 in 16 colors with one dragon, that dragon is throttling your capacity, and the path runs through a lakehouse plus bronze, silver, and gold, where drinking from the bronze layer kills you because the water is dirty. The rooms include a Duke of Dax and a Jeff who wants Excel help, picking up the report on the desk tells a Pro user that 2.3 gigabytes needs a better license, and they will do a build-through if the comments ask for one.

• Data Goblins — Kurt Buhler is the Data Goblins name people already know him by, and the site’s line is that he helps people make useful things with Fabric and Power BI. Mike pointed the chat at two recent Kurt videos, one on building reports with AI and one on getting started with agentic development for reporting, alongside this site. Kurt said he is easy to find across socials, that he is about to go quiet for a few months on another project, and that he stays open to people who feel overwhelmed.

• Kurt Buhler on the Tabular Editor blog — Kurt is Head of Innovation at Tabular Editor, and he told listeners this author page is where most of his writing shows up. The page currently carries his guides on managing context for agents, picking a model for agentic development, and getting started with agentic development for business intelligence. On the show he pointed to a Tabular Editor series with Eugene and Ruben that walks through the building blocks as slowly as they can, and he asked anyone with questions to reach out.

Main Discussion

Topic: What it takes to have an agent build a Power BI report

Kurt’s premise is that coding agents such as GitHub Copilot and Claude Code can build lasting BI artifacts, and that reports have been the stubborn case. Semantic models already have tools. Reports are rigid, slow in the interface, and full of properties whose names sit far from the words a person uses. They reach that argument after Tommy’s read of the first piece in James Serra’s three-part series, What Does It Mean to Make Data AI-Ready?. Tommy’s takeaway, which Mike agrees with, is that clean data is only the start: duplicates, missing values, and inconsistent spelling still trip companies up, and an agent that hears “customer” still has to be told whether that means bronze or gold, a lakehouse shortcut, or the HR version versus sales. Tommy’s line is that an organization without a working data-governance practice should wait on AI. The rest of the hour is how you get from that ambiguity to a report an agent can build.

• Use the agent to build the thing people open every day. Kurt separates two jobs. One is a conversational interface over the data. The other is an agent that helps you create artifacts that stay in the platform: semantic models, where tools already exist, and reports, where the work has been harder. Mike has been on Microsoft for aiming the demos at answers. His use case is the report, the bar chart, and the system someone checks every day. Kurt’s boundary is the executive decision: exploring data with an agent is fair, and relying on the agent’s answer for a big call is the person’s job. Tommy hears the same split in the market. Ask a company what agentic AI for data means and the answer is still a chatbot. He wants the significant role to be development.

• The report format is an engineering problem an MCP server still has to solve. Kurt’s complaint about Desktop is the click tax. The format pane exposes the properties, and reaching them still means nested menus, especially on the newer card visuals, where properties overlap. Setting a number format on the visual and setting a format string in the model are different jobs that feel like the same job. Tommy puts visual design at maybe 30 percent of a good report and points at conditional formatting, plus the gap with something like Fabric’s Quest, which is largely TypeScript and sits in a language models have already seen. A bar color is “data point fill” on some charts and a different property on others. Kurt’s list of reasons a report agent fails is specific. The Power BI report format is new and thin in training data. Some things cannot be controlled. Some properties can be set and still fail to render, and browser control will miss a share of those. His advice through the week before the show was narrow: simple scenarios, brownfield edits, and a pile of templates. A nice report from A to Z was a bad time.

• That line moved between Tuesday and Sunday, and he does not pin it on one model. Kurt says a threshold crossed in the days before this episode. Designs he would not have attempted three weeks earlier, including ones he calls outlandish, were landing in half an hour to 45 minutes. He had stopped enjoying report work and would rather live in the semantic model than spend four hours on a card. Over the weekend he sketched an idea, handed it to Claude, and had the report in about 25 minutes. The longer path: a first success in October 2025, just before skills, from a context store of the report format; then the pbir-cli he and Maxim built, which is programmatic control and carries no design knowledge, so design has to be encoded in a prompt or an example; then, at the start of 2026, mocking layouts in Excalidraw, Figma, or on paper, and moving the build to the web once agents had better browser control. He had stepped away from Desktop around January or February. Asked what flipped in that one week, he listed tuning of the pbir-cli, a classifier he and Mike call Jev that he used to optimize the CLI and then dropped into the agent loop for classification, and a suspicion that he had been routed from Opus 5 to Opus 5.5. He tested that last idea on Codex and on Qwen, got the same results more slowly, and dropped the model swap as the cause. Once a few examples of success exist, more success comes faster, and his warning is to widen what counts as success or the agent settles into a narrow hallway.

• Requirements are the scaling activity, and the model for iteration can be a small one. For a report that stakeholders will use to make decisions, Kurt wants the same planning as before AI, done more carefully, because the documentation now feeds an organizational context store, the ontology, and the next mockup. He remembers fighting for one or two weeks of design-sprint time. That time is easier to justify when the notes pay forward. Rapid prototyping still belongs. The mistake he sees is reaching for the largest model when a report iteration wants speed and consistency. Opus low is where he has had the most success. He keeps the larger models for planning and for work that crosses many complex pieces. Skills, in his use, describe processes and workflows. A folder of finished examples does the visual teaching: build a report you like, set it aside, and tell the agent to reference it. He prefers a tool such as the pbir-cli to raw edits of the metadata, which he finds slower, more expensive, and more likely to be wrong, and he flags that preference as bias. Mike reads “scaling” as the moment the requirements exist and cheaper subagents can take a page, or six passes at a bar chart. He had felt, back in January and February, that agentic building had passed Power BI by. FabCon Atlanta and the work Kurt and Maxim are doing are what turned that for him.

• Five building blocks, then a short list of prerequisites. Tommy’s version of the same idea is the environment: planning, MCP, tools, skills, and the harness. A prompt, or a single skill, is half the plane. Kurt’s north star is the business user, and his failure mode is the person who keeps rebuilding the airport. He has done it. Prompting, for him, is how a session starts — a slash command, a budget bump, a loop, one agent prompting another — and instructions are context. The Tabular Editor series that started in August names the blocks as the model, the context, the prompt, the tools, and the environment. You need to know the options the way you know the options in Fabric, and you pick them up when a decision model or a loop is the right next move. Mike’s answer to the chat, which was asking which tool to learn while the stack keeps shifting, is that nothing has won yet and the patterns will come from people building in public. Kurt will not hand out one lineup. Memory files differ, agents differ, and he prefers the terminal, so an exact recipe will not reproduce. What he will require for reports is a Power BI project in the Power BI report format, and Git on Azure DevOps or GitHub, because a bad edit has to be reversible. Visual metadata changes by editing the files, or by code that edits the files, and that code can be a command-line tool or an MCP server. He knows of few report MCP servers and prefers the CLI. The agent then has to learn the CLI and your examples. With no example of a good report, defining good is the first task.

• Tommy spends the time in the context harness. Kurt keeps the agent out of the plan. Tommy puts about 80 percent of his agentic development in context. Notion is the second brain: master-sequence pages that Fabric and Power BI skills write, so a later Claude session is told which phase is finished and which instructions apply, plus a data dictionary read from the semantic models. On a large project he opens Claude Code by pointing at that page. He calls the build loop the execution harness, and he is wary of prompting straight at a project file and expecting the gains. Kurt’s planning rule runs the other direction. The agent does not write the planning documents. He has watched it bend a plan in small ways, so the plan stays his. The agent researches, challenges the opinion, and then carries out the work downstream.

• A palette demo is a proof. A business report needs the organization’s own examples. Mike loved the report Kurt sent that morning, where choosing a color palette restyles the page, and he asked how that becomes a business report. His wish is a shared library of how Kurt, Tommy, Mike, and a given company actually build visuals, so an agent can pattern-match a style. Kurt treats that as an organization problem. Design is context, the same way definitions are context, and it needs no special tooling: paper or a whiteboard is enough if it says how this company looks at shipments. His example is monthly shipments against budget and forecast, shown as an accumulating month-to-date line by workday. That note scales into report building and into querying, and he wants teams writing it now, in the open, with alignment, rather than waiting for the next project or sending a form that asks people to define shipments. Mike hears ontology in that description. His own picture is closer to linked notes, and he has found little value so far in Microsoft’s ontology tool for making agents better, with room for the tool to improve. Kurt is holding his opinion on Microsoft’s implementation until he has spent more time with it. He finds the argument about whether ontology sits inside the semantic layer often pedantic, and he is making ontology his priority for the next six months: the explicit documentation, and the implicit knowledge of where data comes from, how it is transformed, how it is communicated, and which questions people ask. That material has to reach the agent. Mike wants to keep going on the intersection with governance, data quality, and the enterprise semantic model, including a Fabric workload he has started called lineage view. They parked a bonus episode for ontology, and another for MCP and skills.

Looking Forward

Kurt’s close is that agentic development of reports is possible, and that it takes real effort to assemble the building blocks. Keep a Power BI project in the report format under Git, and spend the time on requirements and on examples your own organization would recognize. The person who stays with the business problem and with the people reading the report spends less time formatting cards and more time on the impact, which is his answer to the job anxiety in the chat. The Tabular Editor series with Eugene and Ruben is the slow walkthrough, and his invitation is to reach out, because the pace is too fast for everyone, including him. He also asked anyone who disagrees to say so, since the setup stays different for every team.

Episode Transcript

0:02 up the sky. Dance to the day to laugh in the mix. Fabric and A. I get your fix. Explicit measures. Drop the beat now. H feel the crowd. Explicit measures. Hello and welcome back to the Explicit Measures podcast with Tommy and Mike. Tommy, good morning. How are you doing? Oh, Mike, I’m doing great. How you doing? doing? Have you slept at all, Tommy? Because

0:32 Have you slept at all, Tommy? Because yesterday, Tommy and I just ran a game called Fabric’s Quest. If you want to go check it out, I’m going to make sure I put the link down below here. This is a game that Tommy slaved over for for many, many weeks., Tommy, give me the link if you don’t mind so I can put it here in the chat., yeah, it’s in notion but or it’s in our agenda, but yeah, I’ll put it in the chat, too. Okay. I just I’m trying to find it and I can’t put my finger on it. I got I got there. Okay, I found it. Thank you. All right. So, this is this is Fabric’s Quest. If you want something fun to do this morning, go play Fabric’s

1:03 fun to do this morning, go play Fabric’s Quest. This is built on Fabric Apps. It’s a fabric app. Tommy does have a SQL database behind it, and all the prompting that you’re seeing is coming from the SQL database. So, go check it out. If you like the game, if you got a couple laughs out of it, which I think you will, we hope you we hope you do enjoy those. Let us know in the comments if you’d like us to do a build through. We’ll build through the game. will show Tommy will come through and just show you how it was created, how he got it all figured out and what prompts or how he was

1:33 how he was using and building through the game. So anyways, let us know in the comments below if you think this is useful to you, if you had fun, explain what it is a little and then we get into it. So for those out there who have ever ever grew up in the 80s or 90s and do you remember those tech most of our audience grew up in that 80s and 90s? That’s I know I know I know but if you haven’t there was this idea back in the day of these fantasy like games where you would have to te basically use commands to go around and basically do actions for example it would say you’re

2:04 actions for example it would say you’re in a cottage you see a mug and you would say get mug you got a mug and you basically need to use those mugs and they’re very cumbersome those games they were always frustrating and combine that with this old thing called Homestar Runner which they’ve de it was this which was this card and cartoon on the web and flash I think and they did the same thing where they created a game called Peasants Quest but the idea was the game was very snarky at you thing and it was hilarious to me to me combining that with Fabric and PowerBI

2:35 combining that with Fabric and PowerBI the whole game is basically trying to get through because there is a dragon that is throttling your capacity and you have to figure out the way and there’s a ton of scenes that are all fabric or PowerBI based. I’m very proud of this Mike. We had the we had a data model. We had the the king of DAX or something like that. We went in the Duke of Dax. There’s a guy named Jeff who keeps showing up who wants you to help him with his Excel. We had to export Excel throughout the game. we went to the lakehouse. We we

3:06 game. we went to the lakehouse. We we went to the lake. and we also we also ran through multiple layers of bronze, silver, and gold. If you drink the water from the bronze layer, you die. It shows you it dies. It’s too It’s too It’s too dirty. It’s It’s not clean enough. And there are a ton of Easter eggs in there in terms of if you try to get something it’s very very comprehensive there where if you try to lift something goes you ain’t lifting that. That’s 2. 3 gigabytes. You’re going to need another license. Yeah. Tommy had created a report on

3:37 Yeah. Tommy had created a report on the desk and I tried to get report was one of the commands and said this this report is 2. 3 gigabytes. You don’t you’re a pro user. You you’re going to need a better license. Yeah. Yeah. You’re need I thought that was really cute. Anyways, fun game. Check it out. Links in the description and also inside the chat window as well, just in case you want to check it out. Fabric’s Quest on Pulia BI. I think you’ll have fun with it. That being said, Tommy, our main topic for today, let’s just quickly touch on that and we’ll do a little bit more news. Main topic for today is AI or

4:09 more news. Main topic for today is AI or agentic development of PowerBI reports. So, a lot of it’s a word, it’s a word salad., if if my family was listening to me to talk, they’re like, “What are you saying? What words are you putting together? They just make no sense.” So, we’re going to talk. We’re going to unpack these words. What does it mean to use AI to help you build a PowerBI report? What does this look like? Okay, that being said, Tommy, you’ve got one more news article here and then we’ll get into the main topic. Yeah, real quick here. So, this is from our friend from James S. James. James. Yeah, he’s been in the game a long time

4:41 Yeah, he’s been in the game a long time and he’s actually just started a new series. This is the first of three on what does it mean to have data AI ready, which I love this conversation because I don’t know if anyone knows and I think Jeff even says too that AI ready is not just clean data you and I don’t I don’t think we’re ever going to have this really holy grail. Obviously, there’s conceptual semantic models, but some of the key takeaways here, Mike, I’ll run through it is there’s data quality is a starting point. Your duplicate records,

5:12 starting point. Your duplicate records, missing values, inconsistent format and spelling, but that doesn’t mean you’re AI ready if you just have clean data. It’s awfully where, let’s be real, many companies are still struggling with this step. I know, I know. We’re still trying to get clean information and dduping records and things. So we actually have a proper list of a unique list of distinct customers like I yeah so and yeah the big part is key takeaways is AI means the business meaning this goes back to the idea for the semantic model creating a business

5:44 the semantic model creating a business vocabulary on your what your acronyms are or overloaded items like gross margin and general manager could mean GM descriptive names definitions and that you actually your concepts are precise terms that seem obvious may mean different things across departments. Mike, I was on a call yesterday, six different types of customer rollups. So that’s so important. How metadata is critical, relationships are critical, and you need governance. This is this is the problem, Tommy, that I see here with some of this. So So I like this article. This is a good article. Solid concepts.

6:15 This is a good article. Solid concepts. I do agree with a lot of these things. The challenge is there’s we’re starting to get some more of these general ter. So let’s you said talk you said you were talking to a customer and there’s many different definitions of what is a customer I talk about it this way you talk that way what does that all look like look like inside your business I think there needs to be like you can think of the customer multiple different ways that should be allowed allowed and that’s okay yeah it’s how we document that and where does that information go and if I think about

6:46 that information go and if I think about systems are being built we have some depends on what your system is some can be more complex than others Right. I’ll give you the complicated one. We have data bricks table for customers which we shortcut to lakehouse for customers which we put into semantic model for customers. And if you give access to your lakehouse for users to build on top of, we could have n number of additional customers tables in other semantic models that are owned by the business. So we’re now able to

7:16 by the business. So we’re now able to push or or bend the definition of customer to the will of the user of that model or what the business unit needs. This is this is the challenge. Do you lock it all down and say no, you can’t do that. This is the only way we look at customer and these the only fields that are in there. Likely someone will complain. So how do how do you balance flexibility with also standards? And so I think this is where I feel like it’s very difficult to say customer.

7:47 very difficult to say customer. And what the AI needs to understand is, hey Tommy, you said the word customer. Did you mean bronze? Did you mean gold tables like may better? Right. Did you mean the shortcut in the lakehouse or are you talking about these three models that have the word customer in them or the the customer table? Those are all still customer at a high level. But then there’s always like this second question that should come behind this is what customer do you mean? What what what are you talking about the HR version where it customer or the sales version or the

8:17 it customer or the sales version or the ops version? There’s a lot of different ways you can define customer. So anyways, that’s where I think this is going with the AI prep stuff. We have to be able to describe things in a way that makes sense for other and that’s where the data governance is so important. And I I mentioned this on a previous episode, Mike. If you do not have a semblance of data governance in the organization, don’t do AI yet because those issues were always going to arise. I I I am more and more solid on this fact that

8:47 more and more solid on this fact that part of the foundation for be an organization being AI ready is you have to already been doing data governance or have to have a very healthy data governance program because to your point with the six customers I I’ve talked to organizations with and like well why can’t we have a name for each one like c sales customer or member you sales customer or member rather than customer because of know rather than customer because of these different logic well no well we use customers like they’re all the same thing I Don’t go to a baseball game. A baseball is different than a basketball. Do why? Because we’ve given a

9:18 Do why? Because we’ve given a different name and it’s a different object if it is. That is a my own kind object if it is. That is a my own riff there. But we have to be able to of riff there. But we have to be able to get the co the company flexible to that to be able to name things properly. And it’s the ontology. It’s the conceptual side. It’s the semantics. It’s it’s a lot of talking between teams and aligning on what you think is the definition of this thing called customer or member, right? What what is this defined as? And to your point, Tommy, I think some

9:48 And to your point, Tommy, I think some at some point you can have a consortium of people, but if you always have a governing body to always manage every single term and definition, it’s going to take a while for you to get that figured out. Like sometimes people just need to make some sometimes people just need to show up and say I’m just going to make the decision. Here’s where we’re going with this is what we say. Use it, right? Some we need some of that sometimes. Interesting. Good article. I will put this article. Sorry, I didn’t put in the chat yet. So I will take this article. I will put this in the chat if you want to

10:18 will put this in the chat if you want to read more about preparing your data for A or what does it mean to prepare your data for AI. James Sarah’s got a great article here. Check it out. Tommy found the good read on this one. There’s some really neat and interesting points in here. helps you evaluate how far along you are. Are you in a place where you’re actually able to push a little bit more AI because you’ve got some of these foundational pieces set or at least maybe this will help pick on some of these more foundational pieces so you can better use AI. Good. 100%.

10:48 100%. All right, this being said, Tommy, I have a sad note. I have a sad note here. Oh, no. Oh, no. The sad note is today is our last day with Kurt. Kurt is back again for our final day. Sad. Yes, Sad. Yes, Kurt, welcome back. Sorry the news took a little bit of time here. Kurt, welcome back to the show. We are bringing Kurt Data Goblins,, as you may know him online., prolific writer, incredible thinker, doing all kinds of fun things., , thanks guys. Welcome. Welcome. Happy to be here. Happy to be here. All right, let’s jump into the topic for

11:19 All right, let’s jump into the topic for today. This this topic was highly suggested from Kurt. So, I would love for for this example here, Kurt, I’d love you to do the introduction on this one. give us a give us a little bit of runup here on what we should talk about today with building reports with AI. What are you thinking here? Sure. Sure. So, as we all know like AI is it progresses and we’re working with coding agents like we talked about this in the context of fabric data apps and other contexts as well but you can use agents like GitHub copilot

11:50 you can use agents like GitHub copilot or cloud code or other things to be able to facilitate the development of BI artifacts. So instead of using the agent as a conversational interface to be able to query your data, you can use it to help you build persisting artifacts inside of your data platform or elsewhere as your choice. So you can of course use it to build semantic models and there’s various tools that help with that. but for reports things have been canonically a little bit

12:22 been canonically a little bit more challenging. So as we all know who’ve created reports and we’ve talked

12:26 who’ve created reports and we’ve talked about this also that reports in PowerBI can be very rigid and inflexible. It’s very time consuming to go through the user interface. So it is very desirable to be able to cut down on that time to be able to make it a easier and more streamlined process especially as some of us have started to taste this already with HTML or with facilitating basically web dashboarding and we see like how you can have that

12:56 and we see like how you can have that pixel perfect control of the dashboard. So the question is can we achieve not quite pixel perfect control in a self-service BI tool but can we facilitate the development of the report using an agent so that we can remove the friction that isn’t helpful of the user interface and focus on things that are a better and more valuable use of our time. So that is a bit the subject. So,

13:26 bit the subject. So, I feel like desktop, again, I’m just going to poke on desktop here for a bit. I feel desktop has been very click tax heavy for quite a while, while the format pain is useful and it gives you all the properties that you need, there’s a lot of clicking to do to get stuff done. So, it’s if you compare it to something like Figma for instance, like where in Figma it’s driven to be able to make changes,, a lot more efficiently and to be able to make those bulk edits and it has come a long way. It has improved. So agree nothing there

13:56 It has improved. So agree nothing there nothing wrong against that. So but but indeed it is it is challenging and if you do have to like if you use for example like these new card visuals like it is very challenging to be able to go through all these different nested menus just to find properties and sometimes it’s confusing about like okay what you want to do but what is the property you need to change in order to get that and the properties are intersecting and redundant in ways that is very confusing. like for example setting the format of numbers versus doing it as a

14:27 format of numbers versus doing it as a format string in the model this thing. So it’s there’s there’s a lot of of challenges there that are not necessarily something that if you you necessarily something that if you if you get good at working with know if you get good at working with that and learn that that’s not like that’s necessarily going to help you deliver better insights or whatever like that’s just overcoming the challenge of the UI. So that’s a really good point. Yeah, Tommy, go ahead. go ahead. No, and I was going to say and you’re just talking right there just from the design, the the visual display or the

14:58 design, the the visual display or the visual properties, right? which is what maybe 30% of the battle when you’re thinking about having a good report. There’s obviously the data that lives inside of it and not just evaluation context, but I was watching your video you just posted trying to do conditional formatting is an issue there and understanding those concepts where building a game like Fab Quest like and I’ll bring that up because it is more or less straightforward TypeScript for a lot of it., it’s a little image based, which is more

15:31 it’s a little image based, which is more also too, there’s a much more volume of examples out there that AI can learn from. from. Yeah, Yeah, you’re dealing with a previously binary thing, a f a a language timole that we don’t have all these examples of people posting yet and we have these ambiguous properties. And I I love that you also mentioned the video bar chart shows fill for a color, but that’s completely different if you’re doing it in a different visual. So you have all this ambiguity that can cause an issue.

16:02 ambiguity that can cause an issue. Yeah. No, exactly. And it’s it’s it’s one of the reasons why basically that if you’re using a agent and you want to build a PowerBI report and you have let’s say you have a good experience with the MCP server or a command line tool and you’re like okay this is great for the semantic model. now I want an experience like this for reports. So I hear this all the time and people say well I just want an MCP server but the thing is that the reality is you cannot just slap an MCP server or something on top of PowerBI desktop or

16:32 top of PowerBI desktop or PowerBI reports and it will just magically work. There are some very specific and substantial engineering challenges to be able to make it work. Like you talked about the properties that’s one thing that the properties can be ambiguous. they’re not exactly connected with the semantics. So like the example you give, if I want to color the data bars, I have to say I want to change the bar color and the agent has to understand ah yes, that is data point fill. But that’s that’s only true for some charts, not all charts.

17:03 true for some charts, not all charts. Yes. Yes. And so but on top of that like the PB format is quite new. So it’s not like it’s,, rich in the training data and there’s a lot of things that you simply can’t control with PowerBI or there’s things that you can set, but then it’s not necessarily guaranteed that it’s going to render. So the only way that you can know that it will render is let the agent see it or you tell the agent and even if the agent sees it, it’s probably going to miss it,

17:33 sees it, it’s probably going to miss it, for example, with browser control or whatever. So there’s there’s a lot of reasons why it’s challenging. and I used to say up until last week I used to say that for PowerBI for agentic development simple scenarios make sense and like brownfield scenarios of adjustments and iterations make sense or if you have like a big pile of templates you can probably get a little bit better results push around some stuff. Yeah. Okay. Yeah. But but otherwise you’re probably better to start thinking

18:04 you’re probably better to start thinking in terms of just templates themselves like not thinking about AI necessarily but but something changed in the last week actually and where suddenly in the last week I’ve been able to make pretty much any design I want in PowerBI and suddenly there’s been some threshold that’s been crossed where I can create things that are even quite like outlandish ish like stuff that I wouldn’t even dare try with an agent like 3 weeks ago and now suddenly it

18:36 like 3 weeks ago and now suddenly it works in half an hour 45 minutes or whatever. whatever. Yeah. Yeah. So So yeah. yeah. So I want to just double I’m going to hang on that point just for a minute. So you just did a video Kurt around like an hourong video just kind around like an hourong video just a hey an ad hoc build and of a hey an ad hoc build and you just went through is this on YouTube? Is that it is you put it on data goblins? Yeah. So,, Kurt just did a really good video, so I got to give a lot of credit here, right? So, Kurt, you’ve been one, you’ve been pushing on this pretty hard.

19:06 you’ve been pushing on this pretty hard., two, I would say you’ve been really exploring heavily. You stopped pretty much using desktop as of when? When when did you really step away and say, I’m only going to focus on primarily AI building of stuff? So I would say probably January, Februaryish is when I really stepped away from it., I started experiencing like success in October of last year when around the time when just before skills

19:38 around the time when just before skills came out came out and I had built out this big context store of the PBI format and stuff like this and that was the first time I had success and then Maxim and I built the PBR CLI and we we were we were having success with that of course but the the reports that we were making using the CLI were not like particularly spec spectacular because it’s a CLI. It doesn’t embed any design information in there. It’s just programmatic control of the report. So you you have to encode

20:08 the report. So you you have to encode that,, in a prompt or in other ways. So, so usually what I was doing like at the beginning of the year when I wanted to create a PowerBI report is I would do like I would draw the layout. I would do a lot of like mocking it out first like an Excala draw and I would think through like the chart types and layouts and this is still valuable now of course during a requirements gathering exercise. But I just mean like for an ephemeral prototyping or brainstorming like I would never go to PowerBI first. never like back then

20:39 PowerBI first. never like back then I would start on the whiteboard or on paper or in Figma or Excaladraw whatever whatever and then I would take that and it I would do some of like the visual placement and stuff and then I would work a bit in PowerBI desktop and then go back but then at the beginning of the year I switched to doing everything in the web because I didn’t see a reason why I should use PowerBI desktop anymore especially because the agents got better browser control and And then it it just worked.

21:10 then it it just worked. Okay, I’m making I’m making I’m making faces at Tommy here. I’ve been That’s why I’m saying Okay, so I’m going to pick on your point here. So for many many episode one of the podcast, I said everything’s going browser. Everything’s going to be in the web. You’re going to be able to edit in the web. Everything will be web based stuff. This is 2021 as I’m calling it. I’m like this is going to be the future. And Kurt, here you are, a person who’s much smarter than I am, way more into this AI

21:41 smarter than I am, way more into this AI space than than and and studied than here than I have been. So, I’m gonna high I’m going to say Tommy after five six year six years I gain some credibility for what I said. First time. Yeah. No, no. I’ I’ve harped this a lot, but but I really do believe like I think your observation here like the web getting so good. not requiring a downloadable thing. It can go on any machine. you It can go on any machine., there’s computer use is going

22:11 know, there’s computer use is going through the ceiling right now. I just saw a really interesting and I’m going to open a can of words because I know we’re not going to go too far down this unless Kurt says we can. Jev, which is another model. No, no, no, no, no, no, no, no. There’s another type of model that’s not just so there’s another model that has appeared recently that is allowing us to be able to like make very cl very quick classifications at very cheap rates rates and I’m seeing people now throw computer

22:42 and I’m seeing people now throw computer vision vision at this new algorithm called Jev and so what it does it takes the whole screenshot almost in real time and it’s classifying everything on on the page and giving you bounded boxes and really really useful local model classification. So where agents when they go to websites they would just read the page. They’d look at the code and figure out what’s there. Now they can physically see pixels and items on the page. So it’s not they’re not looking at the page like we are with an

23:12 looking at the page like we are with an image but they’re using code to visualize the page with a lot of other data that’s going there. So there’s a really really I think we’re here at we’re seeing another inflection point. This this is what I felt probably back in like January, February when something happened and agents became I can use them for code. I could they became useful. I can build real things with them. I’m I’m feeling like we’re in another ramp up of this new world that we’re living in now as well. And I think Kurt, maybe what

23:42 well. And I think Kurt, maybe what you’re seeing here is,, part of this going to the browser, getting better results from the report with an AI agent. I feel that as soon as you get some system or some framework around this, this will be the only way people want to build reports. They they will never want to use desktop again. I I totally agree with Yeah. Go ahead. It’s it’s it’s crazy because I used to there was a time like a few years ago when I just stopped and I know this

24:12 when I just stopped and I know this doesn’t sound very nice, but like I just stopped enjoying making PowerBI reports. like I just I fundamentally did not like it and I would much rather be spending time in the semantic model and it was largely just because like I felt like I was wasting my time when I was having to dig through all these properties and I had in my head an idea of what I wanted and an experience and I felt like this would be this is what this user wants or this is what I want or whatever but more importantly what the business needs and but I can’t do it and or I I can

24:45 but I can’t do it and or I I can But then I’m literally going to spend four hours formatting a bloody card. And so I was I was so frustrated and but

24:53 so I was I was so frustrated and but then over the weekend it was I was enjoying creating PowerBI reports for the first time in literally years where I was just doing it just for just just for the laughs like and I I shared one with with with you guys just before where I just had some spontaneous idea and then I just had I sketched it out a little bit and then I gave it to Claude and I said go and it was there in like 25 minutes and it was that’s that’s been such a breath of fresh air and I think it’s already possible. But the thing is too like it’s a bit of a

25:23 thing is too like it’s a bit of a challenge because because of these challenges with the PowerBI format the PBR format, it’s not going to work out of the box for everybody. So what I’ve already noticed with a lot of people is people have this really diverse experience where some people have a horrible experience and some people like are able to make really good things and there’s no clear tutorial to follow. And that’s so so the thing is that this is inherent to the technology, right? Is it’s non-deterministic. So you can’t exactly

25:53 non-deterministic. So you can’t exactly make a tutorial and guarantee it will be the same for everybody. But I do I do have thoughts and ideas on how someone can start. Sorry, Tommy. Yeah. No, no, I think that’s I think that’s spot on because there there is obviously no clear howto guide that you can really ever create. You can provide a recommendation or suggestion. Yeah. But I think also too to your point, if you just figured out a lot, if this was this threshold happened on Saturday for you, we know we’re still early in the game. Even if this happened

26:23 early in the game. Even if this happened a year ago, this is too early to say this is the universal best way to Oh, yeah, for sure. I I think there’s some concepts there that absolutely work. However, I think this more of a I would think of it more like a recipe at this point in the ingredients that you need rather than having a technical like this,, install this, do this, and then it’s going to start running. Yeah. And yeah, I want to hear because I think from your video, Kurt, there are a lot of things I absolutely relate with

26:53 lot of things I absolutely relate with and nothing I at least to the point I got didn’t disagree with, but I think there are those elements, right, that really make this shine or at least set you up for success is maybe a better way to put that. So, so I think like the fundamental thing is if you’re doing a real report for a real project, right? You’re doing a BI project. This is for stakeholders and users who are going to use it to actually read and interpret data for decisions. You have to have a good planning and requirements process. And there’s nothing special or new about

27:23 there’s nothing special or new about that compared to the era before AI. You have to have your stuff together. You have to see and actually talk with the users and you have to start to document things. And now what is different is that documentation is so much more important. like requirements gathering is insanely important because it’s a scaling activity now. So it’s something where that documentation builds its way forward into the organizational context bank or whatever you want to call it and it’s it’s it’s what’s going to lead to help

27:53 it’s it’s what’s going to lead to help you build that ontology. It’s going to help you build the documentation. You’re going to create the mockups and you should be spending most of your time there. Whereas I remember doing these design sprints and stuff like this and having to really fight to be able to get one or two weeks of time to be able to focus on that. whereas I feel like now that is way easier to justify it because it pays its way forward so much more. more. but so so the planning is is

28:23 but so so the planning is is really key. But that’s not to say that you can’t do like rapid prototyping and stuff like that. And I think that’s where a mistake that I do see some people making and again everything I’m saying is from my subjective opinion and I’m not providing any like this is the way like this is just nor should we at this point it’s too early. early. So I see a lot of people who are like it’s not working. I’m going to go Fable Max or like Astra Max or whatever. It’s like like dude you don’t need that for PowerBI.

28:54 dude you don’t need that for PowerBI. Opus low is fine. like faster is better and iterate use like the big models for planning or like architectural activities that touch a lot of different complex things. Sure. But if you’re going to iterate on a report, you want speed and consistency. So opus low is for me where I’ve had the most success, especially with Opus 5. 5. Holy crap, man. man. So So awesome. Good to know. But great great but but you you you also need some kind but but you you you also need some tooling. Like a lot of people are

29:25 of tooling. Like a lot of people are also like focusing on like okay I’m going to create skills for bar charts and then I’ll create skills for line charts and stuff like this. And it’s like okay I get it. Like but for me skills are more about the processes and the the the like the business side of things and the workflows and stuff like this. And then just just keep a nice organized place where you keep your examples like build out a PowerBI report. Take your time and then once you’re like okay this looks good. This is going in the right direction then you

29:55 is going in the right direction then you document it. You put it aside and then you can just tell the agent like reference this example like this is in the direction that I want. and I do think that tools and I am biased because like we built the PBR CLI. I think the tools are better than raw editing of metadata because it’s slower, more expensive, and more likely to make mistakes., but, yeah, I’ll pause there because I’m starting to ramble. No, this is good. Yeah. Yeah. Many points I want to touch on. Tommy, you want to go first and then

30:25 Tommy, you want to go first and then I’ll come point. The only thing I was going to say is I would love to focus on that planning side, but Mike, I want to hear hear your thoughts. thoughts. So, that was that was where I was going to take. to take. Yeah, I was going to take that point there. So, things that stuck out to me, right? I wanted to unpack your comment around which I think I really agree with here which is getting requirements is a scaling activity. Okay, I want to I want to unpack that phrase. phrase. I’m also trying to sus out here. Is this another t-shirt that we need to

30:55 this another t-shirt that we need to have like my requirements scale? Maybe it’s a Marcelica. what is that in English? muscle shirt. It’s a muscle shirt. Yes. Exactly. Exactly right. So So getting requirements is a scaling activity. So, I I really like this idea and I think what I what I’m hearing you say in that statement that that that to me is there’s a lot to unpack there. How I read this is the fact that I can then spin up an agent or two agents behind me

31:26 spin up an agent or two agents behind me or have one agent think through with me requirements and what does this look like and then I can spin up three or four sub agents to go use these to your point Kurt Opus low right a less expensive model and now we’ve just get a lot of other release models Grock 4. 7 just came out Luna just came out with six like we’re starting to see these like lower-end token like there’s a really a really local models. Yeah, local models is also something that is in there, too. I’ll add that in there as well. I haven’t, again,

31:56 there as well. I haven’t, again, admittedly, I haven’t had enough time to play with a lot of local models. I want to do better there., I don’t have a big enough graphics card to pull, I think, the models that I want at this point., I have a 3090, but I’m that’s old. So, in general, I’m trying to like still push around these ideas of of what’s going here. the scaling activities. I like this because that means without those good requirements, I don’t have the ability of making an agent per page and having to go build stuff. I don’t have the ability to say I need six iterations of a bar chart agents go

32:27 iterations of a bar chart agents go figure out give me just give me like I know what I like to look at, but I necessarily don’t always know exactly what makes sense on the chart itself. And so I want let me I’m gonna I’m gonna be very honest here. Okay, Okay, by nature I’m always honest. In January and February of this year, I was so disheartened by the PowerBI system in general. So disheartened. I I was not enjoying it. It just felt like I was seeing all this proliferation

32:59 like I was seeing all this proliferation of like agentic development, agentic creation. And I really felt like I was looking at the world going, “P PowerBI just got passed by.” Like we we we lost we lost vision for where we were going and all this I’m creating websites. I’m building PowerPoint with with HTML. I’m like doing all these things. I’m making full apps. I’m building games. I’m like all this stuff is so exciting to me. And I’m looking at like my world. I’m like I I don’t want to be here anymore. This is not fun. And then for me, I think things started to shift a little bit inside

33:30 started to shift a little bit inside March with Fabcon Atlanta, which was a good thing. And we have been slowly picking up pace and now I’m very encouraged by Kurt what you’re doing with Maxim. I think those are those are amazing tools. Skills for fabric have come out has slightly helped. So that’s a bit more process related. Not exactly what we’re talking about here, but all of these things I believe are starting to help. Microsoft is starting to move the ship more towards we need AI to build the things. I don’t need AI to give me the answers.

34:01 I don’t need AI to give me the answers. And Tommy and I have harped on Microsoft a lot about this. There’s so many blog posts around this. There’s so many examples that Microsoft themselves are giving out. That’s the wrong use case for AI. Do not use AI to give you answers about your data. Use AI to build the report. Build the bar chart. Build the system. build the thing that you look at every single day in systems that are commoditized. We need we need to be cautious about this and I’ll I’ll mention here Tommy then Kurt. Yeah.

34:31 Yeah. like little asterisk do use data and agents to explore your data though but but but don’t don’t use it to like the executive is is relying on the answer from the agent like it’s like very fair. Yes. Sorry 100%. Yes, use data with your agents, but don’t ask it to give you answers about your data, right? right? For big decisions. That’s that’s you to think through. Tony, go ahead to you. Well, no. And I honestly that’s perfect because I think that’s the biggest misconception right now. When you think

35:03 misconception right now. When you think if I were to ask you what agentic development is, you’re going to give me a much bit different answer compared to a stakeholder or rather what does agentic data mean or agentic AI for data. And most companies are thinking still on this conversational analytics, right? Where it is a conversational chatbot. But I think we we all know here that’s a per like a role but that should not be this most significant role and I think the whole term agentic development

35:34 think the whole term agentic development for reporting is to your point Kurt I almost called you Keith is is changing completely. If I were to ask you a week ago what does agentic development for reporting mean compared to today? I gu does it sound like a different universe. It would have been a different universe. I would have literally told you it’s not possible. Like I would have said like if if I would have literally said like okay brownfield yes making changes making adjustments bulk activities but if you want to create a nice looking PowerBI

36:04 want to create a nice looking PowerBI report from A to Zed with an agent you’re going to have a bad time. If last Tuesday you would have asked me that I would have said that. So, and then suddenly suddenly in the in the span of that week from like Wednesday through to Sunday, like it’s like it like went in a straight line. Like it was insane. Like I was it was one of those What changed? Was it a model? Was it your prompt? Was it like like something had to have changed here? Yeah. Yeah. So, I’ve been analyzing it really deeply over the last few days, which is part of the reason why in that

36:34 which is part of the reason why in that like video I put out, I look like I’m like hung over. But you’re just so f you’re you you’re in the zone, the zone. I have a I have a nine-month-old kid and a toddler and I’m also like staying up super late trying to like figure out this stuff. But so so there’s many things. So I’ve been going like really hard on tuning the PBR CLI and stuff like this. So Jev came out and that was the initial incident is around the time Jev came out I was going

37:04 around the time Jev came out I was going through an activity where I was preparing for a project I’m about to start not a BI project but it is a project like parallel and so I was I was adjusting a lot of things in my agent setup and facilitating agentto agent

37:17 setup and facilitating agentto agent communication and agents that could call specialized smaller models. So that could be a factor but a very small likelihood. And then I Jev came out and I I used Jev I had an idea for Jev which is it which it’s too complex for me to get into here now but I I used it to optimize the PBR CLI and so I started working through that and that was taking quite a long time. So I I worked started working over that and then I start I

37:48 working over that and then I start I also incorporated Jev into the agentic loop where it can be part of certain decision- making and classification processes ad hoc. and then on top of that I think I started getting routed to opus 5. 5 when I was using opus 5. So however I did I did eliminate the possibility of that one because I started testing once I became suspicious of that. I started testing on Codex and with Quen and I I I didn’t get the

38:19 with Quen and I I I didn’t get the results like as fast but I got the same results. So so it’s something about the combination of those factors that that and and that and the crazy thing with agents right so this is also something that I really want people to understand is that once you start to have success once you start to lift off you you go you lift off like because once you start to have some examples yeah once you get some examples of what success looks like you get more success

38:49 success looks like you get more success faster faster and however ever at that point it’s really important to taper yourself and and pump the brakes because otherwise success becomes too narrow and and then success becomes like this really narrow hallway. So at that point you immediately have to recognize okay the the plane is off the ground so we have to pull in pull up the wheels but now is the time that you have to you but now is the time that you have to that’s where the critical piloting know that’s where the critical piloting comes in and you have to start giving it other other perpendicular definitions of

39:21 other other perpendicular definitions of success. So, it’s like, okay, we’re we’re having success here. We’re jumping over here now, and let’s let’s hit that point over there, and it’s on a roll, so it’s it’s going to be a little bit more likely. And once you hit that success in a few areas that are farther away, it’s like scouting a map. Like, did you guys ever play those games like Warcraft or Starcraft or Bunker? I’m not going to I am not creating a PowerBI game for that, though. I love those. and the dungeon crawler games like I did all those I’ve tried to recreate those with AI

39:52 I’ve tried to recreate those with AI that that’s been one of my tried to passions I want to I want to rebuild I am not doing a fabric quest we’re going back on the rails we’re going back to it we’ll come back to that come back to that come back that was your fault though though you didn’t know that was you induced this one this is your fault that was my fault that’s my fault I should have seen this coming but so so so there’s a fog of war right like so when I when I’m exploring an area with agents and I start to I start to sniffle. sniffle. I start to sniff that there’s success.

40:22 I start to sniff that there’s success. It feels like I’m on like a Warcraft or a Starcraft map suddenly where I’m like scouting it and I’m trying to find like where’s the enemy base? Where’s these little resources? And that’s for me mentally what it feels like. But which is why when I’m doing something with agents, I try as fast as possible like the fail fast mentality. as fast as possible figure out if you want to spend time on this map because if you can kind time on this map because if you can immediately figure out, okay, we of immediately figure out, okay, we shouldn’t explore this map, you move on to a different map. But then then you start exploring it and once you find a

40:53 start exploring it and once you find a few things on the fog of war and you like find some like radio towers and it illuminates the area around and then it suddenly like most of the map starts to get filled out and then you’re like, “All right, let’s go.” And then it’s and then it’s on. Now we harvest a ton of stuff and I make a hu huge horde and I just start dominating the rest of the map. That’s how this how the game goes. And watch out for the zergs. Yes. Completely. So, but I I love team human not team or I like the team human. human. Taran taran taran protos and zerg. Okay.

41:25 Taran taran taran protos and zerg. Okay. Yeah. Yeah. Canon. Canon. I I want to go back a little to the the plane analogy too because what I’m hearing and it’s driving very well with my own experience when it comes to what honestly it’s your AI environment or your agentic environment that you’ve created for yourself. Yeah. being and that structure that you have to sense to drive the plane and the same way the planning the MCP and the tools the skills right the harness that you’re using these

41:56 you’re using these I think for a lot of people they just start running with a prompt and or they may have a skill that they’re using using but you’re you’re flying only half the plane you’re only doing half the things I think that are necessary yeah so speak to how important that environment setup is I guess it’s it’s it’s also important to know that like the goal is not to get the plane off the ground. The goal is to fly somewhere. So like we have this tendency in the PowerBI community to really lock in on tools and be like oh man like this tool

42:28 tools and be like oh man like this tool this this MCP or like this this new feature like amazing and stuff like this but like okay right yes okay but like where are we going with this? like the goal is how do we how do we how do we build the thing that’s going to make the impact for the people, right? so so so that’s that’s that it’s really important that that is the northstar like the people you imagine your business users and you you hang up their little picture behind your monitor and like but but then indeed you you

42:59 and like but but then indeed you you do have to build the airport for the plane to take off like you can’t just have a John from like we’re gonna the plane we’re gonna just the plane is going to take off there’s some magic to it like it does require work and engineering and and design and thought and it requires trying but at the same time you shouldn’t your goal isn’t to build the airport like I I I have seen a lot of people in my ecosystem who have spent like they just spend too much time iterating on

43:29 just spend too much time iterating on the airport and they forget that they have to fly a plane and so and I have I’ve done that too. I’ve done that lots. I’ve done that too. I think we’ve all that’s normal. Oh 100%. but but there are things like yes like prompting I I don’t really find is that important personally but like for me prompting is just when you start the session but like I refer to everything else as just context like how are you like your instructions and stuff like this like for me I never think of those as prompts. Like for me prompting is am I going to use a slash

44:01 prompting is am I going to use a slash goal? Am I going to say ultraink to like bump the budget? these like little things., and, or am I going to let the agent prompt another agent or am I going to set it on a like a cron job or loop?, but, like it’s the building blocks like in tabular editor, we have this series that we’ve been starting since August which is walking through the building blocks of agentic development. So those building blocks are what you use to build your airport. Like if you if you take those

44:33 airport. Like if you if you take those ingredients, so it amounts to the model, the context, the prompt, the tools and the environment that encompasses those things. So if you consider all five of those things and you don’t need to know everything, but you need to know just like in fabric or PowerBI, you need to know what’s out there, what are the options, when do I use that, when would I think about that? And you don’t need to go into like super detail. I’m an expert on like this or that, but you just need to know like, okay, this is probably a time for me to

45:04 okay, this is probably a time for me to think about like a decision model or this is a time for me to think about like a loop or whatever., and and and that’s, you , and and and that’s,, in my mind, in my subjective know, in my mind, in my subjective opinion, which could be totally wrong, like how I think of this airport construction, let’s say, deep on that metaphor. metaphor. Deep on the metaphor. Well, and Mike, do you have anything there? Obviously, I don’t want to I don’t want to steal anything from you. Well, I I just want to There’s so much to unpack here. Sorry.

45:34 Sorry. So, no, no, this is good. Good. This is great. I want to un I want to So, someone in the comments her her hip is commenting here and the comment is really Let me just summarize this comment., we talked about,, Jev and then we talked about all these different other tools. We have Kurt’s MCP CLI. Maxim’s working on something with you, Kurt. Like there’s a there’s a whole bunch of tools that are here, right? right? How do which one to use? The whole setup is takes a little bit of time to adapt and as soon as you

46:05 bit of time to adapt and as soon as you get something set, it feels like it shifts again. It’s it’s shifting. So my comment really here is the tech right now of how everything’s being built. We’re in rapid acceleration and development mode. I really think right now everything is very fluid and it’s very difficult to pin anything down right now. So if you’re if you’re just now stepping into this, be prepared for a fluid conversation. There’s going to be a lot of decisions to be made. Not any one of them is right at this point. We haven’t really landed

46:35 at this point. We haven’t really landed on what’s working the best out of all the tools. Then we lean a little bit on the community to be honest, right? This is where we we really do pull in the community to figure out what these patterns should look like. Kurt, you’re learning things. you’re building stuff, you’re making your own tools. So, we leverage it’s you let people explore and discover and we have to learn from that from and from those then we start building patterns and cycles and we can then start flying planes to destinations as opposed to just trying to get it off the ground.

47:05 to get it off the ground. Yeah, indeed. To you, Kirk. And so, so this like circles a little bit back to what I said is like this is one of the challenges with agents is it’s very bespoke. It’s different for everybody. And one of the reasons why I’m very stubborn about like not telling people you should do this is because they have different memory files. They’re using a different agent. I,, I prefer to work in the terminal. They might be working somewhere else. So even if I give you an exact lineup, very simple lineup of what you should do, I can’t guarantee that your results are going to be comparable because of these

47:37 going to be comparable because of these factors. And that was already something like back before AI when someone would ask like how should I do this with my semantic model it’s like well I don’t know your business or data but this is roughly what I would do now introduce all of these unknowns and all of this change with agents it’s even harder and I really sympathize and empathize with people that are like getting really frustrated of like okay how do I start but but but so the best thing you can do is to just start and you can also work with your agent and you can discuss with your

48:07 your agent and you can discuss with your agent and you can just get a lay of your lands and you can like survey what’s out there., so so for example for reports like there’s you example for reports like there’s we know for sure the report has to know we know for sure the report has to be you have to have a PowerBI project you have to have the PowerBI report format. So that is a mandatory prerequisite. So it’s also strongly recommended that you use git Azure DevOps or GitHub because you want to be able to save the changes and see the

48:37 able to save the changes and see the changes and revert the changes when you have a boo boo and you will have a boo boo that is just the reality. So then if you want to change the the visuals you’re changing the metadata and there are two paths to do that well three sorry. So there is directly modifying the files or modifying the files via some code which could be a command line tool or it could be an MCP server and currently there to my

49:08 MCP server and currently there to my knowledge there’s not many MCP servers like we have a command line tool that is my preference yeah and that is that then you you iterate and you iterate from there and the last thing before sorry I’ll hand it to you Tommy is is that you context. So you need to know like discuss with the agent and say like okay I have this PBR CLI thing familiarize yourself with it it and I have some examples like familiarize yourself with those examples but if you don’t have any good examples of what a good PowerBI report looks like

49:40 of what a good PowerBI report looks like then that’s probably going to be your first task is you need to define good

49:44 first task is you need to define good like what is your expectation and in some way so sorry I no it’s great so I I well Mike actually We found out how many MCP servers Microsoft has for Oh jeez. Well, I don’t know now. Actually, let me Kurt, what do you think the number is? How many Microsoft G? 36. 36. Okay, Okay, good answer. Good answer. Probably give it at the end of the year. It’s probably going to be up there. If you if you just check like if you just check the GitHub accounts from like

50:15 just check the GitHub accounts from like people at Microsoft sometimes. I’m like like there was one time where I was like someone published something and I forgot who it was, but I was like that was really interesting. who did that? And I used my agent to like survey people from Microsoft to see like who published this. And the agent came back with the kinds of things people were publishing. And I was like I was like I was like I’m not sure if all of this is supposed to be published on GitHub. Did you do something you’re not supposed to? to? Yeah. So that’s that’s a bit of the

50:46 Yeah. So that’s that’s a bit of the challenge with having a lot of agent stuff. But yeah, but but there’s there’s a lot of MCP servers. Yeah. So, I’m going to back at last count. So, Matias and I have been doing a separate podcast which is called Agentic Thinking. Matias said at the time of our podcast, this is probably about two or three weeks ago, there were eight separate MCP servers. Okay. PowerBI MCP server modeling server. Well, less less. Well, I don’t know now., not the fact that they’re building things with agents anyways. Now, every a now every MCP

51:16 anyways. Now, every a now every MCP modeling server, sorry, now every MC is getting two of them. There’s like a remote one. Like there’s one that Microsoft hosts and now there’s one that you get. So like you have the MCP modeling server, but is it the remote one or is it the local one? Is it the one that you install or is it one that you use their servers to do? do? So there’s actually a lot of I think flexibility now coming with this. And it it’s good. I think it helps. We also need Jev to help us pick which MC. MC. Okay. All right. All right. So I’m going to go. All right. Enough with Jeff. Build a Je. So you want to talk Ila instead? Is Laya

51:47 you want to talk Ila instead? Is Laya what you’d like to talk about Tommy? you what you’d like to talk about Tommy?, know, another version of Gen. We’re just going to keep saying names. Well, before you go on that, Mike, I I want I want to make an argument about the time spent in Agent 2 development, where you spend your time. And honestly, if I trying to write it down, I would be curious what your own workflow is, but 80% of my aentic development is in that context area. So I I’ve spoken about this on the past where and we were talking about this a

52:17 where and we were talking about this a bit offline. I use notion where that’s that stores all my context. It’s that second brain but it also stores what I call master sequence instructions for cloud where if I’m working on a major project we need to do data discovery we need to do semantic model building. There’s going to be report design. I’ve set it up in such a way that it’s using the skills for fabric custom agents for fabric for PowerBI in notion that actually write what’s called a master sequence page that will tell Claude, hey, we’re

52:49 that will tell Claude, hey, we’re already done with this. Go to this page. This is the new instructions that you’re going to use for this session, especially on larger projects. And I spend most of my time in that development of those instructions because it can be comprehensive. we can use blocks and code formatting and with the notion MCP server the I honestly go into cloud many of the times now more often than not saying look at this page in notion the master sequence for project X

53:19 master sequence for project X and let’s begin to dive into it let’s make sure that we’re on the same page and that what I call context harness harness is where I spend most of my time anything else where I’m actually doing the work is what I’ve been calling an execution harness and that loop I think of agentic development to your point with going back to the beginning the requirements the context the meetings that we had let’s feed this all into what we are actually going to build and actually be able to relate that to what the metadata is because no notion

53:50 what the metadata is because no notion also stores all the data that’s in that semantic model has a data dictionary based on reading the semantic models so it has all the things you can read read and relate to to me I would be weary you can absolutely just start prompting in in cloud code and to build even if you’re using a a project report file but I would say if you really actually want to see gains I think that setting up that environment

54:20 think that setting up that environment is essential yeah agree it is it is important to think through your building blocks and it’s like the same thing it’s like you want to make a cake so you need ingredients to make the cake and you don’t necessarily like need to have all the exact measurements and you’re not going to be a master chef. But the point is like you just just try to bake something that’s edible and and you’ll get better as it goes through. But you’ll also develop preferences and those preferences will lead you to having your own

54:50 lead you to having your own distinct workflow like like Tommy like for example you talk about like having an agent really heavily involved with the requirements process I think if I understood correctly. Yes. Whereas mine is is very much like the opposite where I I don’t let the agent touch the planning documentation ever. Like those are no limits. Like get out of the house. Like those are that’s my house. because I I never let the agent do the writing of the plan because I’ve noticed over time that there’s so many ways that it can just derail the plan in

55:21 ways that it can just derail the plan in very subtle ways. So, I will make sure that it understands the plan and I’ll use it to facilitate research and stuff like this and to to to adversarily challenge the plan. And that’s something I like to do in general is when I hold an opinion is to just have the agent like challenge the opinion. But and then and then to to facilitate the execution downstream. So,, so yeah, but,, so I guess what I’m really hearing you say is in all this situation, you should

55:52 say is in all this situation, you should really be using the Department of Common Sense. If you’d love to get a shirt, we have made it. Oh my god. Wow. He’s been waiting for that for 40 minutes probably. I’ve been trying to put the sentence on the keyboard. You need a full cut. Yeah, Yeah, I better get royalties. This is just literally totally a joke., it is a real shirt. You can go buy it. It is it is available on our store. But that is totally a joke. Last last episode we talked about the the

56:22 episode we talked about the the phrase of the department of comments. But to to be here, right, we don’t know. So going back to your point, Kurt, I don’t want to derail here too much. Right. I love all these things you’re thinking about. I really like this examples. examples. One area that I wish there was a bit more from the community. Okay. So this is one area that I think would be useful here. We need more examples that are agent ready. And what by this, I think there’s a need here around you’re you’re finding great success with building different things. You sent us this PowerBI report this

56:52 You sent us this PowerBI report this morning. I absolutely love looking at it. You’ve got this really cool select the color palette that you want and the whole report changes. Awesome. Super fun. Like love that. But when I look at this and I I step back and say, “Okay, business report.” Yeah, it’s not it’s not a business report. So how how do we start? So what we need is we need to start collecting the the collective knowledge of Kurt, Tommy, Mike, building visuals in your organization. These are common examples of how we build these visuals. And then you can literally say agent because it’s

57:23 you can literally say agent because it’s really good at finding patterns and pattern matching. I want this style visual. Here’s an example. So So in I will push back on that a little bit and I will say this is not a community activity. This is an organization activity. Just like how an organization needs to get their stuff together with context, design is context. So when you talk about your organizational context store or ontology or whatever whatever to me personally a part of that is also

57:53 to me personally a part of that is also defining how you communicate and visualize data. And there’s nothing technical that you need for that. Like you can draw that on paper or a whiteboard, but you need to document that like okay we have shipments. We typically look at shipments on a monthly basis. When we look at shipments we compare to budget and forecast and what how we expect to look at that data is for example in like accumulating month-to-ate line chart by workday. So

58:23 month-to-ate line chart by workday. So this is all essential information that again is a scaling activity that compounds for not just building reports but querying the data and building other things and this is the stuff that happens in requirements gathering but it doesn’t need to wait for a new project like you these initiatives should be starting now everywhere where people are going through and they’re documenting this process and making sure that it’s surfacing and that it’s getting alignment., and definitely that it’s

58:55 alignment., and definitely that it’s not like just you send a Microsoft form out to a bunch of people and you’re like fill in what is shipments., , I’m going to open a can of worms here. I know we’re at time so I do want to be mindful of everyone’s time here. Isn’t what you’re describing Kurt ontology? ontology? Oh, Oh, to me it is. Now it is. How Microsoft has implemented ontology? I’m not sure I agree with it. I my version of ontology feels more like an obsidian to me. Like

59:25 obsidian to me. Like so so notes and related ideas and concepts in a graph. That’s what it feels like to me. What Microsoft gave us in the actual ontology tool feels totally not what you’re talking about and doesn’t really help. I have found very little value from using the ontology at this point to incorporate into business and giving and giving any additional insights to agents that are making it them go or work any better for now. Yeah. Doesn’t say that the tool may not get better and they’re going to,, change it so it works better.

59:56 change it so it works better. What do you think, Kurt? I don’t have any like opinions on the Microsoft ontology right now because I need to spend more time to develop those opinions, let’s say. But I do have I do have very deep and intricate opinions about ontology as a general concept. Like, of course, you do need to be able to convey the relationships between various entities and stuff like that. And I I’m aware that there’s a lot of confusion about ontology versus the semantic layer and is ontology part of

60:26 semantic layer and is ontology part of the semantic layer and all of this like to be honest like it feels often quite pedantic. But I think for me the ontology is incredibly important. So incredibly important and to be frank for me it’s also going to be my number one priority in the next six months. So so so to be to be clear about how I feel about it personally. So,, it’s we don’t have time to dive into it, but but indeed the onto bonus episode if if everyone if everyone gives us a, 000 likes on this video,

60:56 gives us a, 000 likes on this video, they’ll do whatever the YouTubers do. Give us a thousand likes on the video and Kurt will come back and talk about ontology. Make sure you play D and D game live. So, also also I’m gonna I’m gonna be very clear like I have never played a D and D game. So Reed Reed has been trying to convince me for I think two and a half years now to do a live stream D and D game where I I host it, but because parenting is hard, I haven’t had you don’t have eight hours of or six

61:26 you don’t have eight hours of or six hours of your time to prayer and play the game. the game. It takes so much longer than that. But yeah, but I so I promise you and I also promise indirectly that therefore Reed we will sometime before the end of this year I think we should do a live streamed D and D oneshot. So,, that would be a data themed,, live stream D and D one shot. Tommy’s already got the,, Fabrics Quest part of it storied out for cartridges, man. We start USB USBs at at

61:58 cartridges, man. We start USB USBs at at conferences with a bunch of games. So, maybe closing remarks just on ontology before we maybe start to think about wrapping up the agent development or reports topic. So, ontology for me is

62:09 or reports topic. So, ontology for me is very important. And what it boils down to is that you have you have explicit information about your data and documentation and all of that. And I think we all have a pretty clear understanding of what that is and why it’s important, but you also have implicit information throughout your organization. And that implicit organization, yes, is semantic in nature, but it’s also coming down to the fact of like, okay, where do we get this data from? How are we transforming this data? How do we communicate this data?

62:41 data? How do we communicate this data? what kinds of questions are we asking about this data? Like all of that is information that somehow needs to be able to be surfaced to the agent experiences and needs to be needs to be something that that that could potentially be,, involved with the ontology. So it’s it’s it is it is information that needs to start be being mined out of the organization in a way that will support people to do more valuable tasks. so so that is

63:11 more valuable tasks. so so that is how I feel about it at that point without going too too much into the too deeply into the fog of war. So while we while we are also here on the podcast I will say I want to continue to unpack this concept. Kurt, I’m actively working with like ontology like conceptual things in addition to the enterprise semantic model. There’s there’s a really interesting intersection between data governance, data quality. Yeah. Yeah. The ontology and then the enterprise semantics layer. Like there’s there’s a really interesting story that’s

63:41 really interesting story that’s happening there. they fit together and harmonize really well or they should they should and I I think that’s the point the the concept is it should harmonize really well but the tooling and the things that we’re getting today is not doing a great job of this and so I’ve already started down this exploration group of I’ve made a workload it’s called lineage view or lineage it’s inside fabric today it starts the fundamentals of what this is doing to some degree and so I’d love to continue unpacking this with you Kurt either on the podcast or another episode I do think we need to have a

64:12 episode I do think we need to have a bonus episode for you at some point in the future. I think around talking about MCP skills. Yeah. Yeah. Yeah. No, Yeah. No, totally agree. Awesome. Let’s wrap this AI agentic space. Kurt, I have linked your videos so there’s two that you just did recently and I feel like we’ve maybe have like I don’t know what it is. The algorithms now know that you’re on the podcast. It sends me now every single post that you were doing. So, thank you. I’m now getting more of your content which is good. Also, I’m also seeing a lot more of your videos come out. So there’s two videos at the top of this list here in the chat

64:42 at the top of this list here in the chat window. I will also try to make sure I include them in the description on the YouTube video in case you want to see them. One is building reports with AI. That’s one of the videos you just did. And then getting started with AI agentic development and reporting. So those two videos, the blog post from your website,, data goblins, I put them out there. Make sure you everyone who’s listening to the show as we wrap these are going to be core principles or starting points of like how we are going to unpack this. This is probably not the

65:12 unpack this. This is probably not the end- all beall solution but everyone is trying to sus out what does this look like? How will this work? At the end of the day all I can say is I already like building things with agents much better than doing it myself or clicking buttons myself. So, the more I push into agent building things, the agent build, I’m going to call the agent build. I’ve been trying to coin the term for this one, but it’s it is the agent build, that’s exciting to me. It’s fun. I enjoy it.

65:43 exciting to me. It’s fun. I enjoy it. Everything should be agent built. Yeah. Yeah. Anymore. So, that’s that’s going to be the all-consuming noise I think we’re going to hear for the next couple months as this stuff continues to mature. I’m not going to Barcelona, Spain. I will not be at the Microsoft Fabrica Fabric Conference in Barcelona. I have to imagine there’s going to be some language around how this is going to start playing out at Barcelona as well. For sure. I think final thoughts, Kurt, and then we’ll go to Tommy. to Tommy. One thing I really want to make clear to everybody because there’s still, again, I know we talked about this before, but

66:13 I know we talked about this before, but like a lot of people are having anxiety about like, is this a risk to me and my job?,, things are changing very quickly. Look, you do not have to be afraid. You do not have to be afraid of your job. Your job is not going to be automated. AI is not going to take your job. If, again, if you’re the person who’s focused on solving business problems and engaging with the people who are going to be the ones consuming your reports, you are going to become even more important because this like I said like requirements gathering,

66:44 like I said like requirements gathering, it scales across everything and it’s be it’s super super important. So you will probably spend less time making reports. Yes. But that means that you can spend more time being able to facilitate the impact. And if the the building the reports is the part that you really enjoy that doesn’t mean that you’re not going to enjoy doing the other things as well because you can make a huge impact. So aside from that so agentic development of reports I

67:15 that so agentic development of reports I would say is something that’s certainly possible. I would say it’s something that’s challenging. So, it’s not trivial and it does take some effort to assemble your building blocks. If you want more information, we are creating this series at Tabular Editor to try to walk you through with Eugene and Ruben as detailed and slowly as possible. They’re also geniuses, too. It’s very good content. I love working with them. Yeah. And so, but if you have questions, if you need help, don’t hesitate to reach out. it’s it’s a very open friendly

67:45 it’s it’s a very open friendly community and we are all overwhelmed. It is going too fast for everybody. I am overwhelmed. I feel I feel like I’m behind. We all feel this way. So it’s okay. okay. And also to to your point there, I’m going to jump on this one as well. If we feel like we’re behind and we’re the ones in the tech reading the stuff, Jev comes out two days later we’re playing with the tool. Like we’re very forward for forward thinking in in AI. If we’re feeling like we’re behind, everyone else is probably feeling miles behind where we are. So, don’t worry.

68:15 behind where we are. So, don’t worry. We’re all in the same boat. We’re all gonna keep pushing into it. Just keep What I will say is plug yourself in. Go f go find a real person. Honestly, go find real people doing real things on the internet and go follow them. Don’t listen to this AI slop. Don’t let people random rip out blog. I can’t stand the amount of junk that feels like people are just writing things and shelling it through an AI and it doesn’t feel like it’s adding any value. Go find real people doing real things, teaching real stuff on the internet. Find who you like, hang out

68:46 internet. Find who you like, hang out with them. Kurt is one of those people that you need to go hang out with. And when you need to take a break, you And when you need to take a break,, play a game, relax, maybe, I don’t know, play a game, relax, maybe, I don’t know, online with Fabrics Quest. No., Kurt, I want to go just I would love to hear where we can find you. I love having you on. We loved having you on. I think we had some great conversations, but honestly, it’s been an honor to do this with you. I’ve been reading your blog for years now, but where can we find you?, I am active all over the

69:17 , I am active all over the place. So, various socials. So, if you just search Kirbieler, you’ll find me., I am about to go a little bit dark for a few months, I think., so where I’ll be working on a different project, but I’m always always available. So, Tabular Editor blog is where you can see most of the stuff that I’m putting out and I do plan on putting out more YouTube stuff., so, but I’m very open. Yeah, I’m very open to helping people and especially if you’re someone who’s

69:47 and especially if you’re someone who’s like you’re really feeling overwhelmed, you feel like you’re at your wit’s end or whatever and like you you you feel free to reach out. I’ll do my best. If I don’t answer right away, I’m struggling to be a parent. So, that’s why No worries. I We We all get it. Especially with young ones. Young ones are much more intensive than the older ones. ones. Listen, what Mike and I have talked about for years on the podcast, and this was all before AI. How the heck you built your blog, man, because you didn’t do that with any AI with the drop downs. So, that was the most impressive.

70:18 downs. So, that was the most impressive. It was all Yeah. Handwritten PowerBI tips is no longer is no longer handwritten by Mike Carlo anymore. It’s all static web apps. I’ve I’ve definitely but I will say all the content I do push out to PowerBI tips is 100% reviewed by me. It’s it’s I am a lot and a lot more of what we’re doing on Power Tips is the full transcripts of this. I’m committing to doing a lot more speaking, talking, engaging. The only thing I can really do that I can guarantee it’s not AI written

70:50 that I can guarantee it’s not AI written is be on camera talking to you guys, having real conversations about real stuff. So that’s my commitment to the community is I’m going to continue to participate here. We’re going to do more things on YouTube. We’re going to continue to explain things. I want to continue the engagement. I’m going to do so here. so that’s where if you want to find me and what I’m working on learning on, it’ll probably be on YouTube and on media. And one last thing that I’ll also put out to people is is if there’s something that we say that you disagree with or that you have an opinion

71:20 with or that you have an opinion about about then then speak up. speak up and say and and say like disagree more because we need more diverse opinions from different people and it’s it’s good to when people disagree because it helps us refine our opinions and your opinion is valid and if you bring a different perspective we need to hear that more. So don’t hesitate if you disagree or you have a different opinion throw it out there. We need more of that. I love that. That’s absolutely amazing.

71:50 I love that. That’s absolutely amazing. I do agree with those comments and those sentiments as well. I hope we’re mature enough to be able to take good corrective criticism and at least diverse thought and evaluate it, right? I will have opinions about things. I will find things that I like or don’t like. But,, we’re all allowed to have opinions. We’re allowed to have some preference here. And I think, you some preference here. And I think,, the collective Tommy’s changed my know, the collective Tommy’s changed my mind on on very few some things. But,, it , it something something took a while. There’s been a thing somewhere. I don’t remember what. No, no. I’m I’m totally joking there.

72:20 No, no. I’m I’m totally joking there. Tommy has very much impacted and and changed my line of thinking. I think hearing from really smart people and pushing on them. We’ve done a number of of people bringing them onto the podcast, not interviewing them, but talking about real concepts of what things are people working through. I think this is so valuable and that that diverse public thinking space is not very encouraged or I don’t see very much of it happening right now. I want more of this. This is this is I think valuable to the community. Mike have been saying this for 566 episodes and I still believe it that if we had no one

72:50 still believe it that if we had no one listening or even a minus one usage, we would still do it because just talking it out too. So and if you do disagree, Mike, I know where people can fi disagree with. Yes, Tommy. If If Yeah, exactly. as as Tommy puts on his his final, his final shirt here. So, if Tommy’s going to land us on some common sense here and Tommy will say, “Look, where can you find maybe Tommy, where can we find the podcast?”

73:21 Tommy, where can we find the podcast?” You can find us on Apple, Spotify, wherever your podcast. Why is there a 42?, because it’s the meaning of life. life. The number 42. Oh, okay. Apple, wherever your podcast, make sure to subscribe and leave a rating. It helps us out a ton. Do you want to disagree with us? Do you want to disagree with Kurt? We’ll let him know. Go over to power. tipsodcast. Leave your name in a great question. And finally, join us live every Tuesday and Thursday to be with us more on all Power Tips social media channels. channels. Awesome. Thank you all so much, Kurt.

73:52 Awesome. Thank you all so much, Kurt. Wonderful episodes. These four episodes have been very encouraging. So, probably some of the best content we’ve ever made. I by far been super fun. It’s been a lot of fun. Agree. And and thank you very much for your time. I know this is a big commitment away from your family at the end of your day. You’re tired. I get it. Thank you so much for the commitment. I do appreciate this. It’s it’s heartfelt. Thank community. Thank you so much as well. Appreciate you listening. We’ll see you all next time.

74:22 Measures pump it up. Tommy and Mike lighting up the sky. Dance to the day to laugh in the mix. Fabric and I get your explicit measures. Drop the beat now. Steal the crowd. Explicit measures.

Thank You

Want to catch us live? Join every Tuesday and Thursday at 7:30 AM Central on YouTube and LinkedIn.

Got a question? Head to powerbi.tips/empodcast and submit your topic ideas.

Listen on Spotify, Apple Podcasts, or wherever you get your podcasts.