chapters transcript notes
click any line to jump to that moment in the video
0:15 All , we're back again with another agentic thinking podcast. Matías and I are here. We're going to just go a little bit different format than we have in the past. , if you've seen podcast in the past, typically it's all words talking, that's all that's 0:25 there. This one's going to entail some demos. We're going to do some screen share. , if you're listening online or on Spotify or wherever you're listening to your you get your podcast things from, we want to encourage you to go to 0:37 YouTube and go watch the video. It's going to be in portrait mode, nice easy to see area. And we're going to do some screen share and we're going to talk and interact with a demo that you have planned for us, Matías. . 0:48 The topic for today is going to be all around GitHub Copilot . and using the Power BI MCP modeling server. Give give us the full name. You have a full description of what we're 0:58 doing. , I had to shorten the title for the video. What was the real name you gave us here? I was very descriptive. , I I was going to show how to set up the the 1:10 Power BI modeling MCP server to work with GitHub cloud agents, which is something I'm using a lot for my daily 1:20 development. And it's not super easy and straightforward. That's why I thought I'd share sorry, I hope you didn't hear my ring 1:31 door bell here. [laughter] [clears throat] , I I thought I'd share, , the hoops I had to jump through and and show how 1:41 exciting and new this experience can be. And I'm going to be also be very note here. This is you've been chewing on this idea for not just it's not 1:52 a one that's a little difficult to set up, but you've been chewing through this workflow, how to do this for months . you've been you spent a long time thinking through 2:02 different patterns, which patterns would work. And this is a a pattern that you're finding that works , it's secure, and it lands what you need to do from a workflow standpoint. 2:14 This is a a pattern that you're evolving from your expertise and your experience here. And I want to highlight this is we're all learning this together. I feel a lot of times I'm just educating myself. I love 2:24 learning from people in the community. People are still trying to figure out what's the path, what's the pattern in some of this. And we got to tease out these ideas because sometimes we'll build something that wasn't secure or that didn't 2:36 really work with my workflow. And , I think this is going to be really informative for everyone. Anyways, we'll get into it. Before we get into the main topic, let's do a couple news items. , 2:46 Yeah. there's been some announcements this week. We want to talk about this first and then we'll go into the episode. first and foremost, we've got to talk about Opus 4.7. Matías, have you played with it? Have you turned it on? 2:58 I haven't played much with it because it was just announced yesterday. I've heard quite a few things about it. I've definitely seen it in in my Claude 3:08 Claude and in my GitHub Copilot. very exciting. there was a lot of noise recently around people getting 3:19 quite upset with Claude experiences and claiming that performance and and thinking capabilities were 3:29 apparently degrading. I can't really back that, but there was a lot of press about that. Lots of people expected that would be the the precursor to some new 3:39 model coming. , it's finally out . but the real thing I thought was worth mentioning was not much Anthropic's announcement, but 3:51 the related GitHub announcement which also came out yesterday. There we go. they're announcing that Opus 4.7 is available as part of GitHub 4:03 Copilot. And if you scroll down here, this one is very interesting. , they're saying it's launching with promotional pricing 4:15 and a 7.5x request multiplier, which is pretty wild. , that's the promotional pricing is the 7.5x? , it's going to be bigger 4:26 than that moving It's , , just to just to recap, , where we're at in terms of those multipliers, Yep. 4:36 most frontier models, particularly the GPT ones, as in Codex, have a 1x 4:47 multiplier , ? ? , which means for each request one of your budget premium requests is being used. 4:58 The outlier far has been Opus, which has had has had a 3x multiplier. Yes. we also have some mini models which have 5:10 a 0.33x multiplier, Haiku and GPT-5 Mini. no, not a a 54 Mini. And then 5:21 and then and then for paid accounts, we also have some mini models that come have a have a zero multiplier. they come for free. , that's pretty exciting. But , there 5:34 was also a time until that was deprecated recently where you had a 30x multiplier for Opus fast mode. I'm 5:46 not sure who's ever used that to be fair. and whether that was worth it. , 7.5 is outstanding, particularly 5:56 obviously the suggestion is that it's going to go up, ? , is it going to be 15x? in any case, I think the learning to take away 6:07 from that is AI is becoming more expensive, , the I think the free lunch is definitely over or is going to be over very 6:17 soon. And , those are signs of this happening. Interesting. I That's an interesting concept there. I I would argue with you, yes, but I I 6:28 think if you talk about the premier models, ? Or the big players, ? Opus, Claude, there's other ones. Gemini 4 was 6:38 announced a couple weeks ago. And apparently, that's pretty easy to run and can run on much smaller machines. And then I think at the same time again, we're talking about very big premier 6:48 models, Quinn 3.6, I think, was just announced or something another one as . These are models that are a fraction of the cost to run. And , I think there's 6:59 there's a bit of a story here. The large players in the space want the game to be about compute. The game for them has to be about the compute story 7:10 because they they harness, they pay access to the compute that they have access to, ? If you bring in a whole bunch of open source models that you can run anywhere you want, even on your local box or machine, that takes 7:21 away a lot of that big corporate power away from these big larger models. , this is going to be interesting to see where this goes. I think I think there's a movement that's coming that's going to 7:32 have to start slowly shift this in a different direction. I don't think we can keep getting bigger, bigger, more expensive models all the time. At some point, companies will say, "That's too much. We we're not we're not willing to 7:42 do it, ? We won't spend we're going to block this model because it's just not cost-effective for us to run it, ?" Anyways, My takeaway was slightly different, . Oh, really? . , 7:54 last time, if anyone followed episode one, if you haven't, you should definitely catch up on it. last time we we talked a fair bit about Agent Harness, ? And and 8:05 how I I think the real takeaway from that should have been. I don't think we were really clear about it. 8:19 the real takeaway should have been it's not much the model you're targeting that matters. It's much more the harness you're using to target that model, ? You could you could for instance, you could be 8:30 using Anthropic's Opus 4.6 or 4.7 for some of your tasks, but depending on whether you're using that out of Claude code or out of Copilot or out of some 8:42 open source coding agent harness, you're going to get , very different multipliers. You're going to get very different experiences. And I think moving forward what matters is 8:53 that people pay attention to the harness and to harness engineering a lot more. particularly when it comes to using 9:03 specific kinds of models as as part of a a bigger agentic workflow, Yes. Don't throw everything at 9:13 Opus. have a have a good routing system in place that uses, for instance, Opus, , for certain tasks, but then Haiku and other smaller models for 9:26 other tasks. And I think this is moving forward where you need to spend much more attention and much more engineering depth in figuring out what 9:38 that that particular setup looks because otherwise, you're going to be you're going to be burning money, ? If you're if if you're going to be doing everything, , with a 7.5 or 15x 9:50 multiplier on on the latest and shiniest Opus, you may get great results, but it's going to really cost you. This is I think my point. I see the 10:01 industry moving more to what you're saying, real-time model context routing, ? There's a router in front of your prompt. Send prompt in, the routing says, "This is a really 10:12 complex problem. We're sending to Opus 4.7. That makes sense, ? Oh, you just said, 'Hi. How's it going?' ? That doesn't require a very massive GPU. Let's use a 10:23 free model and respond back with this additional context." , I think there's to your point I was watching probably months ago someone speaking and educating about harness, and they were 10:33 saying roughly 60% of their performance, they felt against this a a fuzzy number, ? Roughly 60% of the performance of a model came from the harness that was in. You can take the 10:44 same model and put it to many different harnesses, GitHub, Claude Code, something from some other thing, build your own harness, I don't care, ? A CLI solution. That harness really 10:57 impacts the performance of what the agent can do and not do to the user. The the the effectiveness of the user. And I didn't know this. Did there's a whole benchmarking website around harness? I did not know this. 11:08 [clears throat] Yeah. And GitHub isn't even on the list. . There's there's another what's the company Forge? Forge harness? Forge Code? Code Forge Code, exactly. They they lead 11:18 this this one by far, absolutely. Even Claude Code only comes within the first 20 or . absolutely. Yeah. Blown away because I didn't understand 11:29 Bench that what you're talking about. And interestingly, I I I I just read about that yesterday. , it's a massive coincidence that you mentioned. . [laughter] apparently our feeds are very 11:40 linked in the same way. , I'll put the links here as . If you want to go check out another harness, which I'm we'll maybe have to touch test it out on the program here as at some point in the future. But the other 11:50 harness we're talking about here is code or Forge Code is the harness. , I'll put that here as . And then also the other the testing for benchmarking of harnesses, what was the name of that 12:00 again, Matthias? Terminal Bench, if I'm not mistaken. Yep. Let me quickly Terminal Let me quickly Terminal Bench 2.0, I think. It's called Terminal 12:10 Yeah, that's the one, 2.0. 2.0 leaderboard. It had a had a great leaderboard talking about what harnesses are available to you and then how you can measure yourself, what harnesses are good, which ones are 12:21 working . And there's a whole bunch of them. I did not know this existed. just just being aware of this is something that you need to be able to say, "I know what a harness is, and here's how I can go score them and 12:31 figure out which ones are good or not." Anyways, any other news items we should cover off on? We've already talked a lot about this first news article. And we want to get to the demo here shortly. Yes. just briefly, 12:42 the Fabric team announced the OneLake MCP server Yes. I haven't played much with it, that may be a demo for another day, but just wanted to 12:53 mention it here. It's on the Fabric blog. It's all And , this one is GA. it's I landing GA stuff. That's great. 13:03 Absolutely. has it has there been public preview on that? I heard about it for the first time. Interesting that it would come out as GA 13:14 straight away, but there we go. Maybe I missed something. , there's much stuff being announced constantly on the Fabric blog, maybe I just missed it. They probably threw Mythos at it and said, 13:25 ", check it. Is it good to go? Yep, ship it." Let's just make it GA. We're done. We're not going to add any new features. Let it rip. I'm I'm cool with it. Anything that's GA, I'm happy with. That means Microsoft's going to 13:35 support it. It's going to stick moving forward. , I think this is the this is the idea to move forward with. All . Let's with that. Let's go into screen sharing. , if you don't mind, Matthias, let's go into the demo piece. Show us what you're building 13:46 here. Explain through us maybe a high level, what are we trying to accomplish here? What are you going to show us here? What are we looking into? . , I'm a huge fan of using 13:58 cloud using asynchronous cloud agents. not not sitting in a TUI, not sitting in a terminal, , where 14:09 you're instructing your agent interactively what to do, and then you wait for something to come back and you react to it, etc. etc. I'm talking about 14:20 creating very substantial and beefy PODs and then sending them off to an agent that doesn't even run on your own machine, run somewhere else, , on on on 14:32 their own infrastructure asynchronously, and then informs me when when they've when they're finished by creating a PR for instance, ? 14:43 this is this is a a working pattern that I've adopted over the past few months, and I'm very happy with it. And it means that you can scale much better , 14:55 the the interactive terminal experience with an agent means you you you you write a prompt, you 15:05 send off a prompt, and then you have to wait possibly, , 5, 10, 15, 20 minutes for something to come back. But you never know when that is, ? 15:15 it it it may be half a minute, it may be half an hour, you don't know. , how do you decide whether you should just sit and wait or whether you should 15:25 context switch, , to another task. that's the big problem with interactive development. , hope that is 15:36 a useful in terms of introduction here. I GitHub have released their cloud 15:48 Copilot agents a while ago, very mature product . I love it, use it all the time, use it extensively. and I wanted to show how 15:58 to combine that with Microsoft's Power BI modeling MCP, which we all love. in conjunction with an 16:08 announcement that's on my screen . Because the modeling MCP server was recently released as an NPM package, which means 16:19 it's very very it's much more straightforward installing the MCP server on different platforms 16:31 and bringing it into different coding agents. Previously, it was only released as a VS Code extension. , if you wanted 16:41 to use it outside of VS Code, you had to do a lot of hacking. That's no longer necessary. And yeah, we can we can share the link to the NPM package. It's currently 16:53 0.5 beta 3 and was last updated a couple weeks ago. , with that one, I wanted to show this repo, which by the way, I'm 17:04 going to make public. just need to, , add a bit more read me and all of that. , in this read mode read in this repo, 17:14 I don't really have much in place as you can see, ? There are two files here, and they don't really have anything useful in them. But what I have done on it is I've set up the 17:26 modeling MCP server. And I can switch over to the agents tab. if you're not familiar with that, this is your interface to 17:36 Copilot cloud agents. And I'm going to send off a task here, and that obviously that's going to take a little while. And in the meantime, whilst we wait, I'm going to 17:46 show what setup I've done for it. ? , while you're doing while you're doing that, I'm going to interject here a little bit. This GitHub experience, this is an agent 17:57 running in another machine, a VM, something that's on their runtime and not on your local machine. This is one thing I had to get my head around. This, in my opinion, and maybe 18:08 Matthias, you can correct me if I'm wrong here, this is another form of a harness for an agent. A harness a website-based harness. You have a chat box, you have some, , controls about picking what repo you're attaching 18:20 yourself to, what code does it have access to, all these really great things. But this is a new a separate harness. And what I'm trying on the on the same note, I'm as I'm unpacking this, I 18:30 have new language in my vocabulary. I want to interact with an agent here through github.com. I also want to have an agent on my local VS Code 18:40 harness. I also want to be able to work on it on my mobile device. , one of the advantages of this harness here is this agent session will begin, and you'll have access to it on the web browser here as as the mobile app 18:51 for GitHub, which I think is pretty dang sweet. And you can Absolutely. chat with it in both places once the agent is in the cloud, which is kind of cool. , just want to point that out, too. Totally. And , when I 19:04 made the point about, , asynchronous cloud agents, that is one of the big advantages. , you can be on on on public transport, you can you can be waiting 19:19 for a bus somewhere or anything. As long as you have your phone, you can fire off your agents. , this one happens to run on the GitHub Actions 19:30 infrastructure. , if GitHub Actions, if the VM model behind it, and and also how GitHub Actions runners are configured, , they they come 19:41 very conveniently with with a whole range of development tools pre-installed, it's the exact same infrastructure you're getting here. , what I'm doing here is I'm saying, 19:51 "Connect to the sales model in the Git demo workspace and list all tables with their row count, ? this seems fairly innocent, but there's a 20:02 lot in here that the agent needs to understand and translate into real actions. But before I send that off, just going to show quickly over here, 20:12 if you haven't if you've never used you know GitHub cloud agents, we've got this drop down. I'm showing that particularly because it 20:23 links very nicely back to our introduction. , you can see the 7.5x here. If you if you if you choose to select Opus 47, you can 20:34 also see how there's a large 1x selection including Sonnet 46. I because for that reason I use Sonnet 46 a lot 20:46 in when it comes to GitHub. powerful and does a really good job of being a lot of reasoning even does a good job there. I really it. Yep. 20:56 Also to point out a GPT 54 which is definitely very very close to Opus when it comes to coding is available here 21:07 with a 1x multiplier. it's cheaper than Opus and more importantly in GitHub GPT 54 comes with a 400,000 21:17 token context window whereas you only get 160 for the cloud model. cloud models are more expensive and you 21:27 have a substantially smaller context window. Think about that. . let me fire this off and 21:37 you have a really nice UI here. this is initializing and I can click on that. I get a dedicated URL. 21:47 If I bookmark the URL, I can always go back to that exact same session and it's starting a VM in the background. It's installing various things including 21:59 the modeling server. We can come back to it in a little bit. What I wanted to show was Before you leave, real quick. I had a question come in from the chat 22:09 which I think is a really relevant question here. We've been talking about these premium requests and you showed the multipliers and things. Yeah. If we're using a a lot of this is 22:19 obfuscated away from normal agent usage in a harness. When you go to cloud code, you just pay a subscription and you just talk to it and it says, "Hey, you've reached your threshold. 22:29 You've sent too many tokens." in other languages or other worlds, we're talking token count and here we're talking in premium requests. Can maybe you just speak very briefly around 22:40 You have at the very bottom the the window's moving away from you , but can you scroll down to where it says one premium request there? It was showing 22:50 you a little note there. It's showing up here . they're up there. . There you go. This is you sent a message into it. Yeah. It's thinking about it and it's doing a lot of generation of stuff behind 23:01 it. , server's thinking, writing text, , you can see it outputting its logic as it's thinking through these things. The premium requests and the 1x, 3x, 7.5x 23:12 are just telling you this message I sent in that's what it's using as the premium request. Correct. the 23:22 GitHub co-pilot billing model is very interesting. Really interesting because it's a real steal from my point of view. 23:32 They may change that at some point because I don't think they're making any money with that, but I'm on a pro plus plan which is 39 bucks a 23:44 month. nothing and I get 1,500 premium requests. Which means I can send off I I can fire off those kinds of cloud 23:54 sessions 1,500 times in a calendar month and this one here is very trivial and it finished already as you can see in under 2 minutes. However, it 24:07 frequently happens that I send off very beefy prompts and they take 25, 27, 35 minutes and I'm still getting billed only one premium request. , 24:19 economically this is awesome because it's very very different from for instance the cloud code model, ? In cloud you're getting , you've 24:30 got a subscription and it has an unknown amount of usage in it, ? there's no transparency at all with respect to how Anthropic is 24:40 calculating your usage. It is somehow token based, but you don't know what the formula is. It may change every day or by hour of the day. You don't know. All you 24:52 can see is how what percentage of your current 5-hour window you've used. And how you got there is completely unknown and is 25:03 completely intransparent, ? Whereas here I'm using Sonnet 4.6. I'm sending a a request. I'm starting a session. I'm I know that 25:14 one premium request is going to be deducted. If I had chosen Opus, it would have been three premium requests. If it if it was Opus 46, if it was Opus 25:24 47 would have been 7.5 premium requests, ? it's it's a different bill. I just want to be clear about what this is happening here and to be also very honest, this is all tied to the 25:34 subscription you pay for Copilot, GitHub Copilot. And there is a $20 plan and there's a $40 plan as of . The $20 plan gives you 300 a month. I think 25:45 I'm on the $40 plan. I get about 1,000 a month and I think I misquoted myself there on LinkedIn. it's it's not 10,000 on your plan. It's only 1,000 25:55 that you get on the Yeah, you get 1,000 when you're on business. You get 1,500 when you're on a individual one on a personal plan. It's the same price, 39 bucks, but 26:08 on a personal plan you get more. Yeah. With business it's it's it's with business you you 26:19 you you get one pool. And that is filled with the number of seats you have. it's very nice because 26:30 you you're not forced to the the the 1,000 requests are not allocated to one particular person. 26:40 They're allocated to your business. If you have 100 seats, you get 100,000 premium requests. Who's who's going to use them within your business is irrelevant, ? it's it's a 26:51 different model. Nonetheless, . We've talked enough . it's a real steal I would say because one, 27:01 you get to use the really with an arguably very good harness, the co-pilot harness getting better every day and also with 27:12 very very neat UX integration into the GitHub experience. All . Should we see what this one has done? 27:22 it says setting up environment and then in here we can see at the bottom Here we go. Start Power BI modeling MCP server 27:35 version 0.5 21 tools and then those are the tools listed here, ? I'm going to show you shortly how I've set this up, but at this point we can 27:46 see the MCP server was installed correctly and it's in scope here. And then the second thing is it 27:56 translates my request to connect to this particular model into a connection operation and as you can see down here, that's the input and 28:07 this is the output. And I've got the XMLA endpoint for that workspace and I've got the connection string for that particular model. 28:17 it's translated my my prompt and it also successfully connected using the credentials I've provided. it's listing tables. Here we go. 28:30 And as you can see each table from the list table operation already comes back with column count. it doesn't even have to do additional 28:41 DAX requests to find out how many columns are in there which is why the whole session finished after 28:51 not even 2 minutes. there we go. that's that. Let me just show you what's needed to configure that because sadly that's not super straightforward. in 29:01 that repo I've I'm going into settings. Then over here I've got co-pilot cloud agents. cloud agent is the technical term for what we're using here. There are lots of things you can 29:13 switch on and off, but the main thing is at the bottom there is a JSON document that 29:25 allows you to configure any number of MCP servers and the the format here is quite similar to you 29:35 know, the various MCP.json configurations in cloud code etc. etc. And the one thing I wanted to point out here is 29:46 I'm using three secrets to specify tenant client and client secret for my credentials and 29:58 what you need to know is that first of all any environment variables that you want to make available to your MCP server they need to be explicitly 30:08 specified. But most importantly if you want to reference a secret from your GitHub configuration those secret entries need to be prefixed 30:19 copilot_mcp_ If you don't do that they will not be visible here and sadly this is something which is not very prominently 30:30 configured. you need to first of all we've got two layers of indirection here and secondly we need to know that this prefix is necessary otherwise it 30:41 will never be passed to the MCP server. the whole co-GitHub copilot setup is very restrictive from safety 30:53 first point of view. you need to you need to configure your firewalls you need to make sure that a certain 31:04 secrets for instance are explicitly passed into your copilot environment. And that's that and I'm going to share that publicly in a readme 31:14 you don't have to figure it out yourself and then the only other thing is for those secrets to become available you need to set up an explicit environment and you 31:26 which has to be called copilot and then down here you can see the the three secrets we've referenced earlier they are listed here again. If you don't have 31:37 the copilot_mcp_ prefix it's not going to work. Wow . Very specific on that that account. Good good to know. This is the tricks 31:47 that you need to need to hear you can get it to work correctly for you. given that this one finished quickly why don't I give it another prompt 31:57 I'm going to give it a very very brief one but one that will turn out to be extremely powerful. look at that. I'm going to say please review the model and 32:07 propose suggestions for improvement. and I'm just going to fire it off and this is going to consume another premium request. Yes. 32:17 ? this whole exercise has far cost me two premium requests although that's two out of 1500 not a big deal. And Another part I want to 32:28 really hang on here too ? A lot of I think when we talk about MCP servers and what we're used to thinking about mental model here ? We're used to thinking about it has to be on my VS 32:39 code on my machine there's a server physically being spun up there. This whole solution here is allowing me to lift this work directly to GitHub and 32:49 the power of this I think this becomes you can interact with this in the website you can interact with this on your phone you can see sessions on your computer this democratizes where 33:01 the work can happen and I think maybe your I'd your impression on this one as Matthias by moving this away from having to sit on my machine this opens up a lot more 33:12 automation . possibilities where we can have other agents interacting with this or doing some automation that comes into 33:22 here and we can start kicking off GitHub actions and other things that are automating additional subsequent business jobs on top of this ? This is to me one of the hooks of that 33:32 I'm exploring a lot is wow GitHub actions are powerful they can do a lot of things and I can keep the secrets away from the agent but the agent can use them ? it it's to 33:43 me this all feels a lot security first better design and I'm really excited to learn more about where this is going cuz this is really powerful it's it's just chewing away 33:53 here it's just making all kinds of things. Yeah there we go. We can see it's doing all sorts of tool calls and this is all coming from not 34:04 even a full sentence ? it's it's clearly inferred my intention very nicely it's looking at measures relationships 34:14 connections security roles and columns and looking at display folders MDX 34:25 availability perspectives cultures hidden columns. I'm expecting some good outputs here I'm not sure how long that's going to take but we'll probably want to 34:36 wrap up at some point but if you're up for it we could just give it another 30 seconds or and see I don't I don't think we can get I don't think we can 34:46 bait people along this far and not let them see the results of this this it's saying giving giving you all the build it up to this amazing thing and then all of a sudden you don't you don't get the request out 34:57 here. [snorts] let's let's keep talking about some of this as . this is a workflow this is specifically for Power BI MCP . The main let's I'm going to hang on a a 35:07 note here. This works because of the NPM package. for example and the reason I'm I'm extrapolating some things here there's other MCP servers we just talked about the one lake MCP server. We 35:19 also talked about or we've heard about the there's a fabric real time MCP server. There's also a fabric MCP server to do make items in fabric. . None of 35:31 those we are aware of that has an MPP NMP blah blah blah NMP package for those sorry NPM I got it wrong even after I tried to 35:41 say it . this is very specific to NPM this particular server. there's I just want to be very clear about what that means and 35:51 if you want this for other parts that Microsoft is tooling you're going to need to pick on Microsoft and and ask for these features to be made cuz this is an additional 36:01 thing and into above and beyond just the regular Power BI MCP modeling server ? This is above and beyond. some of that a few things to unpack here ? as 36:13 far as MCP servers are concerned they fall into two brackets ? You can have HTTP MCP servers that work as a web service and you can have local MCP 36:25 servers that run as a local process on your development machine. for the HTTP ones it's very straightforward they work 36:35 everywhere because the endpoint to it obviously is in the cloud and all you need is to configure with reference and possibly some API key. the ones that run locally are 36:47 much harder and because you need to make sure that you somehow acquire the binaries that it works with your 36:57 developer machine architecture and that you start that process correctly that your harness can communicate with it ? And the the modeling 37:08 server happens to be one of those you know it has to run locally and this is why the NPM 37:19 delivery model really helps here because if we go back here and if we look at the configuration what's happening 37:30 here is we're using NPX which is the NPX package runner. NPX is not 37:42 normally you use NPM install for instance ? Which 38:00 and this one happens to download the package from the NPM registry and then invoke it straight away using those three arguments dash 38:10 dash start dash dash auth mode and dash dash skip confirmation skip confirmation. and this whole model of being able 38:21 to use NPX and and acquire a package and run it immediately that makes this bit of 38:31 needing a locally running MCP server much more easy. can you also speak to one thing that's also on here that I just recently learned about tools at the bottom there. There's a section a property called 38:42 tools and there's a little star a text star that's being shown up in there. this NPM server has a handful of tools that it has access to it can do things. 38:52 you you went earlier back under the NPM package documentation or when you ran the tool initially there's a whole list of tools Yes 21 of them yeah. 21 in the tools area you can add 39:02 additional security elements to this You can restrict it. No deleting there's there's you could you could start pulling the tools that you want it to use. I don't want you looking 39:12 at partitions. there's a partition tool and you can remove that from the knowledge of the agent that it's not using that as . that's another really good yeah here you 39:23 go. Yeah. Yeah there we go. this is there we go. Look 21 tools and you can for instance say I don't want 39:33 you to be able to do anything around calculation groups Yep. or or calendar or security roles and then you just 39:43 drop them from what's allowed, ? Correct. Yep. And that's that little tools list is all the available tools. it's a star option, which means all tools are available to the 39:53 agent. You can use whatever you want. if you want this is another security piece. The reason I'm bringing this up is because in, ? You need to You need to explicitly allow what the agent is permitted to do. Yes, 40:05 exactly. And that's my point there, which is this there's additional security controls as you set it up during that little JSON file that helps you specify 40:15 which tools you're using. shall we see what it's done? Yeah, what it do? It has I want to see a timing here. 40:25 There's no timing, weirdly enough. I've used another premium request. And this time I got a lot more back, ? I got something about 40:36 measure naming consistency, display folders. It's saying 29 measures have no display folder and that should be changed. Consider 40:47 grouping them. And then it even suggests what those display folders should be called, ? Wow. We've got something around using the 40:58 round function inside a DAX measure, known bad practice. This should be done using a format string that you still have the complete accuracy in the return value. 41:11 Rounding is is is a display is is is a reporting concern. It's not a calculation concern, ? Stuff that. I could go on and on because it goes all 41:22 the way to 11. Oh, look at that. We even got some praise. Things done for reference. How How cool is that? 41:48 for instance, we we could grab this whole thing. We could make that a an issue in that repo and we can then assign that issue to the same Copilot 42:00 Cloud agent and then things become really interesting because then you know, those Copilot agents become team members that can work alongside. 42:11 This is human agents, ? I'm not going to show that because otherwise we're way off, but this is where things are headed and this is 42:21 what everyone should be aware of. I I really want to emphasize this because this is lifting work you would then hand over to BI or modeling developers 42:33 and you're giving it back to the agent. The important thing here is tracking what a problem is, getting it logged somewhere and seeing the work getting complete against it. 42:43 to me, this just just reinforcing all the more story. Everything you build, models, reports, all of this stuff must exist inside Git 42:55 repos. And I'm going to choose for this reason, ? I don't see this level of agentic work and harnessing happening in Azure DevOps at this point. Maybe there's something there. I don't see it 43:05 being this robust. this clearly is choosing. takeaways from this kind of session really quick here are definitely get your stuff into GitHub. That's where you want to be playing. You 43:16 definitely want to be really seriously looking at GitHub Copilot. It is really powerful, a really interesting pricing model, a great deal for you to get started and do a lot of work with it. And 43:27 you've got this amazing integration between agent work, issue building, and fixing that with tickets and other items that you can directly task agents back to go do. This is incredible. I 43:37 absolutely love this. . With that, Matthias, let's wrap it here. I'll take the screen share off. Great demo today. Love this. This is the stuff we're going to continually unpack on agentic thinking. We're going 43:48 to keep doing news. The next episode we're going to do is next Tuesday at 10:00 a.m. Central Standard Time. We're going to do another one purely talking. News, what's happening, what are we 43:58 learning? We're just going to talk through that episode and then we'll probably do another demo on Friday again unpacking. Let's get hands on keyboard. Let's go to our screen and go check out 44:08 things and build things in there directly. Any final thoughts, Matthias? to to the audience, , if there if there's anything in particular they want to see more demos about or 44:18 let us know in the comments and , we'll happily accommodate that if possible. Love it. We're going to be building some other things coming soon, you'll be able to interact with us 44:28 directly. You always go interact with us on LinkedIn or YouTube. We'd see those chats directly and we'll and we'll be able to get those comments and we'll respond to your questions if you have any. Also, more things coming. Hopefully 44:39 we'll get you a mailbox form and you can submit questions and things that as . Thank you much, Matthias. Love this session today. Super fun. Appreciate you. We'll see you next time. 44:49 Thanks very much. Agentic Thinking [music]