chapters transcript notes
click any line to jump to that moment in the video
0:14 Hello and welcome back to Agentic Thinking. We have some more demos for you today. We're very excited to continue down this path. Last week, if I'll do a quick recap, we were working on using the F Microsoft Fabric MCP 0:26 modeling server as an NPM package directly inside GitHub. And we're going to go a little bit further with that. But before we get going, Matthias, welcome. Yes, hi. It's been another week 0:36 in AI land, hasn't it? it's unbelievable. [laughter] Even though it's been a week, it feels a month. . Indeed. how are you doing? 0:46 I'm doing really . There's a lot of new things happening. I am discovering things all the time, always learning new things. There's a was was talking with my engineer this week around just Agentic building 0:57 things. And I said we're I'm just doing regular compression with my with my team. Decompress. What did you learn this week? What's happening? What what things are you finding to be more advantage 1:07 than not? . And one of the two observations mainly came out. One is our team is continually spending more and more time on the planning phases and building a proper phase one, phase two, 1:20 phase three of what I want to build. The the idea, website, feature, whatever that thing may be, typically could be built out in phases. doing a lot more work on the ideation phase is really 1:32 helping improve our productivity with working with agents. Totally agree, absolutely. And we've learned this from you, Matthias. we've been taking your guidance and and trying to directly do that as . that's been one 1:43 observation. The other main observation is markdown is way better than any other documentation that we have. from our standpoint, we really making 1:53 markdown of everything. . and Microsoft has a new Python library called market down. And you Yeah. can take any Excel document or Word document and and mark turn it into markdown. . 2:07 The token usage you get by not sending a Word doc versus a markdown file to an agent is phenomenal. It's a it's a pretty big jump on better meaning 2:19 if you send it images and and documents, it's a have much heavier token usage because it has to parse through the file and there's a lot of extra junk and code in there that doesn't need to have. 2:29 going down to just a markdown piece has been a really big advantage for us as . recently I built a skill around using this new library called market down. 2:40 Yeah. And it runs the local Python script on my computer in my repo and I just say market down and then I give it the file path of the the document that I 2:51 want to turn into markdown. And it just does it. It just figures it out. I built a skill around that Python script that runs and installs and then does all the things it needs to do. I don't 3:01 know if you're using that, Matthias, on your side, but I'm finding good value in using markdown. , I had a use case recently where I had to go the other way. I, , Mark it 3:13 up. [laughter] Indeed. yes, totally agree, ? Everyone out there has to become very familiar and very comfortable with markdown 3:24 nowadays. if you haven't been already, because that's just the way you communicate with agents. 100% agree. 3:34 and I did that recently where I had to produce a pretty hefty document, , that went through 3:44 various review stages. obviously I did all of that in markdown cuz it's just much easier to version control and to send back and forth and and interact with your agents 3:55 on. But ultimately, I wanted to send it as a as a business document in in in Word. And I then used a skill to 4:05 take my markdown document and convert it in into docx. And that turned out very very nicely. Oh cool. . the the third attempt, . 4:16 [laughter] Yeah. Thought was pretty good in in in in keeping going on it and ultimately it figured out and I had a was very happy, 4:27 particularly when I had to do another iteration on that Word document later on. I went back to my markdown sources, , did the iterations there and then had 4:37 had it regenerate. there we go. it's one of those things that we wouldn't have thought, , a year ago we we'd be doing on a regular basis, 4:48 but . there was really good things there around again, this is techniques, things that we're learning away. I've also internalized a little bit, 4:58 Matthias, around what you've been doing around research. I've been doing I've been building my own skills around what is a research what does research mean? how do I use research? How do I connect research to GitHub agents as 5:10 as my open claw agents as as my CLI. And if I find wherever I'm working, I I do a little thing where hey, I want to research this URL. I'll just give it the URL and say, go do 5:21 this. And I have some instructions in my research skill that says, . , go get the website page. Go figure out what's going on there. What are the key insights that I care about? Oh, and by the way, reference that 5:31 against another document I have, my business. what do we do? What software do we build? What how do we add value to our customers? I'm describing that to the agent and then the agent can 5:41 take this article and say, here's where this article applies for your business. . How is it working for you? What things should I incorporate this into my process? Do I 5:52 just keep it on the side, maybe reference it later? Is this something I should give a a documentation to one of my clients because they're really interested in this topic and it would be useful to them. it's it's 6:03 being another second brain, I guess, a little bit to me. Where I'm finding much value in a repo with knowledge and document distillation from random web links or things I find 6:15 on Twitter or videos and stuff that. I'm also finding that to be extremely valuable and I'm pushing into that more cuz then it doesn't forget the information. Yeah, I that's something I have very strong feelings about 6:27 and I'm sure in a future episode what what we can do some demos in that space. Love it. Cuz , I've been building a very extensive 6:37 GitHub markdown knowledge bases in Agentic ways over months and with hundreds and hundreds of iterations and 6:47 commits to it. And it does it all. You always go back and And then that's that's something which first of all you can build Agentycally, you can update Agentycally, but most importantly, and 6:57 that's the ultimate purpose, you can use as input to to educate your agents when when you give them a particular task, ? 7:07 that's ultimately what comes down to. Amazing. . Obviously we want to do demo today, but just in in terms of a couple of news items, I think we have 7:17 to we we did talk about Opus 47 a lot last week. We did. Open AI was busy 7:28 coming in and released Codex 55 this week just one or two days ago. it's very very new, very fresh. But I've already been 7:40 running some coding sessions on 55 far. And There was one 7:50 very unexpected disappointment. I I I gave it a very very substantial spec, multiple files 8:00 of markdown, , something big, but then to be tackled in in in stages. and that was a very fundamental architectural design 8:13 choice which I had made and which was also pretty clear in in the specs. And Codex totally ignored that. And it's currently it's currently doing the the 8:23 first I did the second iteration where I've been very clear . Follow my Follow my pattern. [laughter] yeah. Yeah. . not not not 8:37 the not the greatest start, but hopefully it'll be a one-off. in any case, it it definitely did produce stuff that compiled and and was very 8:51 close to almost completing the entire thing which was meant to be multiple sprints, . that was pretty crazy. it before that, you and I, 9:03 it's availability from my point of view is very limited still, ? you you don't see you don't get 55 in in Co-pilot. I haven't seen it in Azure 9:15 yet. I've tried various times. Yes. but it is available at least to me at this point in the Codex CLI. that's 9:26 how I'm currently using 55. , good to know. Typically some of these bigger, more premier models come out and they are very limited to get access. It it kind of rolls out a little bit. The data 9:36 centers haven't quite been able to move fast as fast as the models can move at this point. we'll see where that goes where where they take us. maybe a quick follow-up from last week as . , we we talked about 9:48 47 being quite expensive and really token hungry in terms of consumption. What I've done, I've rolled back to 4.6 for all my work. Really? 10:00 yeah. Interesting. and I'm very, very happy with that. and I think I alluded to that last time we spoke. they don't make it super easy to do that in the in 10:12 the Claude interface because 4.6 doesn't show up anymore, but you can select it either in your settings.json or you select it using dash dash model when you 10:24 invoke Claude. it is available, they're just hiding it from you. For me, that was the choice, you know? I'm getting super high quality at 10:35 a reasonable cost. Fresh. Yep. This is I think going to continue We we talked about this a little bit last time. This is going to continually be the balance of how much we want to spend on you may 10:45 [clears throat] every often occasionally use 4.7 for something very important or very difficult, ? Let it really reason through things. But I think to your point, the majority of use cases are 4.6 is good enough. It's at 10:56 the price point. It gets things done when you need to get done. You can move forward with that. with that though, there's also there's a lot of reports news-wise around tanking 4.6 performance that way 11:07 it 4.7 looks that much better in benchmarking in these things. And someone said, "I had some data to prove it." Micro Anthropic released engineer at Anthropic an update on the 11:17 Claude code quality reports. they spoke to this initially and they found that there were some bugs, I believe, in the harness. I think that's what they were describing here. 11:27 that they were doing a postmortem around that the reasoning effort was changing in different spaces and it was looking it was reasoning less when it should have been reasoning more. and then what they're 11:38 doing to change the reasoning effort inside the model. this is another article that Matthias you gave us here. I'll put that in the chat window as . This is the April 23rd postmortem. 11:48 and they have made changes to their code as of April 20th and you should start seeing better performance moving forward with their changes that they've recently implemented. That just reminded me I'm just giving you 11:59 another link here. it's it's a particular GitHub issue which I think may have been the origin of of that postmortem article 12:10 you were just talking about. it's a it's a GitHub issue that is very technical but somehow went quite viral. 12:21 it was initiated by a very senior engineer at AMD who had done very, very extensive analysis of thousands of 12:34 Claude sessions trying to prove that their reasoning quality had degraded substantially. And then there was some really good and 12:45 and and a very substantial interaction with the Claude team on that issue. anyone who wants to look behind the scenes a little bit, 12:55 that issue is very interesting read. although it probably goes way beyond what most of us would understand or consume on a on a daily basis. Just be aware someone's looking at it, 13:06 it's going to get better. [laughter] That's the gist here. Awesome. All with that being said, any other news item? Can we move over to show your desktop and we'll do demo ? 13:17 Claude life artifact was released, but that's definitely something we can discuss at another point, ? , sounds good. Cuz that's a very interesting one to look at, particularly coming from a 13:27 Power BI background. yeah, let's let's just leave that here as a as a a tease. Full Tuesday. [laughter] 13:37 Yeah, correct. Awesome, I love it. All moving over to screen share here. Here we go, this is your desktop and we're going to go we're going to go further down this Power BI modeling 13:47 MCP. This is your GitHub repo. This is where we're working off of . If you want to go see this repo, I'll make sure I put a link in the description as in case you want to see what Matthias is working on and this will be 13:56 continually updated as we add more information to it. Absolutely. this is all public, ? available to anyone. 14:07 last time we talked a little bit about how to set this all up and how to kick off a cloud agent session which is this one here. 14:21 was quite simple and initially we just did a simple check to see whether we could connect to our remote semantic model in in Fabric. And 14:32 then I gave her a tiny prompt, "Review model and propose suggestions." which Copilot responded to with a very, very substantial, as 14:45 you can see, a list of things that can be improved. that was awesome in terms of cheap to do 14:55 big win. we've got lots to do's . because I didn't want to sitting here 15:07 looking at the agent doing stuff, I prepared something for today. I created a new session here from scratch as you can 15:18 see an hour ago. and I asked her to do everything we did last time and more. I asked her to connect. I asked her to do the review and to propose 15:29 suggestions. And then I asked her to export those suggestions into a markdown file, ? we talked markdown earlier. I also asked her to 15:39 export the full model as tim doll into my repository. and it's done that in 5 and 1/2 15:51 minutes. Not too bad, ? And took me one premium request. very, very cheap. and we've seen that last time. 16:03 then we have 10 items for key improvements. and it's also told me how it's 16:13 exported 19 tim doll files into this particular folder. what I wanted to point out here, which is why I haven't taken this any further, 16:23 first of all down here, you can expand and collapse the list of all files added, deleted, or modified. 16:33 a git diff. you can also click here to see that. But most importantly, I can create a pull request 16:43 directly out of this agent session. this is how everything is quite naturally integrated in in GitHub land. 16:54 and Can I ask a question here on this one cuz this is something I've been also working through with ? Yeah. You're you're creating a pull request, but this is in draft mode ? And 17:06 one thing I just want to highlight in the upper -hand corner, , couple places, ? You can see on the on the upper hand corner it says not ready. Yeah. One. And then on the left-hand side you see it's has 17:16 another little icon called draft. Yeah. And maybe Matthias, could you speak a bit about , you're you're doing from the agent session. We've left the agent session and we're moving over to this draft 17:28 of a pull request, I guess is what you this would be called. Can you maybe explain a bit more what that means that way people can understand, , the steps here from where we were to draft? 17:39 . . first of all, the agent has automatically created a new branch, which in this case is called 17:49 Copilot forward slash connect sales model and list tables. not not an ideal name because the the the the the the key point here was 17:59 the export and the suggestions, but there we go. I can live with that. and that draft is what we 18:12 have the pull request created for. That branch is what the pull request is created for. the draft stage is a 18:25 Co- It's it's a GitHub a a feature, let's say, where you can 18:35 it comes down to to your to your workflow and how you engage with within your engineering team, ? you can say, "I'm creating this pull request here, 18:45 but I'm not quite ready yet for it to be reviewed." Yes. ? I don't want to share it with my reviewer or my line manager or my peer 18:55 in the team. This is for for , this is a private pull request because I may make further code changes. I may review 19:05 this and when I as as the originator think I'm ready for this pull request to go further, I 19:16 then change the status from draft to ready, ? and it's quite easy. If you scroll down here, it says this 19:26 pull request is work in progress. I've got a ready for review button here. And once I click that, I then have the green button which allows me to merge. And it it also says ready to 19:37 merge up here, ? there we go. but but sorry, when when when whenever Copilot creates a pull request for you, it will always be created in draft to 19:48 begin. Yes. Another thing I want to point out if you go back down to the bottom section here, and this is something I was observing as this does not have any automatic GitHub actions attached to it. , or if 19:59 for example, if you're building software, there may be some a build step that you want to apply that would run some actions to do something. While in draft mode, the actions aren't 20:09 automatically kicked off. , another advantage of draft that I saw as I was working with it is you have to as a user, the human in the loop, ? Agent can go build some code. Agent can 20:20 say it, we're ready to go. Agent can say, ", the draft is prepared." And there's a separate box here, a little approval box that says, "Hey, do you want me to run the actions to build this code on this 20:31 branch for whatever app or other thing you're building?" , it also it also pauses some because actions cost real money cuz you're spinning up a runner to do real things. 20:41 And , one of the things I wanted to note here is it protects you against just running a lot of Copilot generated actions because you're just making lots of changes quickly, ? That's it it protects 20:52 you there. , that's a really good point. , and let me just continue from that. , first of all, we we've got a button here which allows us to go back to the original agent 21:04 session. Love it. where we started from, ? And , you just talked about spending money and cost. Yes. And we spent a fair amount of 21:14 time talking and explaining premium requests last time. Sure. , the thing people need to understand when it comes to running Copilot in the cloud, there are 21:25 two costs components. , there is premium request which is what we discussed at length already, but also 21:36 those cloud agents run on GitHub actions infrastructure, which means those 5 and 1/2 minutes here, they would have consumed 5 and 1/2 minutes of my 21:46 budget for GitHub actions. And that's really [clears throat] really important, ? , you have two 21:59 bill counters. Sorry, I forgot the word . you , you've got you've got you've got two different purses. Yes. Money comes out of 22:09 for those cloud agents, and you have to be aware that you need sufficient budget for both of them. Both in terms of Copilot 22:19 premium requests, but also with respect to your GitHub actions minutes allowance. Generally, that's not an issue because 22:29 generally, you have thousands of those, but nonetheless, it can add up. Sure. Exactly. And then the other thing the other thing to say is 22:39 GitHub Copilot is highly sandboxed, and they take security very seriously, and always lean towards 22:50 being restrictive more than not. what you mentioned earlier is this bit here. 23:00 Yes. By default, if if you have workflows set up in your repo, which I don't have in this one yet, ? I'm sure we'll 23:11 get there. But if you have workflows set up that are triggered automatically on pull requests, . ? By default, Copilot authored pull requests do not 23:24 trigger those workflows. And that's this particular setting here, ? By default, this is on, and it says, "You need to expressly allow to run your 23:34 workflows." If you turn this off, then it will just automatically run all of them, but by default, it requires your interaction there. And 23:45 your choice, ? , this can be turned on or off at the repo level. You may want to turn it off, but just wanted to make you aware of what the default is. 23:57 Obviously, this is all about protecting you, , when an agent does stuff on your behalf, you don't really know what may happen, ? All sorts of stuff could suddenly 24:10 end up in your repo. , this is to ensure that you review before then the code in your repo suddenly runs, which 24:20 is what happens when when when you've got a triggered workflow. , . let's go back here. 24:31 We have this PR. first of all, we can go into files changed to see everything that's been 24:41 done. And this one is the one we're particularly interested in . This is our markdown document which 24:52 explains the 10 items which Copilot in conjunction with the MCP server not 10, 14, 15, 16. Oh my gosh. 25:05 18. lot. yes, it's suddenly 18. 18 items which Copilot 25:15 identified for improvement. The reason why I just got confused about the 18 is, and this is I think a really interesting point to 25:26 make. If we go back to what happened last week, . and if we look at the agent session we ran last week, 25:36 look at that, it found 10 . items to fix unless the counting is somehow different here, ? And then 25:46 also earlier on on on a different private clone of that same repo, I ran the same prompt, and I got 14 items 25:59 identified. And the one we're currently using has 18. , this is this is non-deterministic. Yeah. what we say it's non-deterministic, ? Exact same 26:10 prompt, run three separate times. with the exact same model, with the exact same agent can give you completely different outputs. 26:20 That there's nothing wrong with that. If you think about it, , agents AI agents are meant to emulate what people would do. And if you 26:31 ask the same person Yeah. , to to do the same job multiple times in in a row, they may also not do exactly 26:42 the same thing, ? , exactly if you ask me to review this model, the same model three weeks in a row, I mean, I might find something I missed the first time. I might look at something and , 26:53 again, this is also where we would make we could do analysis off the offline here of , ", we ran it three times. We have three different piles of things we want to get done." There's probably some very evident and very 27:03 common high priority items that are going to be consistent across all three observations. These little or lower end fixes or identifiers might be 27:15 part of what you would normally have when you work with an employee. They would have a couple extra little things. It's a little bit preference. . All , cool. , let's move on a little bit. 27:25 I thought we would move down to the desktop and go into VS Code today if that's . , I've got this 27:35 exact same repo open here. As you can see, there are only two files because that's still the main branch. , what I'm going to do is I'm going to Can you zoom in just a touch on this one if you don't mind? Maybe bump it up 27:46 control plus maybe one or two times. Control plus, yeah. Or something that. Yeah. 28:01 There we go. Better? Yeah, that's great. Just want to make sure you get everything big for our I know it's harder to work in as a user of it, but [laughter] trying to make it easier for people to see on screen share. I'm going to continue locally, 28:10 and as you can see, I've got this branch which was created an hour ago. , I'm going to switch to the branch. but as Matias is doing that, I'm 28:21 going to say Matias here is using a a feature of VS Code called GitLens. GitLens allows you to see what changes are made in repos by who, and it's amazing. It's a really good 28:32 program, and I'll put the link here in the description as if you want to use GitLens. It's a third-party extension. It's not a Yes. Sorry, it's not built in. My fault my fault. But it it is 28:44 it is a must a a must for anyone who works with Git, which pretty much everyone is nowadays. This is much better than 28:54 the native Git features in VS Code. That's great. There we go. And this extension is comes with a freemium model, you can use 29:05 a large chunk of it for free. not even trying to sell you anything here. cool. , we've got a few more 29:17 items as you can see, including our semantic model. and what I wanted to do 29:29 is first of all, here we've got our issues. , I wanted to use an agent to create GitHub issues 29:41 individual GitHub issues for the stuff that's been identified here, ? , let's just see. 29:53 we're we've got let's let's use a filter here. , some of , why don't we say create an issue for anything 30:04 where severity is medium or higher, ? Sure. All , cool. , I've got my chat 30:14 window here. I really the autopilot feature. it's currently in preview, and it is a a a nice mix between 30:26 prompt me for everything versus going YOLO. , autopilot tries to do the thing in terms of tool approvals. I really love it. 30:39 cuz you are safer than going in YOLO, and you don't get prompted all the time to give permissions. . looking at the 30:55 sales model improvements document please identify all 31:09 items for improvement with a severity of medium or higher. 31:21 and create a new GitHub issue for each 31:32 with a detailed explanation of what 31:43 to fix. There we go. . I have a few typos in here. Generally, the models are pretty good at figuring those out. 31:55 and let's see what it does. it's it needs to read this. It needs to 32:05 perform some filtering, semantic filtering, ? it needs to understand what medium or higher means. . it needs to be able to connect to GitHub using my credentials. It it 32:17 needs to create the description that each issue is self-contained 32:27 and can later be picked can later be picked up by a coding agent for instance. , . 32:37 one of the things I really is how you've got this to-do item that is updated in real time. 32:48 once the agent has figured out that it has multiple individual things to do, it creates them as a to-do list, and you can see in real time 32:58 how it's going through that. It's really really nice, particularly when you have longer-running sessions. it's a very cool UI feature. , Matthias, what you did 33:08 here is you asked it to create something here. Would you do some iteration and and some something I've done a couple times is when I needed to go do some other action somewhere else, 33:18 ? , either take the information that's here in this markdown file and enhance it. , , we have a section called severity, and we have the suggestion that's applied 33:28 here for this particular section. sometimes I'll do a bit more maturing of this one document. , knowing that I'm going to create issues, ? I maybe would go back to the 33:39 agent and say for all the items in this markdown file that have medium or high severity . write up an issue description as if you would use it for GitHub. And have it write 33:50 the information into the markdown file or have it, , enhance this file a bit more . that way when you go to the next step, , great, you can read through it all holistically and 34:00 say, , great, make the issues. And again, just pick up the data from here and go to the issues. Do you find that's a pattern that you use? Do you do you go through a couple iterations on documents this? 34:10 the the what you're proposing has the big advantage that it would be kept in source control, ? Because it's obviously a file that's in your repo. 34:21 you have nice version history for that. , from that point of view, I would definitely it. I just chose something that is relatively quickly , ? , obviously, Bigger projects require a 34:32 bit more rigor around these things. Yes. and we can also see this is using the GitHub MCP server, which 34:43 down here. That's our tool configuration. When we click on that, we see which tools and MCP servers are available, and most importantly, which 34:53 ones are enabled. It's very good practice to disable as much as possible because those tools 35:03 clutter your context, and they can take up a lot of let's call it mental 35:14 space. , your context window. you want to be quite selective. and as you can see, I've got GitHub enabled cuz clearly we need that for 35:25 the task. I've got all the built-in tools, and then I've got some of the GitHub pull request tools enabled, and everything else is disabled. and that's how you do it. 35:37 and we can see even with that selection, I have 88 selected. , it is very extensive. It is is, ? , definitely 35:47 something to be aware of. all , there we go. Look at this. we have seven GitHub issues . And as you can see, not only do I have a very nice 36:00 take summary table, I even have direct links. Oh, I love that. And , look at that. Awesome. this is issue number three. 36:10 obviously, it appears as if I did that because the my VS Code instance doesn't have its own identity. It's using my 36:20 GitHub login. , it then appears as if I did all that stuff, ? makes you extra productive to your organization. [laughter] But it's also something to think about, 36:30 ? , you you you you may want to give your agents explicit identities and and then that way when you when you trace items, , through 36:41 systems your GitHub org, for instance, you can you can tell the difference. , that's why I'm pointing it out. It's it's it's not always also, if you wanted to do some analytics with respect to, you 36:52 know, how am I using agents and how what percentage of my work is being done by agents ? And if they always work under your identity, you can't really do that easily, ? 37:02 That's what sets it apart. Definitely something to think about. all , we're not going to be able to go substantially further today, but 37:13 as you can see, this is substantially more than Wow. Yeah. what the original document had. , , what what you 37:24 alluded to earlier in terms of enhancing the the relatively small prompts we were given from the original Copilot session, this happened 37:34 in any case, but it it happened on the fly, as you will, ? I didn't do it as an explicit step in between. and , 37:44 I'm not going to do that . We we can probably save that for next time, but I could go in and click here and and automatically assign that to a Copilot agent 37:56 to pick up as yet another cloud session. But before we do that, we have to do a bit more configuration here, making sure that 38:07 the cloud agent then doesn't suddenly directly change my the model in my workspace, for instance, ? I I would want to make 38:17 sure that any work that's happening is happening on the files in my repo, and not in my workspace model directly. And , that's why I'm not clicking that button 38:27 because we haven't told it yet that that's what we want. , cool. There there's much more we we can do here, and in particular, 38:38 there are various skill plugins from Microsoft, from Rui Romano in in 38:49 particular, that will help us with further work when we're in this VS Code environment, but I think we're going to have to 39:00 leave that for another demo day. I agree. one thing I'll just want to call out here, Matthias, are you using the website Awesome Copilot? Have you seen Have you Yes, yeah, absolutely. 39:11 Yeah. , the reason I bring this up is because you're saying about skills and custom agents or things that are applied that you would apply to the repo here to have you enhance the agent experience. There's a whole bunch of Power BI 39:22 advanced modeling developers and DAX developers, and there's all these people are starting to build these common libraries of skills and custom agents that you can go They're just markdown files. You add them into your 39:32 project, and we'll we'll probably talk about that later, but I do want to point out the website Awesome Copilot. I'll put it here in the description as , just in in you want to check out the other experiences how you can get 39:43 additional agents into your workflows. Let let me do let me show one final thing which is directly related. if I go into my VS Code settings and I filter on plugin 39:55 I'm going to see first of all I've got chat plugins enabled. Secondly, I can set custom plugin locations. for instance, I can 40:05 create a central folder on on on my local machine where plugins are provided but most importantly and this 40:15 is the thing I wanted to show. We have We we can register plugin marketplaces and I'm showing that because what you 40:26 just mentioned is predefined as a default marketplace which means if you go into the plugin extension window that 40:40 is part of VS Code. Anything that's published on Awesome Co-pilot or Co-pilot plugins will automatically show up. Love this. The third one down here 40:50 Rui Romano / Power BI agentic plugins. I added that one myself. Clearly that is not something which ships automatically with VS Code but that's the one where 41:00 next time we can explore what Rui has made available to us in terms of Power BI skills and plugins and what we can do with those 41:10 in conjunction with that particular project. Love this. Great great another good tie in to the link I just randomly threw in here as . these are what we're going to need continue to be 41:20 exploring. Matthias, thanks for the demo today. This has been really useful and very impactful to unpack what's going on here, how these demos are working and and the whole goal of this series is just to continue furthering 41:31 knowledge around agentic spaces. How do we use them? How do we continue to build on them? And what are we learning? Everything's changing quickly for us. Every week we're coming out with new insights or new information about how we 41:42 can work better with these agentic pieces. Matthias, thanks for the demo today. We really appreciate it. You can join us again next week on Tuesday at 10:30 roughly 10:30 we'll start around 41:54 that time 10:30 Central Standard Time as . we're going to keep doing these things and we're going to be Tuesdays will be more focused around discussion, news, what's happening in the market and then Fridays we're going to focus more of our attention around 42:05 demos and live stream screen shares just to really show you the clicks and how things work and and stitch together. Yeah, looking forward to it. Should be a lot of fun. Thanks everyone for joining and 42:15 watching. Thank you all. We'll talk to you soon. bye. Agentic Thinker [music]