chapters transcript notes
click any line to jump to that moment in the video
0:13 [snorts] Welcome back to Agentic Thinking with Matthias and Mike. Hello everyone and welcome back to the show Matthias. Happy to have you back for another Friday session with us today. We're excited to jump in. 0:24 Let's talk about our main topic first and then we'll get into our details. Matthias, what's on the main topic docket today? What are we learning about today? we're continuing with with 0:34 our exploration of using Agentic workflows with a with a Power BI project. last week, if you if you followed, we got some mixed results, 0:47 let's say. , we we we weren't quite happy with what the model produced there. this week it's all about figuring out how can we prepare the 0:57 project in in in different ways that the exact same prompt will give us substantially better outcomes moving forward. lots of learnings here. I it's going to be 1:10 unplugged, ? I don't really know what's going to happen. Real demos. As Mark Russ would say, we're doing real demos. But I I I have thought about this a fair bit, 1:21 I've got a plan in my in my mind about it. Yeah. What's what's up with you? any any news items to share? Yeah, I 1:33 I do a lot of thinking. the algorithms all know what I . And what I is everything about business, marketing, and AI. 1:43 everything on my feeds, all of the platforms, have learned this is what I at this point. And it's all of them, doesn't matter. YouTube, X, 1:53 LinkedIn, all of them are sending me feeds about this information. , I I'm I'm really engaging with this content. one I'd to share with the community here is just this 2:04 It's a concept of in the age of AI, which is I think what we're entering in . This is This is a new era to build and design. There's this idea or concept around 2:14 founder mode. And [snorts] , in this short I'll share here in the chat window in case you want to go look at it yourself on your own time. Let me copy the link here. I found this to be really 2:25 engaging from a concept standpoint. and it was thinking through if you're a CEO and you are using agents and you have a 2:36 large company, it's going to be more difficult for you to adopt agents and move forward because there's , we got to be safe, we got to move strategically, we don't want to upset our existing customers. There's this 2:46 a bit of resistance to doing things new and creatively. The age of AI is incentivizing, 2:56 I think it's the word I want to use, incentivizing the founder model. Move fast, build with agents, create quickly, go and just build faster, ? And , this 3:08 gentleman talking on this podcast, I think he works for Airbnb. He's reimagining his whole the whole company or or portions of it. I think that's where he's working from. I Don't quote me on that. But, the reason I this 3:18 this short a lot is because this is how I'm thinking. I'm rethinking my entire business. , anywhere I have friction about something I am not wanting to do, answer emails, 3:28 write statement of works, , laborious time logging of tasks and things, I'm looking at this and going, in the same way I had when I saw Power Query for the first time, 3:38 I saw Power Query and I thought, oh, this is amazing. I don't want to go import from Excel anymore. I want to use Power Query for everything. , even if it's , I want to make this tiny little table inside Excel. I don't even 3:49 want to do that. I I want to go to Power Query and then use this really nice experience and automate everything. And I'm having the same mental thinking processing that's 3:59 happening on top of agents and everything I do. All my entire business. And I really this AI founder mode. I'm fully embracing it. I'm on 4:09 board with it. and I that's where this short comes from. Matthias, what's your reaction to this this short? totally agree, , definitely the the way I think about stuff as . 4:19 What's interesting to me, I see insane analogies here between the business world and engineering world. 4:29 with respect to challenges we have with , those those new AI capabilities. when you've got a brownfield engineering 4:40 project and you are trying to AI enable it, you're going to find that it's much harder than 4:52 doing the same thing with a greenfield project where you start off with the project using AI practices 5:03 and and gigantic rock rows from the beginning, ? And it's exactly the same thing I believe when it comes to companies, ? anyone who 5:14 starts a new business a new company in an AI native way Huge advantage. a huge advantage over trying to convert 5:26 legacy legacy, exactly. isn't it funny? Yeah. This is exciting. . And and I think Matthias, both you and I resonate with this really because we are both 5:37 not only are we founders cuz we have our own companies, but we're AI founders as . we're heavily invested in AI is changing how we do things. And as we talked offline, 5:47 a majority of what we spend our time thinking about and building through and creating with is heavily how can I get the AIs to do this? How can I it's not just can I solve the problem? 5:57 It's how can I solve this problem with an automated AI thinking system? Where does the AI need to reason about things? Where doesn't it need to be? Where does the AI create things? Where does it 6:07 Where does it need to be in investing in thought? anyways, really cool topic, great short. I think I would I just would share that as . I do want to be mindful of our time. We don't have a ton of time today. 6:18 Matthias, maybe we just jump into demos. Sure, absolutely. Cuz that would be much more to say about this, but I apologize. You had one more thing I thought was really really important. I 6:29 It slipped my mind. You had one more announcement that I need to go pull up a link for. Yes, 2 days old. 6:39 I'm a huge cloud code user, ? And Anthropic announced something which is quite unusual these days. They They've been giving their users higher usage limits. , generally 6:50 and we've talked about what what GitHub Copilot has been doing lately. Things are moving in the opposite direction. 2 days ago, big announcement on the Anthropic blog. Higher usage limits due 7:02 to their presumably significant compute deal with SpaceX. they they removed 7:12 peak hour limitations. They doubled budgets within 5-hour usage windows. And I can definitely say I'm , this 7:24 is not just a blog post with an announcement. This is real, ? I've I've really I've used it. And in fact, 7:35 I'm really struggling using up my max code subscription. It's really It's a real challenge 7:46 spending all those tokens. , there we go. That's exciting news, and it just shows how competitive the market is. Yes, and I would agree with you there. And I This is the thing 7:57 we This is This is what we've been saying since the beginning of the Agentic Thinking podcast is everything is about tokens. It's going to be It's going to be a 8:07 token war at the end of the day for all of this stuff. , very excited as . Awesome. Very good. All , that being said to our desktop. Absolutely. Let's do that. Excellent. All , let's jump over to 8:17 the desktop here. Matthias, pick us up where we were before. We have RBI model, and what we're going to do next? Absolutely. , cool. , 8:27 that's where we landed 7 days ago. I still have my chat session open here, and this particular prompt 8:38 not very written was saying from the model improvement recommendations 8:48 the agent had created for us previously tackle two of them just placeholders and measure descriptions. Create a new local branch, commit changes, 8:58 and make sure that the GitHub issue which in this case is GitHub number three or number four is referenced. 9:08 it then kept working. In fact, it took quite a while, and ultimately we got a nice report as you do. but when you and I reviewed what was 9:20 done, , there were quite a few things that , we weren't happy junior. It's very junior. Indeed. , first of all, just a little recap. Something which is 9:32 really amazing because we've already committed all those changes, obviously, ? , and , looking through my Git commit history at this point, I wouldn't necessarily 9:43 know what exactly the agent has produced, but , in the VS Code chat window, I have , this alternative diff view, which for 9:53 people need to understand this is different from our Git diff, ? This is the the diff that was produced as part of the session output, and , it shows us how , 72 lines were 10:05 added across those five files, and as you can see, , it does look exactly the same as a Git diff in VS Code, you know, where we have , green and red 10:15 color coding for additions and removals. and , , Yeah, I want to I I hang on here just a moment. Also, notice there's some different icons in the folders 10:27 window as . , I'm going to highlight this as . Yeah. , yes. when you change a file versus when there's a new file versus when there's a agent-edited file, . 10:38 there's three separate icons . I believe the square with a circle in it is indicating to you that that file has been changed by the agent, 10:49 but it is not it's not Yep. It's pending it Yeah. Perfect. Love that. , it's it's highly useful, but at the same time, it can be really confusing as 11:00 because if you switch from your explorer view to the source control view, you're going to see very similar things, 11:10 about modified files and their respective changes, and , it over there, the change would say, was it an addition? Was it a deletion? Was it a modification, ? 11:21 Sure. Yes. , , and then the the similarities go even further, when when you look at the diff view you're getting here. , for people 11:32 who are new in that space and , who are not heavily used to using source control, 11:42 it can be quite confusing. And just understand one scope is what Git sees in terms of what's changed in your folder, the other scope is what your 11:53 chat considers a change. Yes. It's a good call out. that one also allows you to explicitly keep or undo. 12:06 Let me call this one out specifically, ? Because obviously as I said before, all those changes that were done as part of the session are 12:16 already committed, ? Which is why looking in the source control bit over here, we don't have any pending changes. They're already in my Git commit history. In fact, they're all in here. 12:29 this keep and undo, which we have here for each individual granular change and which we also have here for all global 12:39 changes. And in addition to that, we also have here at the file level. , a file and three different scopes at which I can apply keep or undo. That 12:49 is related merely to the the chat session, ? And , I can for instance, I can say keep this one. And if I were hypothetically, , let 13:01 me do it. If I click undo here, . it it undoes it, which then means obviously, this the undo relates to what the chat did, but 13:12 it it does mean that with respect to my source control, I have a change, which obviously is the removal of that particular line that had been added 13:22 previously, ? , folks just need to be very, very mindful, that we have two highly overlapping Yes. 13:32 diff and change control systems here, and they both make sense if you really understand them, but they , you 13:42 can get highly confused as . Sure. obviously I don't want to undo this, which means I use the Git undo 13:53 [laughter] to undo my undo my chat undo, ? [laughter] yeah, there we go. . which also means down here 14:04 you can see I no longer have four files showing up as changed rather than five as before. . But let's get to the to to the 14:14 gist of what we wanted to achieve here. , Sure. our main criticism , there were there were two main criticisms, I would say, ? 14:24 One is, , and and that's the key one, we noticed that whilst all those measures have had descriptions added automatically by the 14:36 agent, those descriptions were not very useful from a semantic point of view. Those descriptions were mere rephrasing of 14:49 what the underlying DAX function does. Those descriptions are not achieving a business-focused description of what the measure, you 15:00 know, from a semantic model point of view does. And in that respect, , I I said a week ago, this is the stuff I 15:10 would expect from a very junior modeler, who was tasked to provide descriptions for measures because intuitively they may do exactly 15:20 that, ? Read the DAX formula and and just describe what the formula does as opposed to understanding, you know, what does this measure do in the 15:31 context of a dashboard or report, for instance. , that was a criticism we had. And I wanted to explore different ways of 15:41 how we can mitigate that and how we can get to to a better place. 15:51 And , my idea was because all those changes are sitting in this one Git commit here, 16:01 which by the way was our second criticism, but that's a lot more minor than the other one. All those changes sit in this particular 16:11 Git commit, and in fact, they sit in a dedicated branch as , ? , remember down here, we created this branch. in fact, that was part of the 16:23 prompt for the session. , I'm going to I'm going to create 16:34 a new branch with which doesn't have any of the those changes done, and I'm going going to try again, and I want to try a few 16:44 different methods here. . , let me create a new branch here. let's call this 16:54 do do do do I don't know, demo model fix 17:05 attempt two, maybe. Yeah, that. . Cool. , create branch. . , and whilst we're here, 17:17 I I I don't want to I don't want to lose anything. , I don't want to switch to this branch branch immediately. I don't want to lose anything in this particular window. I don't want to lose anything 17:29 for comparison purposes. what I want to do instead I've got this particular branch here. I want to 17:41 create a work tree which again is a relatively new Git concept and it's become quite popular when 17:51 Cloud Code introduced native support for work trees. , work tree in short allows you to check out multiple branches of a repository in 18:01 parallel without explicitly having to create separate clones of of that repo 18:11 yourself. it means that the the the main cloned folder has a reference of all those additional work trees you've created and you can 18:22 easily switch back and forth between them. , I've just said on this particular branch, I just said open in work tree and as you can see, it's 18:32 doing two things. One, it's creating a whole new VS Code instance for me which means , the other window I can keep open in the 18:42 background. I can have that chat still there I can always easily go back. but then also, we can see that down here very nicely. This is This is my new 18:54 folder. This is the old folder, ? Love it. also look at the actual file system folder it's created for 19:04 that. And we can see down here, ? , by convention, , the original clone I created was called PBI 19:15 modeling MCP co-pilot cuz that's the name of the repository. By convention, it then creates a new subfolder where it appends dot work trees, and then it 19:25 appends the folder the branch name I've selected on top of that. , there we go. , that's how work trees work. Definitely a 19:35 concept you should be familiar with cuz it's I'll pull a link from something about work trees and just get a 101 on work trees cuz I think they are very powerful, and especially when we're building with agents, this is a a skill 19:46 that's worth learning. Absolutely. . , which means with this one because again, I I created that branch one commit ahead, we're 19:57 we're back to where we were before, ? , we've got the semantic model checked out here. We've got this original research item that creates an 20:09 inventory of everything that could be modified. and , I want to Let me just do a little bit of an 20:21 let's call them prompt improvement methods. , I want to show three, maybe 20:31 four different ways how we can change that. , the first one would be to use a richer prompt. That's the cheapest and most intuitive one. 20:42 not cheap. It's the most expensive one, but it's the the easiest in terms of getting it done quickly, ? the second one would be to use 20:57 agent instructions. at the Can you zoom in a little bit on that Yes. control plus on that document 21:07 there. Yeah, perfect. Yeah, that's instructions. Thank you. that would be for the repository. the third one would be to use skills, and then the fifth one would be 21:17 to use a custom agent. Those are all valid ways of achieving what we need to achieve here, and it's not the case at all that 21:27 one is better than another. anyone who actively works with agentic systems needs to be aware of all of them. . 21:37 let's just go back here, and the goal is to be able to use the exact same prompt, but 21:48 to achieve better outcomes. , I'm going to paste the prompt in here, and then I'm going to add something to 21:58 the prompt, ? , all the things that we want to be done separately all the things that we want to be 22:09 done differently, I'm just going to add to the prompt here. , what should I say? when composing 22:21 measure descriptions, do not and capitalization generally 22:32 is a very good way of highlighting or emphasizing instructions you're giving to an agent. do not and merely describe 22:43 what the measure formula formula does technically. goodness, I'm really bad typing 22:53 today. instead think about this from a 23:03 business user perspective and describe in instead 23:13 what a [snorts] Excel user for instance 23:25 can expect as output from that measure, ? , that would be one additional sentence or paragraph 23:37 to provide a richer prompt here. The other thing, oh, sorry. I am I sent it off already, but I think that's fine. We we don't 23:49 have to worry about everything else. We just want to see that that it how it behaves and and you 24:01 know, whether it it produces an output that's much closer to what we're looking for. in the meantime, what's the Yeah, sorry. Go 24:12 ahead. Yeah, you see, one of the things that while you're in that text box, this is you , not accidentally, but you hit enter and it sent the message. , it's funny when you're in that little text box window, it feels 24:23 the DAX editor or when you're in making DAX formulas, ? You need to hit, I think it's shift enter or control enter in order to get another a new line to make new lines in 24:34 there. Cuz as soon as you hit enter, it just goes. No, no, no. the real the thing that really happened here is that I constantly switch between different coding agents. 24:45 And it [laughter] happens to be that in cloud, you on a on a Windows machine, you need to press control enter if you want to and a new line. Yeah, that's 24:56 correct. And in most other systems, you need to use shift enter for a new line. And I just accidentally because I've got the muscle memory from from many 25:07 cloud sessions, I just press control enter. Control enter. Which unfortunately doesn't mean anything. It's not in VS Code. that was my mistake. No, it's . It happens a lot to me, unfortunately, but 25:17 there we go. Inconsistent tooling or inconsistent harnesses, I guess, is is what that is is causing the problem with. Yeah, it it gets even worse because 25:27 when you use Claude in the browser, it accepts shift enter rather than control enter. you have to be extremely mindful of it. 25:39 this is something I was hoping for. I remember permissions is always a a weak a key thing for you to configure correctly. 25:51 one you don't want to overdo it and and give too many permissions cuz , you can easily get into trouble with your agent going 26:02 Too aggressive. profile. And on the other hand, if you're giving too few permissions, this is happening, , you're constantly going to be 26:12 Approving. prompted and the agent is not autonomous and you get disrupted, ? finding a balance is key here. It turns out 26:22 that this is not something I'm struggling with. It turns out that the VS Code guys, , who who 26:33 built this harness, have provided us a really good way around this. They've got this auto pilot mode here, which means 26:45 you don't have to go for default approvals, which , which prompt you a lot. You don't have to go for bypass approvals, which never prompt, but allow everything. And Autopilot means every time the 26:58 model would prompt you, it's using another a model to determine how severe that 27:08 particular tool action is, and that model then makes a decision on your behalf, ? , that's called Autopilot. Yeah, reading something is less impactful than writing or, 27:19 creating something new or even, , delete it would be on the highest end of that threshold. deleting things is probably going to , "Oh, this is a severe action. We should really confirm with 27:29 the user a delete is is required here." Exactly. , when I use Claude, I'm generally very good at predefining the permissions in in my Claude settings 27:39 files, and normally it it it's it's very good at then not prompting me when I use Copilot and VS Code Chat, 27:50 that's not straightforward. Unfortunately, the declarative permission system is not as great as the one in Claude, but Autopilot is really good. , there we go. I'm going to switch that. 28:02 it's warning you. But, there we go. but, I still for this particular prompt, I still need to explicitly confirm it. , what's this 28:12 doing? It's looking at Oh, there we go, ? , it's it's using the GitHub CLI. , , that's GitHub CLI that allows you to interact with the GitHub API in a very high-level way, 28:24 and obviously it wants to list the issues I've got on this particular repo because I referenced the issue as part of the prompt, ? , of course I want to allow that. 28:34 All . while this is doing that, I'm going to go back to my original window because obviously I promised you four different ways of tackling the same thing. And , 28:44 it can go on, , with my modified prompt whilst we look at alternative ways of achieving that. , I'm doing the same thing here. I'm going 28:55 back. I'm creating a new branch. I'm going to call it demo. 29:10 to agent instructions. Here we go. Create branch and then I can go here. And I can do the same thing. Open in work tree. Got got yet another window. 29:25 This time I'm going , first of all, I'm making it bigger you can see. Yeah, thank you. Appreciate it. This time I'm going to switch into autopilot immediately. 29:36 I love it. Unfortunately, this is not something which is persisted across VS code session you always have to do that explicitly unless I haven't discovered it yet. if if someone knows something I 29:46 don't then let us know. in the chat. If you if what this is and how to keep it on all the time. This is collective learning here, ? Yes, correct. , . , I've switched that. By the way, , I'm 29:57 using Sonnet 46 here. , your choice whatever model you want to use but that one is very capable and at this 30:07 point you still get it for 1X premium requests unlike Opus 47 which is 15X as you can see. , that uses 15 times more of your allowance. 30:20 agent instructions. What does that mean? We can at the repository level define 30:36 standardized documents that will be consumed by any agent session automatically. and and unfortunately, there are various 30:46 conventions what that document is called and where you place it depending on which harness you're using. , in Claude code for 30:56 instance, it would be claude.md. thankfully, outside of Anthropic's Claude world, many harnesses have converged on a 31:07 convention that's called agents.md. and , I'm going to create an agents.md file here. 31:17 and Copilot will read that automatically. and in fact, agents.md is a convention that also works 31:27 in a scoped fashion. , you can create agents.md at the root of your repository, which means that one will always be read in any new session. You 31:38 can create nested agents.md files in subfolders, and those ones will only be read and only be applied when the agent does something in that particular subfolder. , very powerful 31:48 concept. That's cool. . And , this one means you're putting let's call it static 31:59 additional context into all sessions that run within that particular repo, ? , 32:09 and , this is something which will automatically and completely hidden from you be hidden to any prompt you're 32:19 giving. , for instance, I can say when ever you 32:29 annotate semantic models with 32:40 descriptions ensure you do not merely describe the underlying 32:52 DAX functions. Goodness. Sorry about my typing. there. Go back to there. Instead, focus 33:02 on explain the business logic and the purpose of the measure. Exactly. , yeah. , this this Tab Tab through that one. Absolutely. this is AI yeah. Helping us. 33:13 Yeah, . , if you're looking at the evolution of AI, this auto complete is the very first, , iteration that we had before chats 33:26 and agents came about, ? But Yeah. But nowadays, they're much faster because they can use very low latency mini and nano models 33:37 for you to provide those sub second auto completions. And they are just extremely good nowadays, ? 2 3 33:47 years ago, hm? , could have been Hm. gone one of two ways. Nowadays, I love auto complete. good at inferring your intent. Hm. 34:00 they suggest you stuff that's way better than what you would have written. it And in this case, it turns out, ? I've given it this and it it completely inferred what I wanted 34:11 and gave me a much better version of what I probably would have typed, you know, with various typers . , there we go. , I'm happy with that. this will help users understand the 34:23 intent behind the calculation, how it can be used. There we go. I want to pause while you're going through that one here as . One thing I want to note here as , a lot of what I'm doing , pattern-wise, is I'm doing things this, where I'm 34:33 building some semblance of, , an agent or custom agent, or I'm building specific things here that are going to help me generate better output from the agents. I'm also 34:44 to this end, this is just a file, just markdown and everything else. I'll take a first swing at this. . And then I'll go back to an agent and say, "Opus or something more reasonable, with more reasoning, and say, 'Look, the 34:55 intent here is to make it not just write the DAX formulas, to give real business user value.' Look at this Look at this prompt. 35:05 What should I What am I missing?" Yeah. Add something to it. , I'm using the agents to help me generate better prompts to be used in the agents, that, again, it knows what it wants 35:16 to see from text to get a better output. I just describe more around the outputs I desire, and let the agents figure out the wording it needs to be able to to run better. , even this, I 35:26 will cycle through a couple times, letting agents modify this code here just to get a better output. . And to add to that, , my my 35:36 personal experience, if you want to say, is those repository-scoped agent instruction files, 35:47 they should never be static. They They need to live with your project. And in fact, I frequently, , in in all live projects I'm working on, I frequently 35:59 have a weekly, if not more often, review session with my agent, where I'm asking the agent, "Have a look at agents.md. understand that this is 36:10 a grounding file that will be included into every single agent session in that project. is that still accurate? Does it still 36:20 reflect where the project has evolved to? Does it still point out, , the the the the 36:30 key do's and don'ts? Does it miss anything? , I I do that very frequently. and particularly when you have fast-evolving 36:40 projects, , that go , that do substantial evolution, quickly. It's very important, , this is not something you write once. and also 36:52 the other thing is Claude and other coding agents have an init function. ? where when you 37:03 go into a new project with them, you can you can do {slash} init and then it will auto-generate those kinds of files for you. And you may think, 37:13 this is a job done , but it's highly highly recommended and highly important for you to actively review what's in agents.md 37:25 or claude.md because it turns out the quality of those files has a huge impact on the quality of all your agent sessions you're running later on. 37:35 . , let's do the same thing here where I'm copying that exact same prompt from before and with no 37:45 further additions, with no further modifications, I'm giving that here. I'm going to let this run. And I know we're probably 37:57 We're at time . at time. , let's let's just do a quick review of where our first attempt got to. and then maybe we can pick up things 38:07 here next time round, if that's . this is the one Sorry, this is 38:19 this I I this is richer prompt, I think. This is Oh, what? obviously my prompt included the instruction to say, "Please create a new branch for 38:30 this, ?" that that's why I I got confused. . this but we can see it here, ? attempt two is 38:40 showing as the as the root node in my explorer. there we go. that that that's that one. if we go up, it's the longer 38:51 prompt. And let's just look very quickly, what has it done? customer.timdl Ooh. Ooh. We're not getting any 39:03 comments. Oh. Oh dear. . that's a fail. [laughter] That's a fail. welcome to the real world of real 39:14 demos and real coding. this is why we do this cuz we we we modified something, but we weren't clear enough. It didn't understand what to do. . let me 39:24 have a quick look. Issue three is about descriptions. Issue four is about display folders. . the the the the the the the the the the the let me make all the edits. 39:34 for some reason it's decided not to make any For some reason it's decided Oh here, business user measure descriptions 39:44 added. , it was very selective. sales amount, margin, cost. go to the sales TMDL. You have 39:54 You're picking the store one . the sa- the sales one maybe has it in there? There is There it is. we Oh, look at that. it was very 40:04 selective. It's done it only for three. and interestingly enough Oh . Yes, we didn't see it 40:14 because I didn't scroll down far enough. this one says , this is sales amount due today. Your cumulative net sales revenue from January 40:24 1st of the current calendar year through the last day visible in the report. Compare this against to see whether that's not too bad. I 40:34 the description . verbose, but I think that's fit for purpose here, ? Because we would not assume a technical user for that, ? That's not the audience. , margin percent 40:50 can we find margin percent? Or control F. 41:08 it would be called measure. Whoopsie. Measure. Or if you just go to the If you're on the sales file and not do the search icon there, just control F on the file. It'll just pop up the search window in the file. 41:22 Oh, here. There we go. There we go. Measure margin percent. Here we go. , here. How many cents of every dollar Oh, I love that. How many cents of every dollar in sales revenue remain after paying for the products sold? A 41:33 result of 30% means that for every this. 100 pounds of 100 dollars of revenue, 30 is gross profit. Compare this across product. Love it. , , 41:44 very good. , That was a good description. Even gave an example. I love that. , . , mixed, ? Yes. It Getting better? It definitely did what we 41:55 wanted, but for some reason it's decided not to do it everywhere, presumably because it was missing context and it wasn't able to produce something. Yes. 42:05 But we don't know, ? Yes. I appreciate we're at time, but this is exciting and I'm I can't wait to see what the other attempt has given us 42:15 but we'll have to wait a week to get that. regardless though, this is a I wanted to end on this final thought here. This is a real process. 42:25 what you're seeing here of multiple attempts, refining a prompt, doing it again, doing it again, doing it again. This is where the real work happens. 42:35 we we said AI will take your jobs. No. We've just made a brand new job which is prompt engineering where we're rethinking what do we really want? 42:45 Is the output really what we needed? Do we need to change it again? Do we let it live as a highly customized prompt or we want to reuse it at the agent level? these are the considerations that we're 42:55 redesigning. We're rethinking this whole process to get better output. , why the the question may be why would you do all this work here to 43:05 figure this stuff out? What you I think you'll see here in a in a couple sessions time here is once we get to something that works, we have a solution. we have a 43:16 more catered prompt or engineered prompt to get us a a regular output that we're more happy with. And to our point earlier in the session which was you're going to build skills, you're 43:27 going to build agents and things this and they will need to evolve with your process, with your skills, with the repo that they're a part of and it will be constantly you're going to need to be 43:38 maturing these things as you go through this process. And this is directly what I've been finding as I'm doing. I'm doing a lot more of pause, hone in, build a prompt, get what 43:49 I want as output and I'm thinking about building process to make sure that that prompt regularly gets reviewed and improved and over time gets better and better. And and this is I think 43:59 really the core principle here is we're showing you how you make good prompt engineering elements and then we'll have to refine them over time as the project evolves and gets 44:10 better. Anyways, Let me if I may, let me use another analogy here. , last time I said the the outputs which we weren't happy with were, , 44:20 what we what you might be getting from a very junior model, ? And the process we're going through is effectively the same process you would go through with your 44:30 junior employee Yeah. by explaining more, by providing more context, by training them that ultimately, , after 44:40 having invested a bit of time, they become senior and and they are able to produce the kinds of outcomes first time whenever you give them a new 44:51 task. And that's very similar to what we're doing with our agents here. Love this. Very very good. All , everyone, thank you very much for joining us for another working session around agentic thinking. We're going to 45:01 keep going down this path. There's much to unpack here. We're just scratching the surface. We hope you enjoyed our demo today. Matthias, thanks for screen sharing today and just diving in and building some amazingly cool 45:11 things. We're going to keep going down this path. Thank you all much and we'll see you next time. Thanks, everyone. 45:21 Agentic thinking. [music] Agentic thinking. [music]