chapters transcript notes
click any line to jump to that moment in the video
0:14 Hello everyone and welcome back to another episode of Agentic Thinking. We're going to talk a lot about these announcements that have happened almost since Friday. [laughter] It's been crazy the amount of new 0:24 announcements we've already had in a couple days and a weekend. Matias, hello again. Hello. Back again. Yes, how are you? It seems that the world is very 0:35 generous to us because there's always something big new to talk about, ? Yes. [laughter] It has been. We haven't had to worry about anything. I don't think we're going to have any shortage of topics 0:45 here for anything Agentic for a foreseeable future. Things are moving fast. we're going to get into a lot of news items here. . Main topic today will 0:55 just really be discussing what's going on with all these tokens. We are the free lunch, as we were saying earlier, is about to be over. I think we're seeing the end of the free lunch 1:05 here slightly and we're starting to see organizations really ramping up appropriate costs to what it cost them to run an Agentic experience or AIs and these large language models. 1:16 I feel we're entering a new era of agents and things where we're we're going to start hearing a lot more around optimization. . a tuning and 1:27 working on the best usage for the model, the use case, the actions, the build. There's going to be a lot of tuning I think is going to be need to be needing to be applied 1:37 here in the near future. . What are your thoughts, Matias? Are we entering this optimization stage? we better be, ? 1:48 Yes. we even made a prediction last week or . It , be very explicit here. The big news item is GitHub has announced a very 2:00 radical change in their Copilot pricing scheme. something honestly we've all been expecting because and we even talked about that very clearly on 2:10 the podcast. GitHub has been incredible with respect to how much very very cheap top-tier AI you've been 2:22 getting through their plans. Yes. With the latest announcement, GitHub moves to the other end of the extreme. Yes. I would argue that they're going to 2:33 be the least interesting one from a commercial point of view because they're moving from what they 2:43 currently call premium requests. And again, we talked about that at length when we had one of our hands-on sessions. a model to a 2:53 usage-based model, which if I understand it correctly, it's only just been announced and we've only just had a few days to look at it, but if I understand correctly, they're going to apply 3:04 full retail pricing of you know, the various model providers and pass them on directly to subscribers. 3:15 if you're if you're on a on a $39 plan moving forward, you get exactly $39 worth of API calls to your models. 3:25 which can be substantially different from where things were on on the or where things have been on on the 3:36 premium request model. And let's let's unpack this a bit, ? one of the reasons why Matias you really enjoyed working with GitHub and the premium request was you were able to have it spin for multiple 3:48 thinking iterations. And if you gave it really good instructions, you're able to leverage a lot of that premium request to have it do an immense amount of thinking. I don't I don't really have 3:58 the numbers here, but , for example, if I just said, "Hello model." ? Something very simple or innocuous. I'm not sending a lot of tokens to it and it's not sending a lot 4:08 of tokens back, ? There's just a very minimal amount of effort between that exchange. . However, and frankly it doesn't do very much in terms of reasoning. Yeah, it it could be doesn't need 4:18 to reason at all. It's just a a simple response, ? But the things that you were sending to your your models and and what I'm sending to GitHub . is a lot more on the reasoning side. Hey, I need 4:28 you to reason about this. I want you to go research this website. I want you to go pull this information down, absorb all the code, distill it down to summary information for me. And then oh, compare 4:38 that to my business, reason about something there. that's a lot of extra token exchanges that it's either doing with itself, it's writing files, it's doing a lot 4:48 more movement of the token number. . even though I was using one premium request for my "Hello, Mr. Model." and you're using a premium request for "Do research on this website or this topic 4:59 or this get repo." it's burning a lot more tokens in that other request. Absolutely. Yes. This is just level setting that expectation a lot more. And I've always called it the free lunch 5:10 whenever I talked about, , which which models should you go with and which provider should you use. it really feels this genuinely is the 5:20 end of an era because if you have been a GitHub Copilot subscriber, you've been getting much AI capability for nothing, ? 5:32 20 or 40 dollars a month is is is minute compared to the actual compute power you were getting in in response. Turns out we have just over a 5:43 month left of that free lunch. free lunch, absolutely. and yeah, absolutely. and , I've 5:54 been creating PRDs, , very very detailed product requirements documents that which I used other top-tier models 6:07 to produce. if you if you if you're feeding them into GitHub Copilot on the premium request model, you 6:18 get if you're , if you if you're lucky, you get the whole thing or 80 90% of it implemented with just one premium request. which is exactly the 6:30 same unit of currency if you will they apply to you if you send a very tiny trivial chat message, ? it's always been really clear that 6:41 economically that wasn't going to work. However, as I said before, I genuinely think that 6:52 as a subscription model the this is going to be a lot less interesting than cloud subscriptions or Codex in particular. 7:03 if we're all moving towards this credit or , pay for a certain amount. the reason why there's tiered payment approaches to this, which is again, my assumption here around these large 7:13 companies where they're investing a lot of infrastructure, they're building data centers, they're trying to leverage the chips to bring them in-house they can do this, ? the idea of this 7:23 is if you pay for a certain minimum amount of something, ? They are able to then better price in or pay for the systems they want to put in place, 7:34 ? it makes it a regular hey, Mike is showing up. You're going to pay the $200 cloud price or the $40 GitHub price. we'll be interesting to see what this 7:45 is going to look from a comparison standpoint cuz there is not a $100 GitHub or $200 or even $300. I know Grok uses a $300 plan as . These 7:55 different AI platforms have different price points, ? the the highest price price point we're seeing with GitHub is 40 bucks. from a what plan you can buy perspective 8:09 we'll have to see how this is going to distill out cuz while they're talking about AI credits for things we don't know for $40 a month what is that going to give you? are 8:19 you going to get how many tokens are you going to get? What is that going to look ? How are we going to monitor and meter that in our applications? What does the harness look to help us show us how 8:29 much of the spend with that we've compute or used at this point in time. this is something else that's going to be quite interesting to see how this shakes out as . , here's my 8:39 understanding of it, ? if you look at the the general per million tokens costs of of the various models out there, , 8:51 very stable and and -known and -documented. as an example, GPT-4 5, you pay $5 for a million input token, you pay $30 for a million output tokens, 9:03 which compares to Anthropic say office 4 7 where you pay $5 for a million input tokens and $25 9:14 for million output tokens. same input price, slightly cheaper output, ? those those are the those are the retail prices per million tokens. 9:24 And there's generally no variability about that. No matter which provider you're using, you it's the same model, you get the same price. My understanding is 9:35 GitHub AI credits moving forward will just apply, , that exact same pricing model directly to the 9:46 budget you have, ? if if you if you happen to accrue a $39 worth of API calls, no matter which model you're using, you've used up your $39. 9:58 In that respect, I don't see any advantage at all of having a GitHub subscription moving forward because you may as just be on a 10:09 pay-as-you-go model, ? Whereas the other subscription models, Codex, Claude, you name it. Yeah. With them, even though 10:20 they have a substantially larger monthly or annual cost , you have a substantial advantage over those API prices, ? if you if you 10:31 pay $100 or $200 to Anthropic for for Claude subscription, you know, you know, you're getting a lot more in terms of those API calls than 10:44 than what what you're paying to them. And it it looks to me the new GitHub model is no longer offering that exact same advantage. 10:54 And it's this is also I think what I observe happening in Microsoft Fabric as , ? There's this kind of concept of inside Microsoft Fabric for premium licensing, they want you to 11:04 pay a certain threshold, ? You're an F2, you're an F4, you're an F8, you're an F64. That is what you're committed to pay pay for and that you can use anything you want underneath that governed bucket of compute. And I 11:16 think this is again, going back to the finance teams, I think the finance teams are really driving a lot of this is we have regular costs we're going to have. We want people that are regular paying a consistent income stream on those 11:28 services. And then when you go above and beyond, you could do the API above pricing, but then it's just, , you're not paying per million tokens, that price goes up per million tokens. 11:39 it's almost , "Hey, if you commit to us and subscribe, we'll discount a little bit more and they'll take a bit hit or hit on the margin because you're a regular customer and you may or 11:50 may not use up all those tokens or not." this is where I think they're going. Also, I I will point out here on the I think it's the fourth link I've put down here, 12:00 there is a link that says, "How do AI credits work?" And I do want to call this out. . There's There's two kind of points here. Each Copilot license comes with a monthly amount of included 12:11 AI credits. Copilot Business will have 1,900 AI credits. Copilot Enterprise will have 3,900. 12:21 I'm not sure why they didn't just round up to another number, but apparently that's what that's the credits are using. They do have a promotional amount for any existing customers. if you're an existing customer of Business or Enterprise, you 12:33 get a huge jump for a period of time. from June 1st to September of 2026, Copilot Business will be 3,000 total AI 12:44 included credits and GitHub Copilot Enterprise will be 7,000. those are those credit numbers that you're going to be looking for. again, we're we're trying to interpret what we were using before premium 12:56 requests and we're going to these AI credit experiences as . It'll be interesting to see how this is all going to pan out and what the spend of things will change to. coming back to your earlier question, 13:07 I think what we're going to have to talk a lot more about moving forward is optimization, ? How do you optimize? What what does optimization even mean in this context? 13:17 and that's probably where we're going to see a lot of innovation, ? let me just start here, ? for my point of view, there are two 13:27 things. One is what's the model for the task, you know? If if you are if if you're if you're going to go, , to to your 13:37 neighboring town that's 30 miles away, you're not going to want to jump on a a bullet train, ? 13:48 that's something you do if you want to get to the other side of the the continent. But that's the equivalent of using the latest Opus model for for 14:00 every request, ? Which far, because generally it's been pretty cheap, may have been the default mode for many 14:10 people. that's that's one thing to reconsider moving forward if you want to be more price conscious. figuring out how to break down your eight agentic 14:23 workflows that some portions of your tasks are routed to top-tier, most expensive models and others 14:33 are going to cheaper models, , a mini model, a nano model, a Haiku, things that. They They come with amazing capabilities, but also at a 14:45 tiny fraction of of the cost. And then the other one is really thinking about how tokens are accrued, ? Generally 14:55 input tokens are the thing we're worried about as in, , what are we what are we sending up to the model? And 15:07 if we're thinking engineering context, , having having a repository with files and and all of that, there's a lot 15:18 you can do around optimization even when it comes to optimizing your Claude.md or agents.md instruction files, ? 15:29 Many people don't really pay much attention to them or they have them generate once when they when they start a new agentic session, 15:39 it turns out that what's in those files gets appended or prepended to every single request you're sending across. if there's a lot of stuff in there that's too verbose or if there's 15:51 stuff in there that's outdated or if there's stuff in there that's not relevant for you to include in every single request, you are being extremely wasteful, ? And 16:01 it's things that that you may want to look at much more consistently moving forward. I don't I'm I'm not sure if we really landed on the term of what this 16:11 optimization world's going to look for us, but already I'm starting to see things. , Microsoft has been making moves inside Azure AI Foundry. Azure AI Foundry is I'm guessing 16:21 that's the service that they're using to serve this to GitHub Copilot. That's I'm guessing there's some part of that where Microsoft internal is using that system to help build guardrails 16:31 around your large language models to to manage and and validate the the amount of usage of AI from GitHub Copilot. Maybe there's a different system that's running there, I'm not sure exactly what they're using, but 16:42 inside the Foundry system, there's this idea of a multi-agent workflow, ? You you send something in and then there's , , multiple agents or even I guess what 16:54 we've been talking about on the podcast here a little bit is custom agents. . Custom agents in our world, we can command a certain agent to do something and that agent can in the 17:05 instructions of that agent, you could say use this model, ? For this SQL task we're going to use, ? We can leverage a 17:15 planning agent, build out what we want to build and then hand that over to the SQL agent and then use a a different model. I think multi-agent workflows 17:25 going to become much more prevalent and that conversation is going to become much more front and center around this we need effective tokens. It's no use what we said earlier, saying hello or my 17:37 AI assistant thing that I have built on my computer on my desktop, it makes no use for me to have this really high-end reasoning model if I don't need to use it all the time for every single request, ? Routing that detail 17:48 down. One thing I also I I wanted to note here, Matthias, I I saw an interesting article and I was putting this in my notes as . there is starting to trend, I believe, 18:00 when you start taking models and you instead of, , changing the weights or or adding a rag or some other information to them taking these higher-end models and 18:11 adjusting the weights of them for very specific tasks. This also feels I'm starting to hear some very initial conversations around what this looks 18:21 . Training models on your company's data specifically and making custom model this model is designed on top of the data that lives in our company. 18:31 This could be instructions, this could be, HR documentation, whatever that that is. But what that does is it informs and adjust the weights slightly to be more designed for that 18:41 particular company's information. . What I'm what I'm hearing is this runs at an incredibly efficient and low token usage as . the era of customized models may also 18:53 be coming up here as . What are your thoughts on this? Are you hearing the same thing in this space? Yeah, absolutely. Although , that's something where Azure 19:03 Foundry is is very good in providing capabilities for you to train your own models. In fact, when you deploy standard models, Foundry 19:14 explicitly calls it deploy a base model, ? If if you That's true. You know, if you're if you're if you're deploying, , Sonnet 4 6 or something that, 19:24 they explicit they call it base model with the idea that you may then want to use that as a starting point for custom tweaking. Although generally you require a huge 19:35 amount of input to retrain a model that, ? Remember, they all come with billions and billions of parameters in terms of their training 19:46 data. one more thing that's we're we're part of this conversation should we explore a bit more or further around this? We know we have a couple 19:56 more links here. What are your other impressions around this, Matias? we I'm as far as GitHub is is concerned, I'm 20:08 I'm really curious to see how this pans out. As you can imagine, there's probably a lot of frustration and out there in 20:18 in the user base. , many people have made a big bets and investments into GitHub based workflows. 20:28 Yeah. including, , all the financials linked to that. , you may have to reconsider that very quickly. 20:39 At the same time, they've done an incredible job at providing great tooling as part of the Copilot ecosystem, 20:49 particularly everything that lives in VS Code around the VS Code chat experience. I'm very keen to see where that's 20:59 heading. but I think what we generally will need much more of is visibility with respect to how 21:11 your agentic sessions and workflows consume tokens and how that translates into 21:21 dollars and cents. And I think that's I'm I'm pretty sure the community if not some of the big players 21:33 will look into providing better tooling here, . If if you want to optimize, you've got to have data to begin with. , we're data people, ? We we we we know that very . 21:44 If you want to optimize something, you've got to know where you where those tokens are being consumed and and what that means what that means in terms of 21:57 your bill. The other thing is it's probably really worthwhile moving forward for anyone who wants is using AI 22:08 capabilities at scale to invest a lot more into AB testing. As in identify common workflows, , that 22:19 eats up a lot of tokens within your business and and and then do some test runs of those exact same 22:30 workflows against your current model and compare it to how a cheaper model would perform. And you you may 22:40 find that a model that comes at 20 or 50% of of the cost performs equally or or or just as , ? , I think 22:50 that experimentation and and comparison is probably a really worthwhile investment going forward. And personally, I'm definitely going to be doing much more of that. 23:02 This is interesting. I I I You bring up this concept of experimentation and I think of this is how would I implement this in my company? I wouldn't expect everyone in my company using AI to start 23:12 experimenting with things cuz that would just exacerbate the problem of we have more people running not just building code one way with whatever the the model is, Opus 4.6. we have 23:23 people building code with 4.6 and trying to build code other ways and then comparing the output. , a measurement stick, what whatever the the meter will be, will have to be very 23:35 intelligent to be able to compare the two outputs from two different models requesting the same request. But I don't again Let me pull back here a little bit is 23:45 experimentation of this really shouldn't be done by everyone in your organization. No. This is This is the person who's leading the helm on AI, really into the weeds, understands the 23:57 strategy and and sourcing behind this one. And those are the individuals that you're going to want to leverage to do additional, , workflow stuff. But as you're as you're designing the AI 24:08 strategy for your organization, those are the individuals are going to say, "This is how we are going to implement this in an in an efficient way, ?" When talking about what tools are we 24:19 going to use? Are we going to use Azure Foundry? Are we going to stick with GitHub Copilot? What are our spends and thresholds on spends? I do want to bring this up as in related to this experimentation piece. 24:30 CTOs are going to have a huge task at hand figuring out how this AI cost is going to influence them. maybe in the in the tech industry in general, we're seeing jobs 24:40 leaving. Companies are reducing the number of people at companies to make way for spend for AI. but if you look at this Yahoo article I'll put here in place. I think I've talked about this 24:50 before. I found this to be quite interesting as a talking point here is Yahoo Finance is is talking about Uber's Anthropic AI push hits a wall, that CTO 25:01 says, ? The budget struggles. They have 3.4 billion dollars of spend and that is their research R&D budget for all of 2026. And because many 25:12 engineers are moving towards AI, finding a lot of value from it. Albeit, it sounds there's still value coming from it. But the adoption rate of AI on every single research and 25:23 developer inside Uber has eaten up their entire budget. , they're already done. Their AI budget has burned through their research and development budget for 2026 already. And they got to figure out 25:34 where is the money coming from to to continue to do research and development. Is this less headcount? Is this giving engineers more agentic pieces? Again, I what I'm looking at this going, 25:46 this is the optimization problem. away, I'm going to be throwing a couple really smart AI people at this to say, "How do we experiment? 25:56 What systems do we need to build to start routing the appropriate AI request to the models and start saving ourselves money?" Cuz we we don't want to It feels we've been tricked a 26:06 little bit. we're in a an analogy. It's the turning up the water the hot water on the frog in the pot. Is this the , if you put a frog in a pot of really hot 26:16 water, it immediately jumped out. it doesn't it. But if you put a frog in a pot of water and slowly turn up the heat, the frog doesn't quite understand the heat's been rising this whole time and all of a sudden it's 26:26 cooked, ? , I feel this has been a bit of the game that we've been given to us from the AI industry. and maybe indirectly or maybe intentionally, I don't know. . But 26:37 this is I feel we've been given these free tokens for long and people are , "Wow, this is impactful." And companies are adopting it and we're starting to see culture shift in lieu of this AI space. 26:49 And the water's the heat's getting turned up slowly. And we're starting to see , we've got to really start thinking about does does this make sense? Are we going to apply it everywhere for everything? It is a 26:59 revolutionary technology, but only if the price of tokens continually stays low. . . What's really key here, , at an enterprise scale, observability, 27:10 ? Yeah. , if if if you're using AI capabilities at scale, you you've got to have an observability solution where you 27:22 record and trace all your agent flows that ultimately you have data you can look at and analyze and figure out, 27:34 how much would we save without compromising on outputs if we switched to a different tier 27:44 of models, for instance, ? , that's really really key. Anyone who starts with AI got to think about that stuff early on. 27:54 Any session trace you're not recording is a lost opportunity down the line. And It's very true. that's one of the 2026 28:08 tasks for you, I would say, ? Also, I , any company who used to have a substantial software engineering team in-house moving forward 28:20 will probably have an agent engineering team, , taking care of exactly those those kinds of considerations. All here here's the new role. we're starting to talk about we're we're , "Oh, it's going to take all 28:30 of our jobs away." Already, we're starting to see the AI engineer or the AI architect showing up. And this is part of that process. 28:40 How do you monitor what the agents are doing? How do we handle this at scale? How do we optimize what that looks ? We need to have different paths to figure out what's what's being built here. . And I think this also brings 28:50 up the last two articles we have on on tap here for talking about. The last two articles are talking more around one, there's a an update. , there was I think it was before they announced this 29:01 pause and pricing change. There was a GitHub availability issue. I think Matias, you and I were talking about this. We were talking about this on on on chat back and forth. We're , "-oh, 29:12 GitHub's down, throwing errors, Copilot can't resolve." . we we we were immediately impacted. At least I was immediately impacted by it. And then it it sounded 29:22 you were as , ? Yes. GitHub used to be a system that I was always able to completely blindly rely on. And there've been several 29:33 incidents recently where they've had prolonged outages. And thankfully, I think just today they had a 29:43 big blog post doing a bit of a recap here. , I want to unpack this cuz there's a really interesting graphic 29:54 in here that they put in this item. I think this is also a trend we're seeing as , Matthias, is as we look at GitHub in general, even my own workflows, 30:05 I am heavily committing and running a lot more actions all the time . And there there was a lot of things that I was doing in software development we would build some code, I would run it 30:15 on localhost, and then I would make some changes, and then I would push it, and then let it go through the CI/CD process. Internally, I'm really pushing on our team to commit a lot more, make a lot 30:26 more branches, build have the automation set up. And one of the advantages I've been seeing with working with agents is when you have an agent and you want it to test or build or create apps for you, 30:38 I'm finding an immense amount of value for me around getting the CI/CD process set up very early in my process. Just the very basic I'm going to 30:48 have a repo and it's going to publish to a Azure static web app or it's going to be GitHub pages. Without even doing any design work for what that thing is, I want the very first step, does it 30:59 successfully build? And , what I'm doing is, , I'm making seven, eight, nine branches in a day and 31:09 pull requesting them back into main regularly within a day. That was not the case before. And , if you just take one developer and 5x, 9x, 10x the amount 31:20 of commits going back, the record acceleration chart that they show here on this GitHub page really makes sense because agents are accelerating the speed in which we can 31:31 develop and build the code. And , you know, they're saying merge pull requests up to 90 million . Commits are 1.4 billion. The new repos per month are 20 31:43 million the number of repos that are showing up. We were talking about this on episode one. I feel my new social media is GitHub. . that's 31:54 there's constant new repos being presented, but it's because the agent can build them. And I'm looking at GitHub repos , and I don't know if you can identify this as . Just looking at a repo, I can tell away 32:07 based on the readme file, was this built by an agent? Because when there's a human repo, it's two sentences, this project does this. That's that's the human repo. But when an agent builds 32:18 a repo, it's , here's the folder structure, here's why this exists, here's the outcomes, here's the deliverables. all these extra sections that would be very helpful for a user to understand what the project is 32:29 about. You didn't get those before. they show up and they're all there on GitHub, I can clearly see there's a lot of maybe more em dashes 32:39 [laughter] on your your GitHub page. that's a dead giveaway. agents are building this stuff. , I I think this is really important that GitHub is posting post posting this out. And also, 32:50 I think it's important to acknowledge GitHub is getting a lot of traction. This is becoming a really solid pattern. Automation, continuous integration, 33:00 continuous deployment with agents is here to stay. It's going to accelerate even further. 33:11 It just got me thinking, . If you think about GitHub from as a business model, ? They don't charge you per commit. They don't charge you per pull request, ? , 33:22 it's must be really, really challenging for them because their user base, presumably, is relatively constant or probably has 33:34 the same growth rate as it always has had. But the individual outputs in terms of activities and and ultimately pressure on the GitHub system would have 33:45 multiplied. Yeah. , they're they're not generating additional revenue, but they have substantially more cost. , every 33:55 commit you push into GitHub ultimately is a cost for them, not not necessarily for you. I I would not I would not want to be that business where 34:05 you're completely at the receiving end of the a high volume influx. Exactly. [laughter] Yeah. Where I do think this is interesting though is , you're saying on the 34:15 merge pull request or committing to the system is not charged, ? , to your point there, I agree with you on that one. They're not limiting on how many commits you can make, which is great. That's really the way it should 34:25 be. Where they are getting you though is they're getting you on the agent costs and they're getting you on the actions that are running. , there's this action minutes that they charge you for on that side. 34:35 And again, I don't know if this is going to change the GitHub pricing model at some point. Maybe it will, maybe it won't. I'm not quite sure yet. But the fact that it's ramping up 34:45 fast, and again, I would argue I've changing how I build . How I how I think about building is fundamentally shifted, and I'm doing a 34:55 lot more with continuous integration, continuous deployment. , I'm spending more on the actions side of the world than I am on the other side. that's an area that I'm just I'm 35:06 investing more back into the GitHub platform because I'm using it more. But it's it's a balance of value and cost, ? I'm finding more value by getting more software out quickly, and I can see it on my phone, my computer, and have it 35:17 just deployed in these different branches, and then roll to roll to main versus the price that it cost me to run an action is almost next to nothing. , I'm I'm happy to pay an extra 20 bucks a 35:28 month or 50 bucks a month on getting unlimited actions for my company because of the volume of what we're able to produce on the other side of things. Again, this whole dynamic of this software engineering world is really 35:38 shifting a lot. And I think this message around what Anthropic is doing, which is anyone is a creator . . Tommy and I this morning were talking about Notion. Notion is a really 35:48 interesting program that helps you kind of capture your thoughts and ideas, and it mixes an agent with your note-taking systems and helps you build some automations there. But Notion is 35:59 interesting in that they are heavily pushing into this this agentic space. And and and that's also creating this 36:11 idea of in their company, they're directly saying anyone can build. PMs, leaders, there's a high level of autonomy that's being given to the broader part of 36:21 the organization. Build something that works, we'll test it internally, and if it sniffs good, ? If it if it works, if people adopt it, they push it. This 36:31 is the same development pattern that Anthropic is using, ? They're anyone can create. They're really encouraging a lot of their employees just to build something that works and solves a problem. Solve problems. That's 36:41 what they're doing. across the board in their company. And , what's happening, this is where I think Claude design came from. Claude code, the actual application. Same same 36:51 thing was developed. That harness was created by employees saying this is working for us, this is helping us build stuff. And when it became adopted internally to the organization, they're 37:02 sufficient enough. Let's make it a real product and ship it. , I think this this dog fooding process of being able to build things continually at any level of the organization is going to 37:13 really fundamentally shift what we need to think about. And how we to your point, this probably puts more stress on what you made earlier, Matthias, which is we need better observability. 37:25 ? Absolutely. It's the same thing for Power BI. Yeah. Yeah. , I would say anyone who's currently on on a GitHub plan, 37:35 use the time you've got left till May 31st very, very wisely. It's never going to be as cheap as it is . That's a good point. , 37:45 with that being said, I think that's a good note to end here. , I think the free lunch is officially over. We're going to be starving 37:58 for more tokens here in the near future, the opportunity here is move into optimization. Move into how can you do the same task with less tokens? That is going to continually be a skill 38:08 that's going to be invested upon. Build observation tools. Look for observation tools. There's going to be more out there, I'm sure they're going to show up here shortly and helping us to evaluate these things. , that's 38:18 these are opportunities I think for the agentic community in general to start building and pushing into. And then we'll see how that is then going to portray or be built into most of 38:28 the Microsoft and Fabric world as . Awesome. What What are we doing on Friday, Mike? On this Friday, we're going to have another demo coming. , this Friday we'll be doing another episode on the 38:38 agentic thinking podcast. We'll do another one same time around 10:30 Central Standard Time. We're going to be doing a demo. Matthias, what's going to be more in our demo? What are you thinking for this this Friday's 38:48 demo? back in GitHub whilst we can. [laughter] Exactly. and yeah, we're we're we're going to combine some skills with 39:00 MCP servers around a Power BI and semantic model engineering workflow. And And we've already built some issues in 39:10 GitHub with your current solution. , we've got to crank through all these issues that we just created for ourselves and get our GitHub Copilot on our premium requests, rip through those things quickly before we 39:21 start getting charged more for our tokens. , that'll be our Friday demo. Thank you Thank you all much for participating and jumping into the podcast today. We hope you've enjoyed this one. Hope you found some value from 39:32 how we're unpacking this new agentic world that we're living in and just keeping you up on news around this agentic space and how experts or semi-experts, I guess we're somewhat No one's really an expert 39:43 really. We're all trying to become experts at this point. But as we talk through these issues, challenges, and unpack what agentic building looks for our companies and potentially for your organization as . 39:54 Thank you Matthias. This has been a lot of fun again. Looking forward to seeing everyone on Friday. Cheers. Bye.