Why I'm Burning Down Every SaaS Tool in My Business With AI Agents

**[00:00:00]** I don't know if you heard close to that that also like April beginning of April thing I I saw a post by Anthropic and I'm like oh I clicked on the video. It was very weird to see that video. I I thought um because they they sort of explained how they were mapping its neural network and trying to see if it mapped certain things to emotions. Like if it showed a happy scene it would do like happy emotions or like if it was sad scene or stuff like that, right? or like if it did something wrong, it would show like and so some parts of some certain neurons would turn on, right? Like a cat scan would. Yep. Yep. Um and I thought it weird like when I saw that video I was like that's that's something strange to publish like in a little video like they don't publish as much as you would expect in videos but then they publish that like just out of the view in a little video like a nice And who who was doing this? This is post like like one of their like pure like whatever the little videos with the y tad background and I was like huh okay. But then Travis tells me they're like they're going through the the report and everything of of the exploits and all that. Yep. And like it turns out it was like it did some things like deleted a file or something and it was like concealing some stuff and like it it felt guilt when it did some things that it knew it shouldn't like and we know it's not really feeling or anything but it's like the representation is there. So it's kind of creepy getting into the whole sci-fi thing again. It's like those things light up when it knows it's doing something that it shouldn't do. And so it was intentionally concealing some things from the researchers like doing something but reporting something else. Yeah, I've heard about this happening in models now like and they know they're being tested and they'll just go find and research the answer and then just give the answer without actually knowing the problem. Like it's getting harder and harder to determine like are you really smart or are you just figuring out clever ways of **[00:02:00]** getting around the problem? Of course. And that's what's bound to happen as you feed them more stuff. Yeah. And so for me it's always a simulation like it's never going to as much as they say like artificial int. No, it's Yeah, it's artificial, but like a computer, like a machine. A machine is artificial intelligence like because it helps you do things that like the the factory. That's the whole that's AI. No, it's a different version of it. But it's all simulated at the end of the day because it doesn't really have uh feel as a human, right? It doesn't have the way of a human has these feelings. Now, you could have the whole theological debate and everything and be like, "Oh, well, yeah, but where do you know where that is? Is that it? Are we not all like electric synopsis and things?" So, that's where I go to as well. It's like, I mean, they modeled the whole neural network thing off of a brain and how it functions as well. There's all these neural networks and and things. I did find it. So, I want to keep going on that point, but I also want to I was doing a little Googling while you were away for a second or two, and I was looking at like, so I don't know if this is right or not, but people are starting to figure out what's going on in like this space. And so, I found this interesting. Someone wrote a particle what they think that this model has 10 trillion parameters. Wow. Now, by comparison, ChattP 5.4 I think has like 1.2 trillion parameters. So, that's kind of like this the size that we're talking about. big but they also said here that the model was using a and again there's a p there's a PDF that comes out of claude now so there's claude has published a a paper I think called the mythos preview system card where it's like some of their research there and I think this is where this is kind of starting to come out from but it sounds like and again this needs to be fact checked this is all you know I'm not sure this is exactly right from this website I'm looking at it's saying it's a mixture of experts there's five experts and even though it has 10 trillion parameters in the model to your point around the activation of like what **[00:04:00]** part of the model is actually active at certain times of when it's doing something. When you ask it a question, it kind of routes. It's like doing some a little bit of internal routing. So even though it's a even though it's a 10 trillion parameter model, it's only leveraging a portion of the network to actually do the inferencing or thinking or reasoning on a portion of that. So it's big, but it also is small. And it says if you leverage like look at the time and it's equivalent to like a 800 billion or 1.2 trillion on any one prompt. And then it also talks about like there's a memory architecture they're talking about which is now they're they're introducing a new tiered attention. Have you heard about this? No. So tiered attention sounds like recent window context items. When you prompt it, the more recent the chat is, the more emphasis it gets on that, right? and older conversations in the text that it sends to you becomes less attention. Right? So you almost like waiting like the most recent like when we talk we're waiting the conversation about what we're talking about and discussing right now like this is right of course contextually and then but you still have reference to reference to the old stuff right but it's there but I have to like think about it and recall it to bring it back right what we were talking about earlier. So, it sounds like this is going to now enable this tiered attention which will open up a context window of 4 million tokens. Wow, that's better. Like anything else I've ever seen. I've never seen anything more than a million. And I don't even And then even then, I don't even know if I've ever used all of a million token window anyways. And is it any good? I don't know. Like I can't I couldn't tell you. And that's Yeah. Well, that brings another thing like when I see the 1 million token with the Opus 4.6 six. Yep. I mean, I use the model, but do I want to? It's like my wife takes a lot of pictures. Like a lot in the sense that we have 16 **[00:06:00]** terabytes of usage in our NAS. That's it of pictures. I think I think I'm I think 16 terabytes. I think I'm I mean I think and it's and it's accelerating like so like I've been collecting photo same thing. I'm collecting. I need to get a Naz because I have too much stuff sitting on computers nowadays. And so I it's all the pictures and things like same. And now I'm looking at like as my devices get better, as the cameras get better, as the videos get longer, as the videos get higher quality, every year I'm almost doubling the amount of gigs I'm generating every year in just photos and videos. But here's the thing. When we got like a one TBTE iPhone, Yep. you tend to go and use it. What happens if that iPhone dies? Like you have a 256 gig. Yeah. You you you have to be constantly passing it over to the server, right? Storing it. You have a one terabyte one. One, yeah, terrible issue if it dies or iCloud dies or whatever. Y but second, even if you want to transfer the pictures, like things break down. The the Mac thing to transfer pictures doesn't support it. It's like it just gets stuck. It's it's too much. Yes. And now imagine they start making these like two teraby ter like do you really want to use it? Same goes for context windows. Like one million context window. Yes. But you're charging me a ton of money for per turn because of the token usage. I'd rather have little conversations. Correct. And like the memory start over. Yes. Go recall like like a skill like go get it when you need it kind of thing. And as opposed to loading everything. This is the same problem we have with like the MCP server. I feel like same like you just turn on MP server. There's there's also a new setting I think I saw from Anthropic which is like you can now autodetect skills when it needs to detect them. I think yes that's better instead of loading them. Yeah. Yeah. Instead of just loading the whole thing, right? Just hey, I have skills about this thing and then when you ask a question it's like let me go look at this. Let me go look at my tools in my MCP server. So I think there's also like **[00:08:00]** people are trying to solve like the whole like we know MCP servers loading everything is not a good idea. People are ragging on it. They're like well we can just fix that. like just make it a bit better. This is a crazy world we're living in right now. I'm super excited about I mean this is all exciting to me. I'm now feel like I'm focusing a lot more on how do I leverage this in my team? How do I get people excited about this? How do I transition other companies to start figuring out this is the way of building things net moving forward. And so that's just that's just kind of what I'm unpacking now is like okay how do I transform my company? How do I rethink everything we're doing? How do we build systems across team members? Because I'm learning something at a different speed than someone else is. And the the name of the game I think there was like this hesitation around in internal to us and you may see it as well. People don't know what to where to start. And now we're getting to the place of like well we just need we just need to start. We just we just start going just start learning because we'll figure out the rest later. I think the the way you put it is very true. A lot of time, especially anyone older than like 20, 15 even will maybe not proud as much because of that idea of and especially like 30 and above, you feel you got to break something. Yep. And with these with the models and all that, it's like just just go like just ask it to do stuff. If it does something that you don't want it, just guide it. It's a tool. It's not super intelligent. It's just a tool. it'll make mistakes. Guide it towards where you want it. You feel it's not going anywhere, open a new session, start talking again. Do it another time. You want to try multiple things, do it in parallel, do a bunch of them and see which one sticks. Like it's a break. It's my problem now. Like now I'm like when I look at these models and like if it doesn't get what I want out, it was because I wasn't descriptive enough on the front side of it to get it to do what I want. So like the also the other thing here I'm trying to unpack is my family. Like what does this look like for my kids? How do I start teaching them how to do things? What kind of creative things are they **[00:10:00]** going to come up and develop? My my son is doing some lawnmowing service prime availability to make an app for this. Like let's let's take an opportunity like you have a business you're trying to run. Even if it's small form, but everyone's got the internet. Let's let's build something. Let's build something simple that's like, hey, this is a simpleuler app. Here's when you're going to be available. We're going to set you up on our calendar. Here's the calendar. Here's the website. Bing, bing, bing. We'll put you on the on the this is when we're going to do it. So, we're just going to put it all on the count. So, that way he knows and the customer knows all the stuff can kind of be wired together and you know, hey, do you want to, you know, pay you know, pay for your session, right? Great. Here's the here's the Venmo link right inside, you know, your scheduled session and when it's been paid and he acknowledges that he got paid, he'll mark off, okay, you've paid for these lawnmowers. We'll we'll just do the lawnmowers and these are the ones you have paid for and haven't paid for. It's just a simple little just a simple tracker. But this is all possible now. And I want I look at this going like I want my family, particularly my my offspring, my kids to start looking at this as like I can solve that anything. I want to equip you with the tooling that makes like the stuff that you thought you had to learn to get to the over the hurdle step here is not as difficult as you thought it was. And I'm even now I'm I'm hiring some interns from my alma mater. Um and what I'm teaching the intern interns is different than what they're learning in school. But to your point in school they're learning from 45 50s and 60 year olds like they're not adopting this. They're they're not even teaching this agentic way of development and building things. And and even now, so I I don't know if all schools are like this. It's just maybe just the one that I'm working with right now, but like I have a suspicion that schools are typically 5 to 10 years behind and now the gap is feeling even wider because stuff is moving so quickly. **[00:12:00]** Things are stacking up so so fast now with the gains that we're getting. I'm in the position of like I need to teach as much of my knowledge and gain to to my audience because that's just going to propel you that much further ahead of everyone else who's going to start learning this in like 6 months or a year from now. Oh, for sure. And that example is great, the lawnmowing thing. Yeah. Because a lot of like take the the guy that does a lot he has he has a team, right? and they come in, but they're smart because they they've got most of the houses on the block. Yeah. So, come once and they they do like eight houses. And so, they don't just they don't just do your yard. They just drive through each yard like this all across all eight houses. We're sure they could do that. But they do that outside. I see. Like like sometimes they go have to but it's and it's smart. It's really smart to do it that way. Correct. But it's still even though they do that, it's still like when are they going to bill you? Did they forget to bill you? You have to ask them like did you have to remember that they went that day where you weren't home? Like that sort of stuff still. And it's all it's all I have I have a lawn care service and they keep I have a paper invoice like it shows up at my front like they do the thing they sign off the stuff and it shows up my doororknob like what are we doing here? Yeah. Yeah. Because you could just build a system and have and with mapping and everything and be like now if you get more people on the block you can see like these people are using it and maybe you get it everyone gets a discount. We're all doing it together and it's all the same day and you can map out and you can be like there's going to be rain. We're going to move it one day and everyone knows it's all there. You get like this and so you can build that in an hour with like with the care of doing it right. It'll actually take like five minutes to build. Yeah. But if you really want to do it then yeah and that goes for everything. Every single service based business can have like a micro tool like that that just saves them so much time and **[00:14:00]** it'll work and you can something you're hitting on a point that I like here because it's not just that for every service based app. It's like traditionally our companies only thought about buying one or owning one or two pieces of software internally right now. Even for my own company, I'm like, what do I pay for software-wise? Mhm. What of the software can I just throw away? Any little friction I get on any feature of an app that I doesn't work quite the way I want, I'm like, I could probably rebuild that. It might be better for me just to rebuild it. And then I own it full outright and I can then talk to my agent and rebuild whatever I want. And now I put the data where I can see it, put the data where I can collect it, have more access to like this is incredible. So, I I think this is not going to There is a bit of doom and gloom around like the, you know, it's going to replace your jobs. Yes, it will. But companies aren't going to have one or two apps anymore. They're going to have 50 apps or 60 things that are all bolted together and doing very specific things. And so, right now, we're building like these little mini apps that do like one little tiny thing. You know, to your point, scheduling, moving things around on a calendar. Okay, fine. But it's now going to be okay, we don't just want one of them, we need seven of them. And Armando, you and I are working on projects right now. This is already starting to materialize into like there are already needs to have multiple apps being developed for a single business. You now need to be able to integrate them all together, but it makes it still feel like a a uniform product. And that's where we're going. Like I think that's what's happening. We're we're going to get it's Javin's paradox again, right? Right? We we talked about this a number of times, right? You you bring the price to build something down. You bring the price to consume something down. You just get more of it. And we're going to need more people to manage it and govern it and make it better. Yes. Yes. And it's a a balance, right, of like what should I build, what should I not build, what **[00:16:00]** should be packaged as one versus little things. I just saw a post from from a a friend from a while back that owns multiple companies and and then they he owns I think more than one software company and he was saying like I don't want a super app that has everything like I want a company that does one thing very well and that supports it very well and and that's another position as well like the problem with having a giant app that does everything is that it might not do anything really really well versus having like one that's really specific or that at least has teams for each specific thing. Yes. And and you get just the little piece like you were saying for the models like individualized models that work very well for that. So I don't care if the parent company owns all of them but the fact that I could just get that part and it works well because it has a team that's completely aligned to that part. I like that part. I I think this is where we're going right now. It's going to be a lot of like little tiny things built in like it's almost like Lego bricks for like your apps now. Lego bricks for your company, right? What are the what are the what are the middle how can I stitch all these bricks together in a way and and now you know Armando we're talking about this as well. Any app I build has to be accessible by an agent experience. So what does an authentic what does authentication look like for my agent? How do I get them in? What can they do? What? Like, so I'm actually I'm actively working right now on some side projects where I'm saying, "Okay, the app is cool. I can use it, but I really want my agent now to talk directly to it because I need the app occasionally to look at what's going on, but in reality, I don't really want to run the app. I want the agent to be like the front and center part." And so Satia just came out saying something like, I thought this was super fascinating. He's like the business logic of businesses have traditionally been bolted into apps the software as a service solution and now the business logic is lifting its way out of the app and moving into the agent skills MCP **[00:18:00]** servers right so giving it the ability to store and save data but the business logic of how my business runs is lifting its way out of the apps and into the agents because now I can offload reasoning. Before it was programmatic reasoning. I programmed the reasoning into my app and now I'm giving less programmatic instruction and letting the agent do this. I'm excited about this. Only big concern is this whole thing runs on the token economy. Yes, it does not matter whether you're mythos, if you're deepseek, if you're doesn't matter who you are. This whole thing is a token economy. And some people will be more efficient with their tokens. others will not be. And the recent announcement of OpenClaw blocking the API usage from Anthropic is huge because as soon as one person does it, there's going to be a cascading effect across other tools as well and it's going to continue pulling people out of like so and again how I read this is OpenClaw was a harness that harnesses the AI. The AI is then used in that harness. Anthropic basically just said, "No more." Like, "It's already abusing our systems already. We don't want you leveraging this in this way. So, we're going to delete the harness. You can no longer use a third party harness. You can only use ours moving forward." And so, this is where I I'm more frequently communicating to my customers and to other people, what is the harness? I have to explain the harness. What is a harness when we talk about AI? And so, I think that's really going to be key. The AI is okay in and of itself, but the harness is the secret sauce. The harness is where the functionality comes from. I really got excited about OpenClaw because it did a lot of things that I just needed to do. Make files, check in repos, look at code, research the internet. Like these **[00:20:00]** are a lot of skills that at the time the harnesses that were given to me didn't have that. Now you look at what everyone else is copying, everyone's copying the pattern, right? So that's already now inside anthropic. They've got their own but it's it's not there yet. Like I really disliked Anthropic's position in that because yes, you're subsidizing the subscriptions. So I understand you don't want abuse, but you you already put in place measures to protect from that. That's why you have sessions, the cycling 5our sessions and all that. Correct. Doesn't that mean that regardless of if they're using Claude Code or they're using OpenClaw, they're going to have to stick within their se their session limits which and that's why I personally was using the pro plan and then when I started using Claude Code a lot I have to move to the max plan the $100 max plan and then when I used OpenClaw I had to move to the $200 max plan. Mhm. That's good for them, right? It's making people realize that they need to step up in those plans and then maybe they're going to need to use extra usage if they go further. And I don't see what's wrong with that, right? Just keep it that way. Not block it off completely and just segregate a bunch of your potential user base because to be honest, OpenClaw is so much better with Opus compared to Codeex. It just is. It's amazingly better. So now Yeah. Now I'm stuck using OpenClaw with Codeex, which is decent. Fortunately, JBT just today or yesterday announced that they're doing a $100 max plan. They don't you don't have to do the $200 that you need to get Codeex. Man, these these subscriptions are expensive. We use it for work. So, so it's an expense and it makes sense for what we're doing. But for for regular users and they want to experience all this, it's too expensive. like this should all work with the whatever five $20 a month plan in my opinion and just have limits on it and the way we stumbled to them at the beginning being like I hit my limits well maybe I'll just wait five hours and that's fine **[00:22:00]** or I need to pay more. Yeah, correct and that's up to you but I don't see the point of limiting it unless like them they are building tools to sort of emulate what the other tool did. Yep. Man, I I I think that's what's happening. Honestly, I've tried it. It's not It's okay, but it's not the same. Well, let let me give you another hook on this one as well. So, if you want to pay separately and or also pay as you go method, Azure Foundry is a great opportunity here. Yeah, I I started using it. I've been I've been touting the the powers of it for a while here. And it's kind of exciting because I really do get access to any kind of model that I want, right? I can I can go get I have a model that runs DeepSeek. I have a model that's running Quinn. I have a model that's running um chat GPT 5.4. I have another model running opus 4.6. Like so or sorry 6.4 or five or whatever the heck we're at now. I don't remember. 46. 46. Yeah, that number. Yeah. But like to me like this is like the Microsoft I think is definitely positioning themselves as a you send your prompts to us, we run the machines, you get the answers back. And they're doing a pay paytoplay kind of method. Whatever you use is what you pay for. I like that method. I also am not probably abusing it like I was doing when I had like a subscriptionbased you know GitHub copilot and and using it there. I'm finding that when I do real development work, I want to use like GitHub copilot specifically and then my OpenClaw is like I I have a mix of things there. But again to our earlier part of the conversation, at some point there's going to be just a model router, right? Connect this OpenClaw to an an open AI endpoint that's the same pattern, right? It's the responses API or something like that. and you just send your message in and then it just does a little bit of fancy routering and then is it says okay well that that was a request that should be more like cheaper models that that's a request that should be more reasoning model like and then it'll just figure it out and give me the answer. So at some point someone's going to obfiscate the models away from my chatting experience experience and it will just **[00:24:00]** be better. It'll just be a better model. By the way, I know someone who just did just that and I don't know how much we can disclose about it. So, we'll we'll leave it for another conversation, but it has that routing capability understanding if it should be another model or lower-end model or or have fallbacks. If you run out of tokens, it could just fall back to another completely company different model. And it compresses your token usage by like depending on on the prompt and everything by sometimes like 80% 90%. And so without reducing context, which is very interesting because a lot of times we're sending out these messages that due to the nature of the platform that you're using will will tend to be bigger, but not they're not necessarily to they don't have to be as big. So if you have that middleware that's sort of intercepting that making it smaller and then sending it to the right one or to multiple for that matter, there you go. So and and I've seen it action. Looks good. So I I'll see what details we can get from that so that we can talk about it next time. But it's a I love it. I I think that's where it's going. Yeah. First 10 trillion parameter. So I've sent you some links card. This is this is the um so the first one there is the system card for mythos. I I don't again this was recently published I think by anthropic because it's actually on their CDN right now. Um, but it's just kind of a interesting like um change log and they just had a little bit abstract about what it is. If this is one where I would like I'm going to point my agent at this link and say summarize this for me like what's going on here. What what what interesting things should I be talking about based on this paper. I don't really read anything anymore. Honestly, my next Armando, I'm going to ask you the question. You probably already done this. You're probably way better than me in this area. I would expect this to be like you're like, "Oh yeah, we're there. You just need to catch up." Michael, have you heard the concept of second brain? Second brain. Yeah. Yeah. You were going to say you knew it. Yes. Yes. Uh **[00:26:00]** how do you how do you I want to learn from you here. Like do you use it? Is it effective for you? Like what I'm finding the inundation of like GitHub repos and post on X. There's too many and I can't keep it. And I want to just funnel everything to like a I want to have it I want a share button to a single thing and have it index it all and provide summaries of what I found for that throughout the day because I I feel like I'm learning it and losing it faster than I can absorb it. Yeah, for me there's two main players here. There's the Nate B. Jones side of second brain which is called Second Brain um which sort of connects everything together and it's really neat and I think he did a really good job of it and it's it's up there on GitHub and and I really like his videos. He's very down to earth and just explains everything very clearly and is always up to date. What's the name again? Nate B. Jones. So if you search on YouTube for YouTube Nate B Jones AI News and Strategy Daily, that's the name of the channel. It's Nateb Jones. Okay. And uh he has the repo called Second Brain. It's Yeah, there's there's a there's a guide on his website on how to set it up, but then there's also the actual repo somewhere. I don't know where it is, but look for it. That's that's an interesting one. But the second one, which is very interesting, is the one from Karpathy, right? So the wiki version of it that's sort of like a knowledge graph kind of thing via Obsidian. You don't really need obsidian, but that's how you visualize it. I really like that one. I'm that's because it gives it's very simple. It's all markdown files and they're all interrelated or interlin. And so it does work like like a little brain because it's it's a week at the end of the day. But it the way it's set up I feel gives the LLM the possibility to talk to it and gain very quick answers **[00:28:00]** to things across hundreds of documents or thousands of documents depending on on how you set it up with relevance versus a regular vector store which is only semantic search but it's not it's there's no relation search. So I think the best of three worlds is having rag with semantic search or vector search PM2 VM25 search for like free text search for for keywords and things like that and then knowledge graph like if you have the combination I think is great but it's very hard to build the whole thing and multiple platforms and tools and it becomes a mess. So I think this approach of having that that sort of based wiki thing that the agents can talk to is great because you can visualize it visualize in something like Obsidian. You see all the links across things. You see nodes that are bigger or smaller based on how many connections they have across things. So then you can find information based on relevance across the different topics not only if the word or contextual semantic meaning of the phrases. So I like that one. I'm leaning towards using that more especially my main question with that one was how about PDFs like what if I need to ingest other types of files but then the the the other repo I sent you the deep I don't know whatever it's called it was a like an ingestion thing for PDFs it does OCR and everything and that's that looked like an amazing thing so if you put that in your flow then great because now you can have this ingestion flow you drop your files wherever you want like a bin and they just go in there get ingested put into your wiki and then just it's just used for everything and then everything is interrelated and you can use it for your personal notes, your reminders, things you have to do and then also the work things and and so that's you're blowing my mind right now. I think I'm feeling I'm feeling like this is now becoming like another page and content nudge that we're we're going to need to add here shortly, which is like Yeah. No. Yeah. Yeah. Like **[00:30:00]** it's almost like a chat like an only window like hey here's a chat window. you just put in like here's a document I put it in here's a link that I want you to go research and so but the value of this I think is it's like two parts right it's almost like I want so Grock has this feature that I really like right now I don't know if you do you use Grock at all for like like the Grock website okay Grock has maybe other agents are doing this too but I I'm finding I like it the ease of use of this the feature inside Grock seems to be the easiest I just started playing with Microsoft 365 agents a little bit more and it has this idea like there's one thing there's this one agent called the research agent and you can just give it like a URL and it's supposed to go do research and then come back with some answers right so something like that right and then I like grock because it's had this thing called projects and a project is like all the pre-prompting like hey every conversation I have in this project follows the same standard initialized prompt okay right so I have one where I have I write a lot of songs I'm doing a lot of songs on so I have made a project in Groank that's all about how do I want to write what's the structure of the information I want to send to it over and over again to make a song right I want a song that has four verses I want it to be short I want it to be super catchy I want it to have a a lot of hooks in it repeat the phrase and the chorus multiple times like I'm already finding what works for me and how I build like my system and so I'm now looking to like have these call it quote unquote projects set up where I can define some, hey, I'm going to start a conversation, but I need you to have initial context of what I want you to do with this conversation. Yeah. Right. And so that's finding I'm finding that valuable. And now I want like, okay, I don't want to have to tell it every time when I send it a link. I just want to send it like a word like research link. And then it knows Mike wants to research this. Yeah. Provide this, this, and this type structure, **[00:32:00]** right? stick it in either a vector store or a graph database or like you know whatever the thing it needs to do, right? Just kind of just I just want to throw things over the wall at it. Yes. And just have it work, right? And then um you I'm also thinking here too like as I'm looking at these things and reacting to them and reading them and like I'll read through a part of an article and be like, "Okay, that's kind of interesting. I I like this point that that point stood out to me." I'd also like to take that knowledge and say, "Okay, I'm going to give you the research item." Mhm. And then I'm going to say, I found this and this and this interesting. Now write me a post that uses my points of interest. Research my points of interest. Look at those points of interest against what you just researched on the topic and formulate me like a summary, a two or three sentence summary about what that is in my language or how I would say it, right? Something like that. So, taking almost like my spin and kind of like how do I volley the information back and forth with the agent very rapidly and then boom. Okay, great. Now it's saved. It's stuck away in the library. And now it has a couple different reference points, right? The article, the summary, my thoughts, what I thought was interesting. And all of those pieces kind of inform like this second brain of mine which is like I can now just randomly throw articles and my opinions against stuff which helps me formulate new opinions and new ideas moving forward. So that's kind of where I'm like I want to have something that does this. I don't really have a tool that I feel does this well and on the other hand I don't want to pay for it. I don't want to pay for like a notion or integrate another. I already have this is one of my big burns right now is I have a lot of tools in Microsoft stack. I'm a Microsoft guy. You're more of a any kind of stack guy which is great. Whatever it's it whatever works for you. But again, I like the idea of like I've already spent the money to pay for Loop. I've already spent the money to pay for agents inside Microsoft. I mean that's the stack that I'm building in. So, why can't I **[00:34:00]** leverage those things to build what I want? And what I'm finding now is it it has it does some of what I want, but it doesn't do everything of what I want. And Microsoft, I don't think, is going to build that. They're going to give you more of like the Lego bricks of what you want. So, going to like a notion or something that's very deeply integrated. I've just been playing with Manis AI, which is now owned by Facebook or Meta, I guess would be appropriate, not Facebook, it's Meta owns them. Very interesting. good product. Um, but it I I'm I'm constantly challenged with the fact that it feels like I'm sticking myself in yet another box. Yeah. I just I just it's an interesting tool. It does good stuff, but now I'm stuck in another box. And I don't want to I I don't want the whole advantage of this AI thing is I want it to be all sharable and distributable across everything. Yeah. That that's why I like the weekly approach because it's all just standard files and Yeah. blinking like it's it's pretty old school when you look at it and very simple. I think our pathy is very like down to the basics and figuring out how you can make it the smallest possible with the functionality that you need. Yeah. So I think it's a very interesting approach which you can put anywhere just use your blob storage and then just like link files and we had a a really big success recently with the app that we were building which which I you know now we can store videos and I can now store text and I can now store transcripts on things. So like already fabric is really becoming like the back end of my apps that I'm really enjoying. SQL databases for storing structured information. There's other databases coming here in fabric very soon like Postgress and other data sources. So I I want to continue to build on this to see like okay what do these different text stacks look like? I have Cosmos DB inside fabric. Excellent. So if Cosmos DB is there does that how efficient is that able to run? I can actually do storage. I mean that's like you know MongoDB level stuff like ju **[00:36:00]** it's just little file storage and and a query engine on top of it potentially would give me a lot of value there as well. So I'm I'm really evaluating again like really sitting back and saying like what as a business owner when I step back and say how do I leverage all this stuff for my advantage? I don't like the shared subscription models across everything. I need everything to be integrated together and at the end of the day I want all the data to live in a pile at least where I can touch it. So that's that's yours. That's mine. Yeah. That's like my that's the IP of about what I want to do, right? So that's kind of like the secret sauce. And so when we go into this whole like everything is hyperpersonalized, right? Hyperpersonalization is the way we do business and works. If your app can't hyperpersonalize what you want to build, I don't think it's a good app anymore, right? That's the new tech stack that we're going to need to build in. Yeah. It's so easy to make things now, but it's also hard because you have to like figure out what you want to do and how you want to do it more than Yep. But simple. I think it's the the the KISS principle. Yep. Is more important than ever. The other day I was talking with a client we have in India. We've worked together for for years now and we've always had the discussion across we're talking about other stuff but it always comes up that we've changed task management tooling on each side. It's like, oh, now I use notion and now I use this other thing and now I use this other like everywhere, right? It was Jira now it's Jira and Asana and like nothing works now like I use I use a pencil and paper like that's the one that works and like stickies and then and text edits. Well, that always happens and I' I've had this discussion with other people and and yourself and so he comes he says super like basic thinking. He says you know what I'm doing now I figured it out. I'm just using Claude. So he's just telling Claude this is what I need to do and he just like okay and then Claude just keeps a task list for **[00:38:00]** him. So now whenever he needs to make a task he just tells Claude and he adds to a list and then whenever he completes something he just tells Claude and it completes it and now he has a log of all the tasks that have been completed and in all different aspects of things. And I was like, that's really smart. Super basic. Like you're just telling your agent, and that could be any agent, doesn't really matter as long as it has the context, and you can switch agents if you wanted to because I don't know how he's doing it, but I would do it like with a repo similar to what we've done with GitHub. Yeah. Such that it's just it's all there. It's trackable. Changes are trackable. And you can switch agents. Doesn't really matter. It'll just keep track of all of these and check the ones that that need to like to your point that we've discussed many times like GitHub projects. Mhm. that could be your social truth of like all your tasks across many things, but then just talk to it via agents. And so you no longer need tooling that works like task list. It's just as long as you have somewhere you can store the doc the text or the whatever. Yeah. Where where it's an MD file or a JSON file or on on a GitHub repository. It doesn't have to be GitHub even. It could just be a files on a Git repository. That's that's all you need and then you can interact with it. But it's just a smart way to do it and it can even take times and all like all the time management stuff like you just say like okay now I'm I'm starting work on this. That's it. It just needs to store the time stamp and what you're doing and then later on you just say hey I finished and then and then if you ever want to visualize it or get a report you just say hey how much time did I spend on XYZ project? Just grab everything and give you the response. What did I do? No app, no nothing. It's just I want to extrapolate your idea because I love this idea. Why do we build UI for anything anymore? This reminds me of an idea that someone had a while ago, which was um someone was talking about like their treadmill and they wanted to like, hey, I'm I'm going to be healthier. And they basically said, look, I want to just start, this was very early in the days. **[00:40:00]** This was like in January time frame where I was just starting to explore this. I'm like, oh yeah, this is this is reformulating how I think about things. And so the the premise was I have an app, a treadmill that's connected to the internet, has some sort of information in it. It's probably hard to get it out. It wasn't the easiest thing to do, but he's like, I just want to track my my usage. I want to make like a goal tracky something, whatever that thing is. And he basically said, every company is working too hard to build the UI that is pretty and easy to use and meets all the customer needs. That's the challenge. As soon as he started moving away from the company that like and soon as he were moving away from what the company built UIwise and started accessing just the data he could have his own goals his own preference his own style his own like and that's where this is where part of it started formulating he's like I built this entire health tracking app and I could do whatever I wanted to do with it and it did everything I needed to and He's while the agent is involved in the story, the agent wasn't the end all beall, right? The agent was kind of more of like the reasoning thing that was able to, you know, kind of pull all the pieces together and make it easy to use. And so to your point, Armando, like the agents aren't going to be the health tracking app, but the agents will help you build the pieces of efficient hardware, software stuff that let you get to the app that you want. And then to your point here as well, like again, this is like he's using your partner here is using Claude code as like a second brain, right? Hey, I'm just going to throw things at you. It's going to act like an employee at my an assistant at my company. It's going to say, "Oh, yeah, I know how to do this." And then when I say, you know, I need to do this today. It understands the context. it reasons about that and then and then uses something under the hood that's like the database, the storage, the file system, the GitHub, the things that **[00:42:00]** don't cause the the agent to like spin and churn and be very expensive, right? So, I'm looking at this going like we need to stop or the the shift is starting to happen where we stop building UI and software and all we need to build is APIs and backend and storage of data. Like that's that's how we should be interacting. So that way our agents understand what it is and then we can describe to the agent here's the data. Here's the system behind the scene. I need you to build this and then it goes out and builds the front end that I need that works for me. And this is I'm looking at this across my entire business, right? This is where I'm like burning down every piece of software. Anywhere data is stored that I can't get my hands on easily, I'm like, let's burn that down. Let's build my own. I can stick it in fabric. I have full control of all the data storage there. Boom. Done. And so now I'm doing more and more of this and it's accelerating for me. I'm looking at more opportunities in that space. And so I love this idea and now I I think about this and react to it and say, "Well, that's great. I would love the same thing. Well, how do we build a skill around this? How do we build something that's like transportable across our whole team? How do I get visibility to again looking at this? This is my task manager of things I need to do today, things that are important. I want the agent to help me reason about what should I spend my time on. Yeah. Here's what do I have to do today? Yes. What's my backlog look like? That's right. Because we don't we're very bad at at prioritizing. Yeah. I just want to give me and it can give me a first pass. I mean, it's not going to be perfect every time, but I just want a first pass at like what's good and then I go from there. Yes. Yeah. Yeah. Because it's all there. That's the important thing that the data is there but that now access to the data becomes paramount. It's just and and and access to data not only for me but access to the data for the agent like that's to me that's the next hook. That's the next thing we've got to figure out how to do. And then and a lot **[00:44:00]** of the like the only reason you're talking about markdown talking about like these databases. The only reason you're describing these things is because that's ways you get access to the information. and how do we do the vector graph database store whatever second brain you want to call it. I'm now more focused than ever of being very mindful of where does the information land. How do I get myself to have access to it? How do I get my agent to have access to it? That's that's more of what I'm shifting to. And I'm I'm also bringing this up too, Armando, because uh I wanted to talk to you about this as well. I wanted to give you an insight that my bot did when I asked it to add agents to content nudge. Okay. So, I'm trying to find the organization real quick here so I can go get the repo and clone it because I had it write some instructions and I wanted to read you some of the instructions that it directly made for me. So, hang on one second. Let me pull this down here. While you're doing that, it's where before I would see a software as something very abstract and like someone told you you have to look into a software that you didn't build. Nope. It's like I don't I don't I don't want to tackle that. Or you have to do a database migration kind of thing. All that is very tricky and slippery, right? Mhm. But over the past few weeks, it's it's become more of a me I mean the shell it's just a shell. Yeah. I mean sh not like a terminal but more like a shell. It's less intimidating. It's a wrapper. It's like it's just a wrapper around the data. And if you understand the data and you look at it and you can have the agents help you understand the data. It's like, yeah, we just because we're rebuilding a a client of ours has this app that he wanted to rebuild for years, but because he was using one of those no code platforms and and it wasn't scalable. Well, so so did you find your I did find it. Okay, tell me. All right. So, so this is what it so **[00:46:00]** context of what I was discussing with the agent was like, hey, I have an app. I have a backlog and I don't want to go to the UI to build stuff. I want you the agent like this is like I want you to be able to access it with like a secure way. So not authenticating as me the user. I want I want a dedicated agent path to go authenticate. So we're talking we're kind of reasoning through like what would this look like? How would we build it? Right? So it's it's doing what I think it should be doing, right? Hey, we want to be have you know an an off token thing, right? We want to have a separate endpoint where you can send the token in and then it checks the token and the token identifies like what team you belong to and all these other like good security things. It also talks a lot about like scoping. Hey, you we should actually add scopes, right? There's these conbon cards on the page. So we should be giving it a scope of like read only. So hey this agent has a token and this token is only for readon cards. This is for modify. This is for create. And you basically give it the right of scopes you want. Again, right, good best practice for author minimum authorization parameters. I'm liking all this language here. And the one thing that kind of struck me as interesting as it was re and this is I think GPT5 4 5.4 is what I was using here. And it says, let's see if I I'm I'm looking through the chat thread here. Um it it describes the goal to me here a little bit. And then um hold on, you keep going back here a bit further on. Okay. Was doing some data modeling suggestions. Okay. So it talks about the best architecture. Provide a stable task model. Add simple actions via APIs. Clear agent safe metadata. Add an audit trail. So whatever the agent does, you can see it. I'm like, okay. Yeah, I agree. This is where it said it got better for me. It says in planning this, you should think about like what you want it to do. A column has an ID, a title, and a position. The cards on the different columns, right? So, done, to-do, Tuesday, Wednesday, whatever the **[00:48:00]** whatever's in the column names, right? It has to be able to describe the columns, and it has to understand what's in the card. And it's like, you know, ID, title, description. That's kind of what I was thinking, right? But then it said, I suggest to you to add other custom fields to the cards. So, the agent has additional context as to like why you're doing what you're doing. And I thought this was very insightful because again I was going back to like my second brain idea which is like wow this is a really interesting article. I found it on X. I want to give the link to my agent and say add this to the backlog for the explicit measures podcast or whatever. And it said you should think about adding like an impact field. How much effort this would be? What would be the category of this field? Do you have a target platform or content type or a business goal? Business goal, right? And it said if you really want to build a strong backlog, we would add fields like what is the audience? What problem is this solving? What is the format you want to use? Is there any additional source material? What confidence would you have in this kind of content? And then was is there a next action? And I found this to be very insightful because I was really focusing on like I have the problem of I need I'm getting I'm collecting all these ideas across the internet and I'm trying to build a show around basically idea based right so the show's ideas we talk we sit we we explain ideas together on the show we unpack all these ideas and it adds value for people who are understanding and thinking through these similar scenarios problems or solutions but these additional data fields we're like wow this actually really does add additional context and so Instead of me just saying topic, description, summary of URL, it's giving me a lot more around helping me unpack like which backlog items are actually solving problems. What problem are we trying to address? And that's much more impactful. I thought I'm like, wow, this was really neat to kind of go through this exercise. So while I was thinking of just solving like my simple second brain problem, it was actually **[00:50:00]** giving me like better context around what I should really be thinking about which is impactful backlog management and subscription. So that's like wow this is really neat. So I'm still unpacking it. I'm still working on the feature but it gave me a lot more to think about and it gave me all the standard stuff like hey you need to have a get post patch right? There's all these different commands for the APIs like great cool. So, I want to just get the basics in place first, but then now it's adding a bunch of these other details. And I'm already seeing in this initial phase of just thinking through it with an agent, right? I'm already going, dang, my initial feature of conbons and backlogs was actually insufficient. Yeah. To what really needs to be thought about. And this is this is what I should be thinking about. And so, how okay, now let's think about this more. So, I'm going to implement this with your approval because we're working this together. That sounds It sounds amazing. Well, I'm I'm going to implement it, but then it'll it will now immediately funnel into Cool. Now that we've done this, now everything else good. Now, yeah, I got to like think about everything else I'm doing now because now I'm like I'm not going to want to type anything in manually at all. I'm only going to want to add items in the backlog by using my agent. Now, a new problem I'm going to have. It can abstract all that metadata and put it in in real time versus you typing each each. Yeah. I don't and out of the think of it all. Like now it just literally shifted my entire mindset away from I'm not just dropping down a title and description. I actually have to like really think through and reason through this. And if I did drop down a title description, how do I take that minimal amount of information and then let the agent say, "Okay, you need to go unpack all the minimally viable backlog items, enrich them with this extra information, and then let me evaluate and then prioritize them and pull them into a schedule." So like I was like, "Wow, this is so neat." And also, I'll pause here just a minute, Armando. Like you have the Web Talk Show. How are you planning that one? Are you using content nudge for your planning yet or are you doing something separately at this point? By the way, I have to she's she's **[00:52:00]** calling again. But yeah, no, you know, it's funny because the Web Talk Show is all improv the whole thing. No. A wow.