The Critical Advantage With AI

This is an automatic transcript of the recording above. It is published in full and unedited, apart from correcting names the recogniser reliably mishears. It will contain mistakes.

00:01[music] >> Hi, my name is Tejas. That's pronounced like contagious. Okay? Don't worry, I'm not. Especially on the front row. I'll spit on you a lot, but I'm not contagious. I You pronounce it like pages, you know, or like advantage. It Look, it's a weird name, man. And so I I want to do my best to help you uh say this, not because I expect you to say it. Anyway, um I work as uh as an AI engineer at IBM. And we do a lot of cool things uh with I work in IBM research uh and we do we train models and do harnesses and all kinds of things. Um but maybe more relevant for this talk is uh a a different fact I I studied human-computer interaction for some time uh at Stanford University. And I mean, this is pretty relevant because the way we interact with computers, human-computer interaction, is is profoundly changing and it's so cool to observe. I mean, all the talks today uh alluded to that. And so, it's really really wonderful. Today, we're going to finish We're going to bookend day one of the conference talking about the critical advantage with agentic engineering. The critical And the the reason for this is because there's we live in a day where like everybody now has a hammer. You know what I mean?

01:31But we're not all building awesome houses. Makes sense? Slop forks. You know what I mean? Um we all have cameras in our phones, but you still hire a wedding photographer, or at least you maybe should, right? And so, just because the tools are there, uh doesn't mean the quality is going to follow. There's a very clear line between between slop and craft. Yes? And so, what is that line? What is the difference? How do we give ourselves an advantage when we build things with agentic help? That's what we're going to talk about today. Now, what is the critical advantage? You might be wondering. Oh, any guesses? No? What's it What? It's Pepsi. I'm joking. I'm joking. It's not It's not Pepsi. It's But it's it's more this word on the Pepsi can.

02:19What is that That says smakku. I learned a new word in Polish. Smakku. It means It means taste. Anyway, smakku, that's the that's the critical advantage. Or in other words, taste. And this is smakku, man. What an awesome word, dude. I'm just going to say that. Nice smakku. In German, it's Geschmack. Anyway, um I This this talk is about taste and it's taste through the lens of science. Together, we'll uncover science-backed, evidence-backed protocols where we can increase, enhance, improve our own taste. And the hope is we go to work from Monday and we just we work with a greater sense of taste and create more craft as opposed to slop, okay? And I think Hey, listen, I think this is very timely and very important because the tools are only getting better. Um but they may be getting better faster than our taste. And ideally, we catch up and we create truly wonderful things. In fact, it's not just me talking about this.

03:13Mike, anyone remember the keynote from a few hours ago? Yeah, the CEO of Kal Stack, Mike. He said it himself. He said, "Hey, what about 2028?" We're going to talk about what work deserves to run. We're going to talk about who owns the result. And what does he say? How do we develop Smakku. Anyway, so that's um That's This is going to be a topic. In fact, it already is. And so that's what we're going to spend the next 20-ish minutes talking about. Taste or smakku. Um how do we define it? And look, I don't want to stand up here today and give you like an opinion because opinions are cool, but you know what's cooler than opinions? Science-backed fact as best we can uncover with research, peer review, and experiments. Okay? And so, everything today here is going to be cited. There's going to be resources and references.

04:00You can go please go read the papers yourself because that's how we get better. Like it's just some loud guy with a microphone saying his opinion isn't going to help anyone. So, what how do we define taste for the purposes of our discussion and reasoning? I've I've skimmed I've read a lot of papers. I I studied this stuff myself. And the best working definition we can do for taste in the day of agentic engineering is this. We can define taste as a trainable skill, and I don't mean an agent skill, Jesus. I mean like a human. A trainable skill where we are able to tacitly judge quality based on external standards. This is the working definition of taste today. It's not my opinion. This comes to us from the development of aesthetic experience and judgment all the way back from the year I was born, from 1993. And it it's like the working definition of taste across industry. Uh from as old as we can go, right? It's a trainable skill where we're able to tacitly judge quality based on external standards. You may notice some words on here are bold.

05:00And I think it's worth spending a minute or two just really pulling on this thread because we want to get value so we can take practical things from this talk and go apply it on Monday. So, what does trainable skill mean? It means we can all learn it. And we'll we'll spend some time talking about how do we develop this. That's what I want to give you today. Um where we're able to tacitly judge quality. Who here knows what tacitly means, right? Yeah, no one. It's except him. Like it's it's a country where English is not the first language. It's fine. That's why I'm here. I'm a native speaker. Tacit means fast. It means like instinct. It means like, "Oh, I feel it in my gut." Like I took one look at this and I knew it was slop or or the other thing, right? We just know by instinct. We don't have time to think about it. The subconscious brain makes the judgment. And that's why we have this third qualifier to tacitly judge quality based on what? Based on external standards. Because when you make tacit judgments based on internal standards or internal bias, what's that called?

05:56It's called racism. It's called It's called xenophobia. It's called second We don't do that. No, we don't do tacit judgments based on internal bias. We do tacit judgments based on external standards, and this is the working definition of taste. Uh and so it's very, very important to to include this external standards piece. One tenant of taste is that it's trainable, and this is huge because it presents hope for us. How can we train this? How do we improve our taste? Uh science gives us three special ingredients that help us all improve our taste. Number one, uh immersion. So we we literally like immerse ourself in quality every single day, every single moment that we can. Let's Let's use a practical example here um of immersion, right? Say you want to build like the best computer mouse, right? How do you do that? Well, you just immerse yourself, and immersion is tied to volume. How many mice do you look at? How many mouses, computer mouses, do you experience? If you go play with like 500, 1,000 mice, you start to get immersed. Another way of thinking about this comes from my favorite book of all time, uh the Bible.

07:00Like [laughter] we have immersion defined as this through like the living word of God. It says, "Whatever is true, whatever's honorable, whatever is just, pure, lovely, commendable, if there's any excellence, if there's anything worthy of praise, think about these things. Immerse yourself in these things, right? So immersion, what does that look like? Well, if we're designing, if we're creating a tasteful computer mouse, right? We're going to look at all of these, and we're going to be like, "Whoa, that's a lot of mice." And now, here's the thing. Immersion is not enough. Because you're going to look at this, and you're going to be like, "Yeah, I bet that's kind of cool. Oh, that's weird. That one has balls."

07:38Um and you know, you're just like, "What's going on here?" Uh but you don't really know how it feels to use. You don't know how it feels in the hand, right? You don't know how it like when you move it across the surface, how does the cursor move? Is it fast? Is it slow? Does it have a sharp edge that hits my hand just the wrong way? Like you don't know that. So, immersion isn't enough. Like you could look at 500 mice. You could even play with them for like 30 seconds. It's not going to do much. You need the second ingredient. The second ingredient is you need valid cues. Valid cues, they tie the performance of a thing to the surface level impression of a thing. You need both. In fact, if you watch the best YouTube reviews, right?

08:17They go in depth. They're not just, "Hey, this phone looks great." They're like, "Oh, the refresh rate is 240 hertz. I love it, right?" It's It's It's more detail. It's a higher fidelity image. And so, immersion plus valid cues helps you really develop taste significantly. Um They They They need each other. You can't develop taste with one or the other. You need both. For example, if you immerse yourself in computer mice, again, you use 500, but you never really do anything meaningful with the cursor, then you're not going to understand how does it move the cursor. It's the whole point. Similarly, you could spend your entire life studying the human hand. You could be like, "Whoa." Without drugs, too. You could be like, You could be like, "Whoa, man." And And you could understand the contours, the edges, how does things feel? Are they Are they smooth? Are they rough? You could also study electrical engineering, optical sensors, right? Lights that move the mouse. You could be like, "Okay, so So, the the laser bounces off." You could study the technical of it. But if you go your entire life only using a trackpad, not a mouse, you're missing out. You need immersion, and you need valid cues.

09:22But that's not enough. There's a third ingredient. So, let's go back to our mice, right? You have Now, you've immersed yourself. Now, you're thinking, "How does it feel in the hand?" And some of you may be seeing like some of those mice with like the little side thing, you know, maybe that's helpful, the vertical mouse, right? Um you're thinking more how does it actually feel, not just how does it look. But there's a third ingredient. And that third ingredient is um corrective feedback. You typically need to work with someone more senior, a mentor, someone who can disciple you in the work. Um and so you've got immersion, you've got valid cues, and you need someone to say, "You know, I think that's a pretty good mouse." And you need an expert or someone further down the road to come and say, "Yes, I see how you would get there, but you're wrong." And so you adjust your taste.

10:06Corrective feedback is exceedingly important. This is not my opinion. This comes from Dr. Kahneman and Klein from a 2009 paper titled Conditions for Intuitive Experience, a Failure to Disagree, uh in the Journal of American Psychology. It's a very top-tier journal, okay? So, these three ingredients, immersion, valid cues, corrective feedback. Let's apply them to the mouse. So, immersion, we look at them, they're cool. Valid cues, we think about which one feels the best in the hand, which one's the best. And you know, we may we may say, "Yeah, I think the Apple one's the best. I mean, it's shiny, it's minimalist, yeah?" But you need corrective feedback. So, now you go to your mentor and you say, "You know, I think this is the best one." And you get the corrective feedback.

10:46You know. And your mentor is like, "Yeah, but look at this." And then you go, "You know what? I'm going to adjust my taste." And so then you design a mouse without this nons. Anyway, so that's those are the three ingredients. Again, immersion, valid cues, corrective feedback. There's an incredible book by Pixar titled Creativity Inc. It's incredible. It's how they made movies like Finding Nemo, The Incredibles, really really remarkable projects. And they outline exactly a process that it that implements that paper, where they would take ugly babies. These are their unfinished movies, and they would do a retreat. So, it's a room full of filmmakers, and they would show people the movie. And then other filmmakers would give them corrective feedback.

11:28It's an environment full of deep immersion, right? And they all have valid cues. They all looking and experiencing the movies together. So, those three ingredients, if you take them to work tomorrow, which is Saturday, and and if you implement them, I can guarantee you will start to see significant increments in your taste. Okay? So, we're talking about taste. We have this definition, trainable skill, where we're able to tacitly, instinctively, subconsciously judge quality based on external standards. I'm tired of talking. Uh let's let's just let's make it practical and explore with some code. Yes, we read code still. Um let's explore with some code how we can feel together something. So, uh what I'm going to do is browse the internet. It's an experience that I profoundly dislike. Who here likes browsing the internet? I hate I hate it.

12:21Because when you visit a URL in your browser, you're leaving your comfort zone, your house, and you're going to someone else's house. And usually, decisions on these houses are made by people not thinking about you, but thinking about shareholders and impressions. They're not thinking about you, they're thinking about the EU. Uh cookie banners. It's just not Let let Look, I'm just going to search I want to buy shoes, brother. I want to Here, uh shoes. That's what I'm searching for. Um and uh the first thing I see is this. Now, I absolutely hate this. Um and I don't even speak the language. My device is in English. You know what I mean? So, this is kind of irritating, but whatever. Let's uh I don't know. Let's say okay or or or odrzucić um and I'll click on like the Foot Locker website. Uh and then I'm probably going to have some nonsense Okay, this is actually quite nice. Nice. Um but the web like let's go to nike.com or something, you know? And it's just always Look at that, man. That's wonderful. Um the web feels like this, you know? You go here and it's like, "Please click here." Oh, I click here. Okay. And now some things counting, and I really know what's going on. Choose a password.

13:30Okay, oh, it's not a placeholder. So, um I hate this so and I do not accept. Okay, so I guess I I do accept. My email tejas@ I can't tab, of course, ibm.org.com. Of course, of course .com is like at the bottom, right? Um next is here. I can't click it. Password needs at least 10 characters, one capital letter, one number. Password is not unsafe. Anyway, so this is what the internet feels like, and I hate it. Um So, when we use an agentic application, right? Agent runtimes are awesome cuz we don't have to leave our house. I'm in my agent. I'm in chat GPT. I'm in Claude, and I'm having a conversation. It's so much better. Sometimes I go to conference websites, I won't name any, and I try to find information, but the UI is just like, "Whoa, there's a carousel, and now there's a cookie banner, and I don't like it." What's a better experience? Let's Oh my god, there's more. Um Time is ticking upwards, too. I don't Anyway, um let's go to chat.openai.com, right? And start a new conversation.

14:37This is This is the more, you know, this is the coding expe- the agent experience. And I'll be like, "Who's speaking at AgentConf in Poland, right?" And this is nicer. I can discover the speaking schedule of AgentConf. Um but I want you to tacitly judge something here. I want you to see the taste. So, it's actually doing this for me. That's awesome. It's happening in Warsaw. This is actually so much better in a way cuz there's no friction, there's no UI issues, there's no like carousel or whatever. And I see my name there, and it's all good. Um but can we be real for a moment? Let let let let's be real for a moment and make a tacit judgment of quality, right? And I tell you what, man, I feel like it's nice, but I feel like we lost something.

15:19Yeah, I feel like we lost the branding of AgentConf. We lost the taste of agent conf. There's no call stack here. There's just soulless text, right? So, how can we deliver taste with agents these days? Who here has built an MCP server? Awesome. This is great. So, as part of the MCP specification is MCP apps. Who here has built MCP apps in Okay, one person. Crazy. You're going to learn something today. Um I'll show you a little extension to MCP that will bring taste to all your agentic experiences. And so, like let's go to the editor here. This is If you've created an MCP server, you know this already. Create MCP server. I create a bunch of resources and tools. I have a tool um called get schedule and some metadata. Title, the description, get the agent conf schedule. An input schema, that's a Zod schema.

16:12But the important thing here is the meta. And I say, "Hey, this tool has UI." And the UI lives at this resource URI. Now, this URI, it's not a product It's not a web page or anything. It's just a magic string that connects a tool to a resource. And the resource is just a document. It's text. In this case, it's an HTML document. So, then I'll go to this schedule URI and see which resource uses it. So, I have server.registerResource. And this one, sure, uses schedule URI. So, you can find the resource here. The MIME type, check this out. This is important for the spec. It's text/html with profile MCP app. That says I'm not returning text, I'm returning UI, okay? And my async function, it returns a envelope with text. MIME type is text/html. And it's just text. Okay, what is schedule widget HTML? Well, it's just Node.js read a file, that's an HTML file that I wrote. What's the HTML file?

17:07Well, it's just a big HTML file with some UI. I wrote this. div ID tabs, div I just wrote this. And it's some script here where there's globals on the window object. So, you might see window.addEventListener open AI set globals. Open AI or whatever MCP client you're using, whatever tool call results you have, in this case the list of speakers, the tool call results are passed into my HTML file via a global window object, window.openai.tool_results. And then I have access to that and I can interpolate it in my template wherever I want. Is this tracking so far? Yeah. And so, I just have a tool for my agent and I say, "Hey, this tool is attached to this resource and this resource is HTML." And then it's up to the MCP client, Claude, Cursor, Windswept, whatever you want, ChatGPT, it's up to the client to import my MCP server, understand, "Okay, I get some UI." And then it's up to the client to render the UI. So, I'd like to show you this actually what it looks like in ChatGPT. So, what I'm going to do is I'll npm run my little MCP server. This is running on localhost, but it's using Cloudflare to tunnel this to the internet like ngrok.

18:13And what I'm going to get back is this cool tunnel from Cloudflare. Thank you, Cloudflare. And I'll go add this to ChatGPT. I'll go to chat.openai.com/apps and I will make a new MCP app, right? ChatGPT app. I'll call it agent conf. Yeah, I'll just call it agent conf. The URL is my URL here. There is no authentication, although you support OAuth. And I know it's dangerous, but I know what I'm doing, whatever, right? Um and so now Oh, no way, too many requests is wild. Um I'm getting rate limited. Uh let me Let's Let's start a different tunnel, I guess. Um The The purpose of live demos is that they must break. Uh that's That's the rule. But we'll we'll make it. I I have no doubt. Worst case, I also have a video, so we're fine.

19:00Let's try this one. Um unless ChatGPT itself is being rate limited uh and then we're we're really cooked. Anyway, I think it's maybe working. So, we should see a sign that says add >> [laughter] >> I will do do I hotspot myself? That's on LTE. Okay, are we getting logs? We're not getting Let's try again. You know, every single time I do this demo, OpenAI has down Last time literally it was down for users just in Europe. Um which is You know how unlucky you have to be for that to happen. Okay, it worked. Nice. Add agent conf. Thank you. >> [applause] >> Okay. So, we add agent conf to ChatGPT. Should we do it? Yeah? I'll connect and I hope to God Okay, so now it works. So, now I'll go have that same conversation, right? I'll be like chat GP chat and we'll start new conversation.

19:54Who is speaking at Agent Conf? And now I can at mention my tool, right? And this is going to over Cloudflare um speak to my locally running thing and check this out. Like I get beautiful taste like from Call Stack. It's the branding. It's the speakers. You've got Kent and it look the hover state works. You got me, you got Nader. I didn't even know Nader's going to be here. What is Nader talking about? What is Nader Nader talking about? And I'll at mention it again just for fun. I don't really have to. Um what is Nader talking about? And now it gives me rich detail. It knows who I'm Look at this. It's incredible. Nader Dabit, he's talking at Thursday 9:50 to 10:15 right here. And the best part is you see the hover it works. Finally, I'll do show me the full schedule, right? And now I want to show you this. It's so cool. I don't have to go to some website um as nice as websites are. But check this out. It's It's got the navigation in here um and it's interactive. Watch me click on this. Um and it like totally up I could embed an entire map in here if I wanted to. I could I could embed a credit card form and have you buy something in here securely with MCP apps, right? And so, this removes that soulless wall of text and introduces the taste of actually the company, the vendor. So, in summary, um I hate the web. No, I'm joking. In summary, um we live in a day where anyone can make

21:18anything. It's democratized. However, the key differentiator, the key the key critical advantage is what we do with it and the quality of the things we build. In practice at IBM, we actually make our own coding agent. Um it's called Bob and we try to make it as tasteful as we can. You're welcome to look it up. Um and this is one of the ways This is our coding agent, Bob the Builder. And like when you wave at him, he waves back and I think that's really cool. So, um I hope you leave here today after day one, you go to work, you improve your taste with those three ingredients, and we truly together we create wonderful software. Thank you so much for your time and attention today. >> [applause]

• 2026

  The New UX

  CityJS London 2026
• 2026

  Frontend after AI: The New UX

  Future Frontend 2026
• 2026

  Accessibility panel

  Future Frontend 2026 (Accessibility Panel)
• 2026

  Harnesses in AI: A Deep Dive

  AI Engineer Europe 2026
• 2026

  AI Yesterday, Today, and Tomorrow: Extending AI Systems with Model Context Protocol

  How to Web 2025
• 2026

  How to Thrive as a Professional with AI

  React fwdays 2025