Agents Will Kill Your Ul by 2026--Unless You Build This Instead

This week's executive briefing is all about the future of intelligent pixels. We're moving from product as an interface bundle to product as a durable substrate with pixels as throwaway. And I want to dig into what that means. And yes, the catalyst for this is Nano Banana Pro and the transformation it has brought to the way we think about images. But I look at Nano Banana Pro as really the tip of the spear. I'm not interested in whether you think this particular model is I'm interested in

this as a tipping point and we'll see more models that are even better than this in the future. So what does that mean for our software strategies? So I want to break this into a few moves and we're going to go through them one by one and by the end I think you're going to see where we're ending up from a software build perspective, from a software buy perspective, even from a talent allocation perspective. So let's jump into it. Number one, coherent interfaces were an economic hack, not

necessarily a law of nature. For 40 years, we treated user interfaces as scarce because they were expensive to design, they were expensive to build, they were expensive to QA, to localize, to document, to train on. I still remember the days of onprem in the basement Oracle servers, right? Like that's the world we live in where software and hardware were both very expensive and that meant when you got an interface it had to be shared and serve thousands and millions of users and use

cases. I used Oracle eyes store. Oracle eyes store sorry to anyone out there who's from Oracle is a terrible terrible terrible interface. It is absolutely awful. I have deleted half a store because of Oracle is store's terrible interface, but it had to be shared by thousands and millions of users and my preferences didn't matter. Interfaces had to be durable. You had to amortize the design and development cost for years. So, Oracle Eyes Store stayed the same for a long long time because no one

cases. I used Oracle eyes store. Oracle eyes store sorry to anyone out there who's from Oracle is a terrible terrible terrible interface. It is absolutely awful. I have deleted half a store because of Oracle is store's terrible interface, but it had to be shared by thousands and millions of users and my preferences didn't matter. Interfaces had to be durable. You had to amortize the design and development cost for years. So, Oracle Eyes Store stayed the same for a long long time because no one

wanted to change it and Larry could keep making money. So, that meant we optimized our organizational structures around coherent and longlived interfaces. So we would have opinionated interaction design. We would have navigation. We would have page layouts that had very clear mental models embedded. We had training. We had certifications. Has anyone ever been Salesforce certified? Has anyone been Workday certified? Anyone certified in how to use Jira? This is what I mean.

This also meant there was huge change management overhead for any major UI shift. that made sense when every pixel essentially had to be handtoled. That is no longer true. And we need to recognize that this moment, this 2, 3 week period, this is the tipping point. We've seen signs of it before, but this is the moment it all changed. Generative and agentic waves are making pixels cheap and contextual, and we just hit that tipping point in the last couple weeks. You have three overlapping shifts

happening at once and reinforcing each other to drive this tip. First, generative user interfaces are models that can spit out full screens from text or context. You have UISard, Vzero, Galileo. They already generate multis-creen mock-ups from prompts. Neielson has talked about this as the dawn of cheap disposable UI. I don't care how far down the hype train you go. I think the key is to recognize that ephemeral user interfaces are popping up everywhere because they kind of in fact

there's this entire startup called Wabby that just allows you to make generative interfaces and software for yourself now as a personal consumer. You can have generative interfaces and comet generative interfaces in other smart browsers. So you have generative UI becoming a thing. Number two, ephemeral and generative UI concepts are exploding. And so there's a growing conversation around what hyper contextual applications or panels might look like and how we might create and

destroy them and keep the application state the same. That is different from the technology itself. So generative UI is about the technology and user interface. The idea of UI concepts really the design language is growing and we need a new design language for this this change we're all going through. The third trend is agentic software that drives other software. This is actually funny enough where Nano Banana Pro I think rightly comes in. Google smartly placed their image

generator right out of the gate on an API so that agents can call it and come back with images. People who are enterprising are already using this for interface design from Nano Banana Pro. I am not talking about theory. I'm talking about what I actually see on X on Reddit other places with screenshots with videos. People are using the API call to pass a string of data in a structured prompt query to NanoBanana Pro and retrieve a chart or a graph that they can then display as the past week's

sales, the past days customers, whatever it is that they need for internal metrics, they can just automatically query and get a nice chart back from Nano Banana Pro. That is generative interface driven by Aentic software. Fundamentally, the interface is something that is starting to morph based on user context and it isn't staying fixed anymore. So, if you put that together with the idea of throwaway pixels, fundamentally, you have software that's changing in value. Software is

becoming generated on demand from intent and context. It's becoming private to the user in the moment for that particular ask. It's becoming discarded when that moment passes. One of the most instructive descriptions of vibecoded apps has been a recognition from folks who've done this over 20 30 projects that they are finding that these apps are valuable in the moment and some of them they may use again but some of them they created just for a single use and that was worth it to them. So Nano

Banana Pro is basically a futureleaning version of this. A model that understands UI structures, sketches, diagrams, and it flows well enough that UI just becomes one more output modality like text or code. That's your disposable pixels. Before I get too far down the road, I don't want you to walk away at this point and think Nate thinks software is dead or Nate thinks that software won't exist anymore. That's not true. I think the opposite. But I do want to actually talk through what this

means because I think software is going to profoundly change. So let's look at what disposable pixels actually look like in practice. I want to call out three layers. Layer number one is the system of record or the system of decisioning. So in this sense the things that B2B SAS was good at, they don't die. They just move downward in the stack. So data models, workflows, permissions audits compliance things that we paid for when we purchased the software, things that we pitched when we

wanted to be entrepreneurs and make money off of building stuff. It was this hard stuff, right? That's moving down the stack. Domain logic, forecasting, pricing engines, how you handle uh interconnects, APIs, and web hooks. This layer, frankly, is durable. It isn't going anywhere. Nano Banana Pro is not taking that away and neither is any other image generator. It is very valued dense. It's where Moes live. It's why I'm not super worried about Salesforce for the medium to long term. Layer

number two above that system of record is intent planning and operation. And this is the layer that interprets it says show me if you say show me which enterprise customers in AMIA have renewal risk this quarter and give me a CSM touch gap no longer than 45 days and then please draft an outreach email. That's a series of tasks that an AI agent can pick up pass off to other AI agents and start to execute against the system of record. Layer two is becoming an agentic layer. It's not all the way

there yet, but I don't know anyone who operates a B2B SAS company that isn't working on some version of layer 2. And in fact, most businesses are working on some version of layer 2 for their back office operation because this kind of experience is what we have all wanted software to be and we never got a chance. If you remember back when I said software was something we had to conform to, we never really wanted that. We wanted software to be more personal and

with an agentic layer over the top of a solid data foundation, we finally have that chance. So that means the agent can hit your CRM, it can hit your customer data warehouse, it can run the queries, it can call the email system, the ticketing system, it can decide what needs a UI and what ought to be autoexecuted. All of that can happen and then you can finally get to the UX. Layer three is pixels, but not pixels as the handtoled crafted objects that we had to live with back in the Oracle eyes

store days. I mean pixels as a compiled artifact of intent. So only only when it needs your judgment does the system compile pixels in this model. It might be a one-off panel, right? It may have a rank table of atrisisk customers. It might have an inline suggested outreach. It might have a toggle for send now schedule and design. And it's a transient visualization. It's a specific cohort ch cohort chart or funnel for this question only and a narrow editor UI for exactly one structured decision.

store days. I mean pixels as a compiled artifact of intent. So only only when it needs your judgment does the system compile pixels in this model. It might be a one-off panel, right? It may have a rank table of atrisisk customers. It might have an inline suggested outreach. It might have a toggle for send now schedule and design. And it's a transient visualization. It's a specific cohort ch cohort chart or funnel for this question only and a narrow editor UI for exactly one structured decision.

In other words, we are moving to a world where at least some of the UI does not generalize. I am not trying to suggest that all of the UI is going to be composable. And part of why I'm not is that we are creatures of habit. We have a lot of assumptions around how UI ought to look and we do get used to our software products pretty quickly and we don't like it when they change. I think there are going to be common cores in our software stacks that remain durable

even if they're UI. Think of it as the homepage for a B2B SAS that shows you customer conversations and you want that homepage to be easily navigable and you don't want it to be new and different. I think that kind of UI is here to stay. It's not going to be AIdriven. I think the key is that there are going to be a whole new class of user interfaces that nest under that that are going to be heavily used that are generative that are throwaway that are rendered at

runtime for that particular person. We are arguably already doing this when we create an interface on the fly through perplexity and then share that throwaway interface with one or two other people as a way of talking about a topic. We're starting to do it in chat GPT when we have shared conversations and chat GPT creates an artifact that we both view together. So interfaces are becoming this sort of twoclass object where you have durable permanent interfaces that

may be a common core that has high habit that is the front door of the application and this disposable layer that sort of makes up for a lot of the pages that were handtoled before that never got a lot of traffic. Anyone who has managed a SAS application will tell you that traffic decays stochastically. Traffic decays like this on an exponential curve and your top two or three pages account for most of your traffic. But you have to put just as much work into all these other pages

that only a couple of people want. Those are the pages that I think are largely at risk during this transition. We are going to see SAS applications that only have two or three main pages and everything else may be generated for the user on the fly. Sure, the user may be able to save it in some place so they can come back to it if they like that particular view, but fundamentally they're going to be much more composable than and that brings me, I think, to a chance to talk about the differences

here because I want to be really clear about how different a coherent consistent handtoled interface is versus a disposable pixel pixel interface. If you want to lay that out and talk about different horizons and axes of value, they could not be more different. The time horizon for a traditional interface is measured in months at best. And for disposable pixels, it can be done in seconds. The design target like you often have a lot of people focused on personas, roles, generalized workflows

for your fancy interface. And for disposable pixels, the agent is going to decide. It's not going to be a human. the agent is going to put a user, a moment, and a goal together and go somewhere. Your mental model for a coherent interface app is learn this app. And I think that is actually one I would really like to emphasize from a talent perspective. Most of the talent at tech companies and at non- tech companies still has the mental model of learn this app and they've brought that

with them to chat GPT in the AI era. that does not serve you because the world we're moving to with disposable pixels is more like what AI actually is. State your intent, do the prompt, and UI appears when needed. And that could not be more different than assuming that the app is static and you can learn it. And I think so much of the time we assume that the cost structure, I've called this out, it's so different. Instead of a heavy upfront cost for traditional

software, disposable pixels, they have heavy model training, but the pixels are functionally free. The models have been paid for and you can get cheap, cheap, cheap iteration. Even Nano Banana Pro, which is relatively expensive now and will get cheaper, it's still dirt cheap, relatively speaking. The consistency is something I want to call out. This gets viewed as a concern for a lot of generative interfaces. Consistency value is obviously very high for traditional

software. It is mostly inside the agent planning and the durable state and record layer. It is not in the pixels. And I think that a lot of times proponents of generative UI fail to make this connection. They tend to say that generative UI is whatever you want it to be without recognizing that it has to rest on a durable software substrate that does not change that is not ephemeral. Differentiation or how software differentiates from others is also in and of itself different. So let

software. It is mostly inside the agent planning and the durable state and record layer. It is not in the pixels. And I think that a lot of times proponents of generative UI fail to make this connection. They tend to say that generative UI is whatever you want it to be without recognizing that it has to rest on a durable software substrate that does not change that is not ephemeral. Differentiation or how software differentiates from others is also in and of itself different. So let

me explain what I mean. In the traditional software days, if you were pitching your software as VC era, we were software is better. You would call out look and feel. You would call out interaction design. You call out UX patterns. You'd call out the the smarts of the machine learning inside. You would call out the cleanness and efficiency of your workflow. The way you'd understood the problem. With disposable pixels, you call out the outcomes because the AI agents are doing

more and more of the work. You would call out the speed from intent to action. And as an example of speed, it took me 10 seconds to craft a perfect chart of GDP annually in the US and Germany compared on the same chart in Nano Banana Pro from 1960 to 2025. 10 seconds. You're not going to beat that with a traditional BI tool. The speed from intent to action is addictive and it is driving consumer and business behavior. And I think that we are fooling ourselves if we think anything

else. Look, coherent interfaces are not going to disappear. They're just going to stop being the default shape of software. They're going to become perhaps a fallback when tasks are ambiguous. They're going to become a shared frame for multi-user collaboration. They're going to become a meta surface where you orchestrate agents. It's just going to look different. I want to go a bit deeper here on the B2B SAS side partly because I am very deep in B2B SAS myself and I

else. Look, coherent interfaces are not going to disappear. They're just going to stop being the default shape of software. They're going to become perhaps a fallback when tasks are ambiguous. They're going to become a shared frame for multi-user collaboration. They're going to become a meta surface where you orchestrate agents. It's just going to look different. I want to go a bit deeper here on the B2B SAS side partly because I am very deep in B2B SAS myself and I

think this also hits B2B SAS profoundly and I want to call that out for if you're like buying B2B SAS or if you're a leader in B2B SAS if you're a builder in B2B SAS that should cover a lot of folks. This is a big deal. So the disposable pixel story is extra complicated and I think it justifies a little sidebar here. First I want to call out that right now today a ton of the enterprise value is framed around this idea that we own the primary surface where the job happens right so

CRM think that way ERPs think that way HR information systems think that way PLG analytics systems think that way if the primary interaction moves to an agent or co-pilot surface then your own UI is just a reference implementation it's not the default touch point anymore and so your API behavior behavior, your data semantics matter more than your navigation bar. So the bundling power shifts from is this the system with the best dashboard, which is what sales has

sold on in B2B SAS for a really long time, to is this the system that is easiest for agents to choreograph. And I think a lot of companies don't have a good answer to this. Also means that UI is becoming a product surface that you do not fully control. If customers are using generative UI tools on top of your APIs, they are letting their own internal design systems and their own models render their own views of your data. And then your Canon UI is just one of many frontends. And so you're

competing with internal task panels, with co-pilot generated micro apps, with perhaps a third party universal workspace tool that comes along. In other words, you are at risk of disintermediating the relationship because you get aggregated with many other SAS products behind one agentic interface. And so where SAS still wins is where it's able to be a substrate as a service where you own the canonical state for something, the contracts, the ledgers, the records, the risk models,

whatever it is. And that means that you are embedded in domain flows that track real value. So SLAs's compliance reference data being safe and predictable for agents to call is a way to win. So if you have strong schemas, if you have good safeguards, if you have item potent item potency, say that three times fast, in a disposable pixel world, you become less of a thing with screens, which is what most software has been, and more of a high integrity service that agents and generators can rely on.

Let's transition to the talent side. What happens to designers, PMs, and engineers in a world where we start to have generative UI? For designers, you have to shift the way you think, right? You're the designers on your team, the designers you hire. If you're a designer listening to this, you are moving from owning specific flows and screens pretty rapidly into defining interface grammarss, into defining constraints, into like figuring out safe snap points for generative UI. You are becoming

language designers and safety engineers for human attention. If you're a PM, you're used to a world where what feature or page do we build next is the core question. You're moving to a world where what intents do we support? What state changes must be safe? What decisions need human judgment versus being fully automated? So instead of just creating a static wireframe, you're moving to a world where you're trying to spec out intent, state, and outcome loops. And that's really different.

Engineers, especially front-end engineers, are used to front-end pixel pushing. And now you need to start thinking about building stable interfaces for agents and generators and a thin canonical shell. You may want to build something that enables those snap points. You may want to build something that enables validation logic. You may want to build something that enables a degree of composability within safe constraints. And so your interface backlog for for designers, PMs, and

engineers begins to change here because instead of traditional tickets that come in in Jera, you have new intents that you want to support, new system behaviors, new constraints or invariants, new components or layouts the generator might use. It's not just add another settings page. Now I do want to call out there are places where coherent traditional software still wins. Cognitive mapping is a big one. So humans do like stable landmarks. I mentioned this earlier. If you are doing

complex work like trading, like medicine, incident response, people rely on deep spatial memory of their tools. Completely shifting pixels every time adds cognitive load and risk. This is one of the places where I think perplexity is making an incorrect choice in the finance space. Bloomberg terminal may look like a maze to most people, but it is software that people with a deep spatial memory of the tools rely on for complex work. It is not getting disintermediated by perplexity finance.

Whatever perplexity says there's a floor of coherence that you cannot cross without hurting performance. I would also like to call out audit, training, and compliance is a big flow here. Regulated environments need very reproducible flows. Show me exactly what the user saw when they approved the loan is not something where you can say it was a generative interface. So IDK like that's not going to work with an auditor. Ephemeral UIs make this very hard unless you can capture and version

the UI spec itself as a first class artifact and that gets very very complicated very very fast. I think that the the incentives are strongly in favor of coherent software there. Team collaboration is probably also a space where you're going to see coherent software. So shared work needs shared views. Look at this dashboard. Check this queue. And if everyone has a different ephemeral panel, you need explicit mechanisms for pinning, for sharing, for standardizing those panels.

I am going to go out on a limb and I'm going to suggest I don't know this is true, but I'm going to suggest that Slack has basically this vision for their product roadmap. Slack is becoming a place that is benefiting from the move to generative UI. Not because Slack is itself a generative UI. It's very stable, but because it is stable and it is a place where teams collaborate and know the interface well. It is a place where all those hooks that Slack has built into other tools can become

passively agentified. The agentified benefits can just flow into Slack as a value proposition. And so when people build charts in Nano Banana Pro, the demo videos they do always show them popping the chart back into Slack where the team can see it. That is not an isolated incident. That is where Slack's value proposition is starting to shift as a stable team collaboration substrate in a generative UI world. So the mature pattern is probably a spectrum. You're

passively agentified. The agentified benefits can just flow into Slack as a value proposition. And so when people build charts in Nano Banana Pro, the demo videos they do always show them popping the chart back into Slack where the team can see it. That is not an isolated incident. That is where Slack's value proposition is starting to shift as a stable team collaboration substrate in a generative UI world. So the mature pattern is probably a spectrum. You're

going to have highly standardized and coherent shells for regulated flows, for shared operational views, for team training and onboarding, for team collaboration, and you're going to have disposable pixels that operate inside that shell for exploratory analysis, for micro decisions, for personalized shortcuts, for just for me flows. I want to suggest to you that it is okay that we have both and that we do not have to insist on a binary fight like I see so many times where people will say B2B SAS

is dead and only generative UI is the future. We will never have stable interfaces. That's a terrible take. But an equally terrible take is we will never see generative UI interfaces in serious SAS applications. That is just not true. And anyone who has managed a serious SAS application as I have will tell you that we have hundreds or thousands of pages that we're managing, many of which we would dearly love to make generative because they're so expensive to maintain through the

traditional rubric. And so when I step back and look at the implication of this nano banana moment for builders, for leaders, for talent, I think the thing that I want to leave you with is this. Software really is decoupling. It's decoupling into a substrate that needs to be stable and a pixel that matters a whole lot less. If you are in the business of either pixels or substrates, this is going to affect you. You should pay attention. You should think about your moat. Is your mode on the

traditional rubric. And so when I step back and look at the implication of this nano banana moment for builders, for leaders, for talent, I think the thing that I want to leave you with is this. Software really is decoupling. It's decoupling into a substrate that needs to be stable and a pixel that matters a whole lot less. If you are in the business of either pixels or substrates, this is going to affect you. You should pay attention. You should think about your moat. Is your mode on the

substrate? You should think if you're in talent, if you're in design, if you're in PM, if you're in engineering, where are you at in relation to the substrate and the pixels? Are you stuck in a world where you're pushing coherent software and you don't see a way forward or are you moving to that world where you have the substrate, the agentic intelligent layer and the disposable pixel? I do believe B2B SAS survives as the substrate and there will be coherent

substrate? You should think if you're in talent, if you're in design, if you're in PM, if you're in engineering, where are you at in relation to the substrate and the pixels? Are you stuck in a world where you're pushing coherent software and you don't see a way forward or are you moving to that world where you have the substrate, the agentic intelligent layer and the disposable pixel? I do believe B2B SAS survives as the substrate and there will be coherent

cores that survive up to the UI layer, data providing agentic intelligence layers over the top etc. But fundamentally pixels themselves as the the single coherent interface for a product are going to go away. We have seen that going away for a while as BI teams have leaned more and more into just give me the data for data platforms and data vendors. They don't want the fancy dashboard the sales guys sell. They just want the data. Well, now we're moving to a world where it's not just

the data science team saying that. It's the marketers. It's everybody is saying that. So who wins? Products that are agent addressable. Products that are schema clean. Products that can be composed. Teams that treat UI as a language and a runtime, not as a set of frozen screens. And that goes for you as an individual. It goes for the people you hire. Who loses? Products whose only mode is that your interface is beautiful? Vendors who resist being called by higher level agents and insist

that users live inside their monolith. Like you can only do that for so long. people will find a way around it. One of the implications of nano banana is that a computer use agent that is very good is not far behind. And even if you insist on living in the monolith, you could see a world in 2026 where the user can just get up in the morning, have a voice conversation with an agent, and the agent can use a tool to go and browse the monolith software that you

insist only a human can use, extract the data, and bring it back to the user. The user is going to be able to make their choices. The user is going to be able to choose their interface. This is going to be true for consumer. It's going to be true for business. And it's going to change everything.