--:-- Host (Erik Torenberg): We thought that we were looking at the form factor of AI, which is you're talking back and forth to something. The real ultimate end state of AI and thus AI agents is these are autonomous things that run in the background on your behalf and executing real work for you. The more work that it's doing without you having to intervene, the more agentic it's becoming. Somehow it produces output that it feeds back into itself.
--:-- Host (Erik Torenberg): It's literally just the ampersand in Linux, which is a background task. And it's like the worst assistant in the world. And agentification is just hiring a lot of these really bad interns.
--:-- Host (Erik Torenberg): What exactly is an agent and how will agents change the way we work? To unpack this question, we brought together three people with deep but very different vantage points. Aaron Levie, co-founder and CEO of Box. Steven Sinofsky, former Microsoft exec and a16z board partner. And Martin Casado, general partner here at a16z.
--:-- Host (Erik Torenberg): From the old school idea of agents as just background tasks to today's vision of fully autonomous systems, we'll explore what this means for coding, enterprise workflows, and how whole industries might reorganize around agents. Let's get into it.
--:-- Host (Erik Torenberg): I thought I'd start this wide ranging podcast by asking the very simple, very provocative question. What is an agent?
--:-- Martin Casado: Oh, boy.
--:-- Host (Erik Torenberg): To who?
--:-- Steven Sinofsky: Okay, Steven.
--:-- Host (Erik Torenberg): To me? Exactly. Oh, I go for it.
--:-- Steven Sinofsky: So I actually have a very old person view of what an agent is, which is it's literally just the ampersand in Linux — it's a background task. Because you type something into O3 and then it's like, "Hey, I'm trying this out. Oh wait, I need a password. Can't do that." And it's like the worst assistant in the world. And really it's just because they need to entertain you while it's taking a long time to answer your prompt. So that's my old person view of what an agent is. And agentification is just hiring a lot of these really bad interns.
--:-- Aaron Levie: The interns are getting better.
--:-- Steven Sinofsky: They are, but they still don't remember if I have a password.
--:-- Aaron Levie: Is it possible you guys just had bad interns in the 80s and 90s?
--:-- Steven Sinofsky: We had terrible interns. I have a very high esteem for interns. But now a real answer.
--:-- Aaron Levie: No, I mean, I think collectively we're seeing what these are becoming. So if you think about two years ago, the post-ChatGPT moment, we thought that we were looking at the form factor of AI, which is you're talking back and forth to something. And I think to Steven's point, the real ultimate end state of AI and thus AI agents is these are autonomous things that run in the background on your behalf and execute real work for you. Ideally, you're interacting with them relatively little relative to the amount of value they're creating. So there's some kind of metric where the more work it's doing without you having to intervene, the more agentic it's becoming. I think that's the paradigm we're seeing.
--:-- Martin Casado: The only addition I'd have, in addition to long running, which I agree with, is that somehow it produces output that it feeds back into itself as input. You can actually do long running inference — you can make a video that's really long running, but it's just a single shot video and you just throw more compute at it. I think there are tactical limitations if you start feeding the input back in, because we're not quite sure how to contain that. So I think you can measure things based on how long they run, and you could also measure it by how many times it's actually taken its own guidance, which would be more of an agency.
--:-- Steven Sinofsky: Yeah, because I do think it's important that in this transition, look, what Aaron described is where we're going to be. It's just that what are the interesting steps that happen along the way? Because we are going to need, for the time being, it to stop and say, "Am I heading in the right direction or not?" Because putting aside all the horror stories about taking action without consent and using accounts and data, there is this thing where you just don't want to waste your time on the clock while it's churning away way off in the wrong direction.
--:-- Martin Casado: So the question is to what extent do they have their own agency, which to me means they've spit something out and they've consumed it back up again and it's still a sensible thing. Which, by the way, as you start thinking of these things in distribution, it's actually a very difficult thing to do because it doesn't know if it's going to be spitting something out that's still in distribution when it brings it back in. They don't have that self-reflection. So I think there's a very technical question here to what extent we can make these things have independent agency. But we can make them long run pretty easily.
--:-- Aaron Levie: Yeah, we're good at the long run.
--:-- Steven Sinofsky: The long run is what you get back.
--:-- Aaron Levie: I think the interesting thing is how the ecosystem is solving or mitigating the issues. You're seeing this logical division of the agents. So they might be long running, but they're not actually trying to do everything. The more that you subdivide the tasks out, the more that they can go pretty far on a single task without getting totally lost on what they're working on.
--:-- Steven Sinofsky: Unix is going to prove to be right — you're going to want to break things up into much smaller granularity and tools. And I think to other points that you've made on X, you're going to want to divide things up so that it's like an expert in this thing.
--:-- Aaron Levie: Yeah. And then it might be a different body of code where you go and ask, "Are you good at this thing? Let me get your answer on this part of the problem."
--:-- Martin Casado: It's kind of interesting. The conversation on AGI has evolved very clearly in the past six months. I think the consensus was, maybe not even consensus, but some of the view two years ago was this sort of monolithic system that's just super intelligent and it solves all things. If you fast forward to today and whatever we agree is state of the art, it's looking like that's probably not going to work for a variety of reasons, at least in today's architecture. So what you have is maybe a system of many agents, and those agents have to become very deep experts in a particular set of tasks. Then somehow you're orchestrating those agents together. Now you have two different types of problems: one has to go deep, the other has to be really good at orchestration. That maybe is how you end up solving some of these issues over the long run.
--:-- Steven Sinofsky: I just think it's very difficult to think cleanly about this. I've still yet to see a system where they perform very well and you don't draw a circle that doesn't have a human being in it.
--:-- Aaron Levie: Oh, yeah.
--:-- Steven Sinofsky: So in a sense, the "general" seems to be coming from the human. These things are tremendously good at increasing productivity of humans. At some point, maybe they'll increase productivity without humans, but until then, it's just very hard for me to actually talk cleanly.
--:-- Martin Casado: Well, it's so important for people to get past the anthropomorphization of AI, because that's what's holding everybody back. AGI is about robot fantasy land, and that leads to all the nonsense about destroying jobs. None of that is helpful because you have to dig yourself out of that hole to just explain, "Wow, it's really good at writing a case study."
--:-- Aaron Levie: Right. Which it writes a better case study than all the people that work for it, but it doesn't know who to write it about. It doesn't know what you want to emphasize. It doesn't know what the budget is, what's needed, how many words.
--:-- Martin Casado: Right, but it also turns out AGI just does an awful lot of work.
--:-- Martin Casado: So, for example, someone asked me recently, "Are you worried that if we have AGI, you'll no longer be investing in software companies?" I'm like, "I'm still investing in software companies, right?" Just because you're AGI says nothing about economic equilibrium or economic feasibility. The term AGI does infinite work for every kind of fear we have and maybe every hope that we have. When we tie it down to not only solving a class of problems, but the economics pencil out, we can actually have a more sensible discussion. I think that is finally entering the discourse. I think we're talking a lot more sensibly now than we were a year ago.
--:-- Host (Erik Torenberg): So when people say things, or the AI 2027 paper, when they talk about automated research, recursive self-improvement — does that feel like fiction or fantasy? Or does it feel like even with those things, we're nowhere near peak software and there would just be unlimited demand?
--:-- Martin Casado: I think you've got to go first for each question.
--:-- Steven Sinofsky: I don't want to stop answering the question.
--:-- Host (Erik Torenberg): I need you to anchor us in reality and then we can deviate.
--:-- Steven Sinofsky: Right. Well, I think that first, I'm just not a fan right now of buying into anything by year. Because whatever year you want to buy into, in 2027, we're just going to be having a fight over what we meant by the metrics. It turns into OKRs for an industry, which is just a ridiculous place to be. But I think that everything takes 10 years, but you can't predict anything in 10 years. So how do you reconcile that? You just have to recognize that we're on an exponential curve. No one's predictive powers work. It's just going to keep happening. It's not going to plateau. It's not going to suddenly be done. That's what makes this a different kind of platform shift.
--:-- Steven Sinofsky: If you look at the progress — that same thing went through with storage, with bandwidth, with productivity on computing, with connectivity around the world. Because it's exponential, you can't predict it. It's folly to sit around and try to predict. You could do science fiction and say in the future when we all have our personal AI, that's great. But then you say it's going to happen in 2029, you're an idiot.
--:-- Aaron Levie: That sounds totally correct. Three years ago, you would not have been able to conceive of Cursor or name your background agent writing code. So what is the point of having some date at which you're naming something? We've seen vastly more progress in the past two years of actual applied AI than we would have thought. Does it matter that one or two of the predictions didn't play out? No. So I think it's probably more interesting to think about where the technology is from a classic Moore's Law standpoint. How much compute do we have? How much data are we working through? How powerful are these models?
--:-- Steven Sinofsky: Let me ask you, as semi-old — after AI collapsed and machine translation and machine vision failed, you couldn't find anybody who thought that those would become solved problems. Or after neural nets imploded, and you were literally teaching — if you tried to teach neural nets, the students would rebel because you were wasting everybody's time. In 1989, Hinton couldn't get funded trying to do neural nets. Grad school was this three-volume history of artificial intelligence. Neural nets was like eight pages.
--:-- Martin Casado: Ironically, I remember when ML was the cool thing and neural net was the old thing. Now ML is the old thing and neural net is the cool thing. Or NLP. So the fact that we will return to all of these problems that couldn't be solved — like everyone's favorite one, "Oh, it doesn't understand math." That is a solvable problem because math is solvable. No one put the math layer in to understand what a number was and to hard code it and just build in an expert system for math, which is actually a well-understood thing because we've had Maxima since 1975.
--:-- Martin Casado: I think it's important to describe how hard it is to predict anything. Let's take recursive self-improvement. This is one of my favorites. The theory of recursive self-improvement is you have a box, which is the thing, and then there's an arrow that goes back to the box which says "improve." You look at that and you think, "It works, right?" From an intuitive lay perspective, every time you have a box with an arrow back in it, you think, "Okay, we're done." But if you know anything about nonlinear control theory, answering that question is one of the most difficult questions we know in all of technical sciences. Does it converge? Does it diverge? Does it asymptote? For example, you could recursively self-improve if you're doing basic search, but you asymptote. Saying recursive self-improvement from a deeply tactical perspective says almost nothing. But because we tend to anthropomorphize AI, we say recursive self-improvement and all of a sudden it overcomes energy boundaries and human intelligence.
--:-- Steven Sinofsky: That's how it goes from being a toddler to being an eight-year-old. It just figured out how to learn. It recursively self-improves.
--:-- Martin Casado: The reality is nonlinear control systems, which are feedback loops that are adaptive — we don't even have the math for a relatively simple system to understand what happens. You have to actually know the distributions that come out and go into them. These things are going to improve. Maybe they'll improve themselves. But just because they do improve themselves doesn't mean they can continue to do it. This is part of this entire journey as we're learning about these systems. The good news is I think we're talking a lot more sensibly now than we were a year ago. Hopefully the discourse can recursively self-improve so we're just more sensible.
--:-- Steven Sinofsky: The good news is that involves humans, so we don't actually accept them on.
--:-- Host (Erik Torenberg): But I think you must be seeing this even with customers. Take the conversation about hallucinations. How dramatically that's altered in just the past two years.
--:-- Aaron Levie: On two dimensions, actually. On one dimension, the problem of hallucinations has improved. As the models get better, as our understanding of RAG or the efficacy of the context window has improved, you have technical improvements across the stack. Equally, you have a cultural understanding within the enterprise that these are non-deterministic systems. They're probabilistic. So you're starting to see a culture shift — you can actually implement AI in more and more critical use cases because the employees using those systems understand they have to do the work to verify it. The only question is what is that ratio of time it took to verify versus if I had done it myself, and how much efficiency gained for that workflow.
--:-- Aaron Levie: We went from two and a half years ago where there was instant excitement — "This is going to be the greatest thing of all time" — to a reality check within three to six months because everybody was like, "Hallucination is going to be the massive problem." Now a couple of years later, we're seeing the hallucination rates shrink. We're seeing the quality of the outputs increase. And we understand that you have to go and review the work these AI agents are doing. That takes on a different form depending on the use case. In the form of coding, that means you just have to go review the code.
--:-- Steven Sinofsky: Which you had to do anyway. People seem to be forgetting.
--:-- Aaron Levie: You had to do anyway, but there was probably a little bit of theory as to what part you should go review with extra detail because you knew the person you were working with. But it also implicitly limits the value of AI, which people are uncomfortable with.
--:-- Martin Casado: Right. It just says it helps people that will know more than the AI does in this business. It starts to bisect the utility.
--:-- Aaron Levie: Yeah. It's super interesting — the experts are now becoming the productivity of an expert is outpacing everything else. I think we could have probably predicted that based on historical events.
--:-- Host (Erik Torenberg): We're seeing that where expert engineers are like, "I don't mind that it's a slot machine where I'm pulling it and I see what comes out because I know I can still get 10x productivity. It gives me good ideas."
--:-- Aaron Levie: Yeah, and I get it good enough that it's worth that productivity gain. Whereas if you were not an expert engineer and you did this slot machine, you would try to deploy all the ones that were also wrong. You don't know which lever to pull.
--:-- Steven Sinofsky: That's a big thing — literally knowing what to ask for and what language to use will get you a better response.
--:-- Martin Casado: I think this is an incredibly important point. It gets to the heart of what it means to use a tool. You put me in front of a 12-inch chop saw and say, "Go fix the fence." Really bad idea. I could go buy one. I could cruise the Home Depot. But it's really not a particularly good idea. How these platform shifts happen and why there's so much excitement over coding is that the best way for a platform shift to take hold is the experts — the closest you have to an expert in the new platform — become the most enthusiastic and the biggest users.
--:-- Martin Casado: I've been practicing yoga at the Cubberley Community Center in Palo Alto because the studio is closed. What's neat is that was the OG place for computer clubs in the early 1990s and late 80s. If you ever wanted to meet the computer club, you would go there. It's like "Halt and Catch Fire" — a bunch of people with soldering irons. When something was broken, that wasn't "These things are terrible, I'm wasting my time." That was the whole meeting: "Who can get one of these new discrete graphics cards to actually work and debug the driver? Can anyone print in this new thing called PostScript?" I think that's what's really happening right now.
--:-- Martin Casado: First, it's obvious it should happen with development and coding first because they're the most forgiving and the most understanding of what's a bug and what's a thing that can never get fixed. No one is saying that coding can't get fixed. Whatever it's generating that's bad for a 2x coder rather than a 10x coder, no one is saying, "That'll never be fixed." The next thing that's going to happen is going to be the creation of words — the marketing document, the positioning document, all of this long form stuff. If you're really good at that job, you know the right questions to ask, you know what looks good, and then you can get really domain specific. The next level is, "I need to understand a competitor," which uses real information from the internet in real time, not just statistical. And then they already know what the competitor does.
--:-- Martin Casado: My favorite scenario that constantly has these aha moments is: "Attack this thing I just wrote. I'm not interested in you adding em dashes and making it a little bit better. I want to know what did I miss?" You said on this last one, "Here's my earnings statement. Just improve it. Attack it like an analyst." There are 6,000 hours per company of analyst questions. It knows what they're going to ask. They only ask three questions anyway, on the expense line.
--:-- Steven Sinofsky: Do not watch this if you're an analyst. This is not advice about being an analyst.
--:-- Martin Casado: But this is what's really going to happen with writing, and then with PowerPoint and slides, and then with video. It's really important to call out that you're getting the consensus mean response. In the limit, it's offloading a lot of busy work if you're a professional. If you're a professional, you actually know all of these things, you just don't have the time to go through all of it. You may not remember it. So it's productivity helpful, but it's not solving problems where you are a particular expert. This is maybe why for those that are non-expert, it's a little bit more threatening because it can do that job.
--:-- Host (Erik Torenberg): Steven, you're asking where the enterprise is now. The coding piece — I think where you're seeing this is a clear understanding that what I'm going to get out will be correlated to what I put in. How precise I put the prompt — I think prompting doesn't go away anytime soon simply because the leverage you get on the set of instructions you're going to give the AI at the start is still going to be massive.
--:-- Steven Sinofsky: Wait, what would prompting going away mean? What would you end up with?
--:-- Host (Erik Torenberg): Two years ago, people thought you'd just tell the AGI what you want it to produce. One prompt. You unbox it and say, "Go do something. My agent is the software engineer." That was an open debate. But what is in my head is going to be unbelievably germane to the thing I'm trying to produce. I have to somehow give you that context. There's no world where you have that context without me telling it to you. Now you're seeing these incredibly unhinged prompts that are pages long. The output you're getting from that is way better than if you didn't give it that context. So there's a clear understanding of that side on the enterprise use cases and a clear understanding that you've got to go and review it.
--:-- Steven Sinofsky: We forget that formal languages came out of natural languages for a reason. We didn't start with formal actions. It's much easier to speak in English? It's the opposite. We have this natural language and it's very tough to convey the information that I want to. You and I are both experts. We understand the solution space, so let's communicate more efficiently. That's what jargon is. Jargon is just a formalized way that people who have domain expertise talk to each other.
--:-- Host (Erik Torenberg): The thing that is most fun to think about right now is, maybe you give us a little history lesson on this. When does the style of work change because of the tool versus the tool adapting to the style of work? I'm starting to see some patterns emerge. We thought agents would go and learn how we work and then automate that. Agents conform to how we work. The question is, when is the moment when we conform to how agents are best used? You're seeing this in engineering — people are saying, "I'm going to have agents and sub agents for parts of the code base. I'm going to give them readme files that the agents read. I'm going to optimize my code base for the agent as opposed to the other way around." Within how we use Box with our AI product, you're starting to see people tell the agent its complete job and the workflow is starting to be almost like the agent is dictating the workflow in the future as opposed to mapping to the existing workflow. I don't know what the history is on this of when the work pattern itself shifts because of what the technology is capable of. But I think where this goes has to be some version of that — it's not going to be that agents just plop into how we currently do our work and automate everything. You start to change what the work is itself, and then agents accelerate that.
--:-- Steven Sinofsky: As important as that is, it's actually more important. Because what happens is — to reuse the word in a different way, this anthropomorphization of work — the first tools actually anthropomorphize the work. Every single evolution of computing. How long did it take for Steve Jobs to get rid of the number buttons on a smartphone? They still had number buttons. Or look at cars — until Elon got rid of all the controls, everybody kept all the controls. I don't want to get in that fight. But with every technology shift, if you were to look at what accounting software looked like in the 60s before IBM said, "Stop. We all use double entry, but we need to have people skilled in how computers can do the accounting, not how people can, because we're never going to figure out how to close the books if we have to automate this whole room of people with green eyeshades that have a manual process based on how far apart the desks were."
--:-- Steven Sinofsky: Everything that happened with the rise of PCs and personal productivity started off — I always use this example because I've watched it happen five times now — the first PCs that did word processing, the biggest request was, "How do I fill in expense reports?" So this whole world grew up of tractor-fed paper that was pre-printed with the expense report. We wrote all of this code. "Are you using an Avery 2942 expense report? Or is it a New England Business Systems A397?" You had adjustments in the print dialog — 0.208 inches. You moved little things. You would print out eight dinner, $22. That was all you printed. Then someone said, "We could use the computer to actually print the whole thing." Fast forward, and finally Concur said, "Why not just take a picture of the receipt? Then we could do all of it." The whole thing gets inverted. Every single business process ended up being like that.
--:-- Steven Sinofsky: Then there are things that really change the tools. When email came along, to prepare an agenda for a meeting, somebody would open up Word and type in all the things and print it out. Everybody would show up to the meeting with this very well-formed document. Then email came out and that whole use case for Word evaporated. An email agenda became no formatting, nothing — just, "Here are the eight things we're going to talk about." You show up and everybody's like, "Did you get the agenda?"
--:-- Steven Sinofsky: You know, what's interesting about the AI one is it's kind of like we're seeing the same thing, but vis-a-vis AI. Nobody really predicted the generative stuff. And we've had AI for a very long time. We had chatbots, and you had these kind of AI-shaped holes in the enterprise for a long time. A lot of the mistakes that we see today are people taking the generative stuff and trying to cram it into the old models. It's really a new behavior that's emerging.
--:-- Martin Casado: It used to be you'd centrally sell AI to some platform team, and then they would try to get the NLP thing to work or the voice to work for talking to people on the phone for support. It was very central. A lot of the adoption that we see is much more individual, for example. I think there is a bit of a mismatch, as we're seeing now, that it's getting ironed out.
--:-- Aaron Levie: So I think the question is, are we in the phase where we're trying to graft the agents onto basically what we've been doing for 30 or 40 years of software? And is this going to be the first real step-function shift we've seen in what the workflow itself should look like?
--:-- Steven Sinofsky: Oh, we are. I mean, remember people tried to jam the Internet into Office. It was fun to watch. But everybody was trying to jam the internet into their product because that's the only way you could envision it. And it didn't really work. You were like, "Well, where else would the internet go? There's no word processor on the internet. There's no spreadsheet on the internet." And then other people would try to implement Excel using seven HTML tags with no script. That turned out to not be a good idea either. The best was, "Let's do PowerPoint." You give them five edit controls, tell them their bullet points, and then we'll generate a GIF on the back end and send it back to you as the slide. That was not it.
--:-- Aaron Levie: I think actually maybe the main point is just the durability of Office. It transcends all disruptions. It pretty much rises above everything. But the thing is that where we are now is everybody...
--:-- Martin Casado: Do you think this is similar to the internet in that it's a consumption layer change? I always view the internet as very much a consumption layer change. Instead of going to my computer, I go to the internet. But otherwise, things are the same. AI has this weird quirk, which for the first time I can recall, programs are abdicating logic to a third party. We've always abdicated resources — "I'll use your disks" — but I'm writing the logic. This time it feels like we're changing the consumption layer. When my son talks to an AI character, he's not going to WellsFargo.com, he's going to an AI character. That's changing how we're interacting with the computer. But also, these programs are no longer written by a human in the same way. I feel like the change is maybe a bit more sophisticated.
--:-- Steven Sinofsky: Oh, I think this is why it's a platform shift, and not just an application shift. Each platform shift changes the abstraction layer with which you interact with computing, but it also changes what you write the programs to. Do you remember ever abdicating logic? Here's a great example of how disruptive this can be. The first word processors in the DOS era all implemented their own print drivers and clipboard. If you were Lotus and you wanted to put a chart into a memo, you couldn't because you didn't sell a word processor. You actually made a separate program to make something that the leading word processor could consume. If you were Perfect, your ads said, "We support 1,700 printers." And you won reviews because you had 1,700 and Microsoft had 1,200.
--:-- Steven Sinofsky: Windows comes along. If you were trying to enter the word processing business, step one was hiring a team of 17 people to build device drivers for Epson and Okidata and Canon printers because you couldn't get them anywhere. Microsoft built print drivers and a clipboard for Windows. All of a sudden, there was a way that two applications that had no a priori knowledge of each other could work together. WordPerfect and Lotus got creamed by that because they lost control of their information. A bunch of developers were like, "Wow, this is cool. Now I'm just by myself." When we did C++ for Windows, I would go and show brand new Windows programmers in 1990, "Hey, you don't have to write print drivers and use the clipboard." Literally a standing ovation of 10 people. They were more than happy to let data interchange between products because that was nothing but opportunity for them.
--:-- Martin Casado: Can you... they probably felt exactly the same way as a vibe coder does today. "You've just given me the platform." And it was just a print driver. The writing code for Windows book was this big, but the writing a device driver for an Epson printer was this big. Writing it for a Canon printer was this big. I'm trying to think of the paradigm shift. There's been many times where we've reduced the amount of work a developer takes, but I just don't remember ever where the programmer abdicates logic. SDM didn't? Not logic — I would always say what is correct and what's not correct.
--:-- Steven Sinofsky: I think you undersold it though. This is the thing, by the way, that Martin invented and worked on, but it's a big deal. Maybe we should post-mortem your pitch at the time if you're not pitching this.
--:-- Martin Casado: No, no. Logic specifically. I am writing an app — some vertical SaaS app for a certain customer base. The answer the app gives is based on logic that I've written historically. If I run it on the cloud, the cloud is not producing an answer. It's providing resources. If I'm using your device driver, it's providing access to a device, resources. But if I'm like, "Hey, large model, tell me the answer here," you're actually abdicating application.
--:-- Steven Sinofsky: Maybe you're right. Maybe this is not... I think you're almost playing like an incumbent in the sense of trying to decide "this is abdicating the logic and this isn't." When in fact, it really was a huge competitive advantage for WordPerfect. They didn't want to give it up. They fought against it. The next example, of course, is the browser, where people literally gave up. In Windows or Mac, you could rasterize anything you wanted. You wanted a button that spun and animated like a rainbow — you could do that. But then the web came along, and you're like, "Wow, I have to use a gray button that says submit." And that was like...
--:-- Martin Casado: Yeah, because it's a point where you use a bunch of third-party things.
--:-- Steven Sinofsky: It took a long time for those to show up. Early in the internet, magazines and printed media were the ones who absolutely wouldn't go to the internet because they would not give up their ability to format. This is another part about the tooling and what's going to happen with AI. A huge amount of the productivity software space today is the preparation of output. Office is basically a format debugger. It's 7,000 commands for how to do kerning and bold and italic. It turns out AI not only doesn't care — you could ask it to make whatever you want. You could just say, "I'd like this to be a double-index pie chart thing." That's not a thing, but you can do that, and it will just figure out something that looks like that, and you'll go, "Ooh, cool."
--:-- Aaron Levie: This is where it's disempowering experts and who's not an expert. When productivity software arose, the big thing was that there were people who figured out how to make killer charts. Ben at the Ovens — killer chart guy. And there were people who were like, every meeting started with, "How did you make that chart?" I could be on an airplane and somebody would be making a shitty chart. In this case, the application is actually about what's the way to visually represent the data. Which is absolutely an abdication.
--:-- Steven Sinofsky: Right, and it turns out 90% of the people never really got to be expert at doing that task, even though 90% of the tool is about that.
--:-- Martin Casado: But the programmer didn't have to keep the logic in this case. This is the user. What's the user, what's the programmer in that?
--:-- Steven Sinofsky: What the programmer was doing was inventing things called wizards that would make a whole bunch of choices for you, style sheets or whatever. In a sense, we were making a bunch of choices for the user, which to the experts looked like disempowering the experts who were tweaking everything. There's some Steve Jobs quote he loves about Schopenhauer — "If you've seen the conjurer, it's not a trick anymore." I really feel like this is the third or fourth time that this has happened just in my lifetime.
--:-- Aaron Levie: Something that's really caught my attention because it's the most senior people I know is that a lot of very senior developers are spinning up a lot of background agents, like code agents, and they're interfacing at the GitHub PR level. It's not obvious to me why you do a bunch as opposed to one, and it's not obvious to me why you wouldn't interact directly. It feels like something's going on here, but I'm not quite sure what.
--:-- Steven Sinofsky: My read on it — and then I'll throw out what happens next as a result — is that engineers are just the first to experience this. From talking to folks who are all in on this, it's this mix of basically the context rot problem. The more that we put in the context window, the more it gets confused, the lossier the answers get. You have to have some way to partition what an agent should work on. We see this in building agents internally. The panacea we hoped for was, "You just put a million tokens into the context window, and then obviously..."
--:-- Martin Casado: So you're saying this is almost like a counter trend to the AGI. It's almost the opposite.
--:-- Steven Sinofsky: It's the opposite, but it only works because the models are so good. You're giving more things, more specific tasks, rather than one thing less specific.
--:-- Aaron Levie: Right, but I think this is why it's happening. The craziest version of this is I was talking to somebody in startup land who has all these subagents. What's amazing is it maps one-to-one to each microservice in their code base. They have an agent per microservice, they have a readme for the agent, and that agent owns the microservice. You could have dozens or hundreds of these things going on. You're effectively mitigating the issue where if you just said, "Here's my entire code base, go run wild," to one agent, it will produce worse and worse code over time because of context rot. It won't know exactly what you're trying to do in that one area of the microservice. The sub-agent model seems to be working for that paradigm.
--:-- Steven Sinofsky: I love this counter pattern. Everybody is like, "Models will get smarter and you'll give them higher level tasks and they'll do things longer." This is a counter one. I want to tweet that, but you have more Twitter followers.
--:-- Aaron Levie: We can collectively do it. So then the question is, let's just assume this works in engineering. You have this interesting dynamic — some of the coding practices will be pretty different in the future. We've talked about the individual engineer becoming the manager of agents. This is a supercharger of that concept. Then the question is, how does that translate to almost every form of work? If I am a lawyer working on cases and I can have 20 subagents that each do a different case and then come back in some kind of task queue, the sheer leverage is going to be insane. But the way you might organize the work and the workflows within an organization are inevitably going to change as a result.
--:-- Steven Sinofsky: This gets to the fact that the flow in the workflow has been serialized or linearized based sometimes on knowledge, but other times on tooling. When the tooling changes, you get a realignment of what's truly serial and what's not. If you're planning an event for a company, you have to book the venue, invite people, create materials. They're not particularly gated on each other, but if you have an events person, they're gated. Now an events person can start spinning up all of these different elements, and they're going to come back. "I've gotten as far as I can on collateral until I get a logo for this event. I've gotten as far as I can on invites until I get the date and time and venue." There's no reason you can't spin off all those in parallel. How does that happen today? If you're a company and you use Box, and this is your 58th event, you have a folder called Event. People take the folder and go Event 59, make a copy of it and all the stuff in it. That workflow is exactly what a series of different background tasks or agents could do.
--:-- Martin Casado: But there's also a bit of an indictment on the ability to give it a high-level task. It kind of suggests that the human being needs to be giving them more granular orders. Otherwise, to start a company, you'd issue one prompt, go to the beach for six months, and come back with a full company. This is almost re-anthropomorphizing — it turns out we did figure out division of labor. We figured it out in the context of a lot of physical analog limits that agents won't have. But there's no total free lunch. You have this context rot issue — you do have to subdivide the tasks at some point.
--:-- Steven Sinofsky: It may not be context rot. The Occam's razor here is you need to give them specific instructions for specific tasks. If you give them higher level instructions, independent of context, they just don't know what you want.
--:-- Martin Casado: This gets to the formal language part. At some point, if you tried to use the uber frontier to get the whole thing done, you have to tell it the whole thing. That seems like a lot of work. Whereas if you have to tell it less because the part of the model you're using knows more, it's a different way of thinking about templates or starting artifacts or scoping the context.
--:-- Aaron Levie: It might be the right architecture in general if you assume we're never going to get to a point where the model is 100% perfect. It might be the right design because you don't want an agent or a set of agents to go so far down a path when there was a step it needed to check in with you on, due to the compounding effect. You need to subdivide the work also because if you have gating moments with a bunch of dependencies, the agent needs to know at what point to roll that back up to the user.
--:-- Martin Casado: Against the common narrative, it seems the trend is prompts are getting more complex, not less. We're seeing more agents, not less, doing more narrow tasks. This is almost a counter-AGI narrative. These are much more specialized and much more deep, with much more specific instructions. There's a history of this — "Maybe we can actually solve it if we're specialized." Like expert systems. At first they thought expert systems would just be experts and they would just know. By the time you got to the actual published research at Stanford, it was, "This is an expert system in deciding on what type of infectious disease you have, as long as you have one of these seven."
--:-- Steven Sinofsky: There is one big difference, which is the model itself is packing in the inherent intelligence or capability to solve all of these. We are benefiting from the fact that you can build these all on Claude 4 and GPT-5. Let me try to demonstrate this with an old person example. Early in the PC era, there were word processors and spreadsheets and graphics and databases. A lot of people were like, "Why are there these four programs? There should only be one." My answer, which often involved screaming, was, "Have you been to an office supply store?" Because there's paper with numbers, blank rectangles of paper, transparency paper. This has been around a really long time. There's some physical human context for that.
--:-- Aaron Levie: How many minutes did it take you to know Google Wave wasn't going to work?
--:-- Steven Sinofsky: Zero. It was instantaneous. There was a product, ancient Mac product called ClarisWorks, which was lauded by the industry. It did everything — you could have a spreadsheet inside a word processor. My first reaction was, "Have you seen a person use a spreadsheet?" Their monitor can't be big enough. They want as many cells as possible. And you're sitting there saying it has to fit on an 8.5 by 11 sheet of paper on a Mac. These lenses that humans bring to specialization really matter. If you think about the medical profession, from a GP to a radiologist to a specialist to a nurse practitioner, they're each going to look at and use AI in a different way.
--:-- Aaron Levie: That level of specialization and division of labor emerged over a hundred-year period, driven by a lot of physical constraints and realities of how organizations emerge. The question is, in a post-agent world in 10 years, do those divisions of labor look exactly the same? Or do they shift because the agents collapse some of the functions? Is there some blurring? Clearly there's a role emerging in a bunch of organizations — "I'm the AI productivity person. I have a way of creating all new forms of productivity with AI." We'll have a bunch of new roles, but is our current division of labor going to collapse in some interesting ways because of AI?
--:-- Steven Sinofsky: If you stick with the medical example, we're just going to wake up and there's going to be way more people with way more specialties. AI will have created more jobs.
--:-- Martin Casado: Do you think AI causes more specialization over time?
--:-- Steven Sinofsky: Absolutely. Every human is going to be way better. More knowledge will accumulate. This has really happened with computing that people forget. There used to be this morass of marketing and R&D. There used to just be coding. Then there was coding and testing and design and product management and program management and usability and research. All of those had their own tools. Go to a construction site. I remember growing up, our neighbors built a house. There was Clem, the carpenter. You built a house with a guy named Clem who used all the tools. Now you build a house and it's a 20-person list of subcontractors, all with whole companies that do nothing but put in pavers. That's what it's going to be.
--:-- Martin Casado: There's been a long disaggregation in the history of IT. Everything in the same sheet metal. Then you disaggregate the OS and the hardware. Then you disaggregate the apps. In the last 15 years, we saw independent functions get disaggregated. Almost everything became like an API would become a company. Twilio, Auth became a company, PubSub became a company. It may very well be the case that every agent becomes a whole new vertical and a whole new specialization. You can actually build a company around it. Today, just like with APIs, one company will have a whole bunch of agents. In the future, a third party might provide that agent as an independent company.
--:-- Aaron Levie: The opportunity is really there for that. It used to be that the impedance to creating a company and distributing was infinite. It used to be ridiculous to think that a single API like Auth could become a company, but then of course it became a company. Or that you could build a whole company out of signing documents. Then you realize the addressable market for that is huge, and it's way bigger than signing because of all the stuff that got done that was just baked into a company causing headcount and waste and fraud and abuse.
--:-- Martin Casado: You can kind of underwrite thousands of these companies emerging. Jared Friedman had a tweet about going deep on a workflow — do the job of some part of the economy, like a payroll specialist, and then build an agent for that. It's not obvious that there aren't literally a thousand of those, by every vertical and every line of department.
--:-- Steven Sinofsky: I love this because it's literally the anti-AGI. It's following the long arc of computer science where as the market grows, the granularity can create a company.
--:-- Martin Casado: It's also economic growth. Take the payroll example. The idea of having a productive Salesforce used to just be a consultancy. The only way you could ever fix it was hiring a consultancy to show up, analyze what everybody does, do a report that says, "This is how you need to reorg," which usually meant go the opposite of whatever you have, and then they would leave. There was no cloud. To build CRM, you had to do all that consulting work and then roll it out. Then it was static and you couldn't maintain it. Then all of a sudden there's Mark Benioff and a whole way to do all this. Not only that, people actually like it. They think they're better at selling because they're using their phone and putting in a few notes about this client, which helps everybody. That's what's really going to happen with all this. Something that looks really small becomes a whole thing because there's no problem with distribution, no problem with customization. We'll have ways to solve security and privacy just like we solved reliability.
--:-- Steven Sinofsky: Look at the stuff you're a world expert in, the stack of internet technology, networking technologies. If you had asked me 15 years ago, "Would CDNs be companies?" I never would have thought it made any sense. How could you have a company that's a CDN?
--:-- Aaron Levie: I think people are probably way too afraid of the model providers eating them. It was a phenomenon in the first wave. If you were just doing basic stuff on GPT-2 and 3, where it was a text interface that produced more text, yes, ChatGPT ate you. That clearly happened. But since then, most enterprises want applied use cases for AI and AI agents. It's not obvious that the current crop of companies — if you're doing AI for healthcare, AI for life sciences, AI for financial services, AI for coding at the right parts of the stack — will get eaten. AI for coding may be the one asterisk area that will be hyper-competitive, simply because the model companies don't want to use somebody else's product to build their own models, so that almost forces them to get really good at AI for coding. With that as the one exception, I think we're just in a five-year period right now where you're going to have to build agents for every vertical, every domain, and there's a playbook starting to emerge of what that needs to look like.
--:-- Martin Casado: There was a technical head fake that happened early on, which was pre-training. Pre-training was a 10 out of 10 technical innovation. Two years ago, I had a friend building their own aging model — a text-to-image model — and they wanted to make it so old people looked really good. Then the next version of Midjourney comes out and does a better job. With pre-training, you're consuming all of the world's existing data, draining all that energy, and it perfectly generalized. But it feels like technically that's passed, and now we're more in post-training and RL, which is a lot more domain specific.
--:-- Steven Sinofsky: The moment that you have access to some set of data that is only for that enterprise — who gets permission to access that data? Who gets permission to do the workflow on it?
--:-- Steven Sinofsky: It's going to be applied companies. Yeah, so yeah. And if we had an infinite number of tokens, then the models would just continue to generalize. But it's pretty clear that that's not happening. So now we're going into something we all understand very well: companies have to choose which domains they go into, solve the long tail of problems there, and get access to the data, et cetera.
--:-- I also think that the shadow cast by large companies — the threat of "we're going to put you out of business and stomp you" — it's ridiculous. It has never in any technology wave lived up to the fear people have. Look, if you built a new word processor in 1995, you were an idiot. That was not the thing to go build. But just ten years earlier, companies built standalone spell checkers. It was just a thing. You went to the store and bought a spell checker, and it had more words than the other spell checker.
--:-- The thing that's not being said now — and we should do a whole episode on this — is: what is the actual platform?
--:-- Aaron Levie: Yeah, this is a great topic.
--:-- Steven Sinofsky: It's all well and good to say the large models will subsume every application. But the minute they start doing that, no one will be in their platform. No developer is going to sit around and say, "If you're going to subsume me, then..." There's a phrase for this — it's "Sherlocking" on the Mac, in the Apple world. It has a real chilling effect, and that's one of the things all the model people are going to learn very, very quickly.
--:-- There's a chilling effect, but there's also just a problem of it being hard to go deep in 50 categories. You just can't. I think everybody's scared because pre-training was actually the one thing that was good at that. And now they have to actually...
--:-- Martin Casado: Yeah, I agree. At some point it becomes purely an execution issue. I don't know how anybody would set up a company to be able to beat 50 startups across 50 different domains. It's ridiculous.
--:-- Steven Sinofsky: No, it's ridiculous. And in fact, it's only good because what happens is that the big company raises awareness of a whole category. Then you just swoop in and say, "To them, I'm just a feature. But to you, this is my whole life." And you're going to win.
--:-- I always come back to this: there's a whole company that just signs things. I cannot believe there's a whole company that just signs things.
--:-- Martin Casado: I have so much to say about this topic. Even minimally, if you graph the cost to produce versus the willingness to pay for an inference versus the cost to serve it — for most companies, for most spaces, 20% of the inferences are 80% of the cost. So the problem of the application is just to choose those ones on, which tend to be more domain-specific.
--:-- Aaron Levie: This is the problem of inviting the three of us on here — getting us to shut up is the trick.
--:-- Erik Torenberg: Guys, thank you so much for coming on. This is fantastic.
--:-- Host: Thanks for listening to the a16z Podcast. If you enjoyed the episode, let us know by leaving a review at ratethispodcast.com slash a16z. We've got more great conversations coming your way. See you next time.
--:-- [Closing music fades in]
--:-- Host: As a reminder, the content here is for informational purposes only. It should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a16z fund. Please note that a16z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com forward slash disclosures.
