Aaron Levie 与 Steven Sinofsky 谈 AI 工作者的未来

We thought that we were looking at the form factor of AI, which is that you're talking back and forth to something. The real ultimate end state of AI, and thus AI agents, is that these are autonomous things that run in the background on your behalf and execute 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.

It's literally just the ampersand in Linux, which is a background task.

Okay.

It's like the worst assistant in the world. 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.” It's really just there 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: agentification is just hiring a lot of these really bad interns.

I thought we'd start this wide-ranging podcast by asking the very simple but very provocative question: What is an agent?

Oh boy. To who?

Steven. Okay, exactly. Go for it.

I actually have a very old-person view of what an agent is, which is that it's literally just the ampersand in Linux, a background task. 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.” It's like the worst assistant in the world. 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.

The interns are getting better. They are getting better, but they still don't remember if I have a password to Notion. It's just—

Is it possible you guys just had bad interns in the ’80s and ’90s?

We had terrible interns.

I have very high esteem for interns.

But now, a real answer.

No, no. I think collectively we're seeing what these are becoming. If you think about 2 years ago, the post-ChatGPT moment, we thought that we were looking at the form factor of AI, which is that you're talking back and forth to something. To Steven's point, the real ultimate end state of AI, and thus AI agents, is that these are autonomous things that run in the background on your behalf and execute real work for you. Ideally, in an ideal world, you're interacting with them relatively little relative to the amount of value that they're creating. There's some kind of metric where the more work that it's doing without you having to intervene, the more agentic it's becoming. I think that's the paradigm that we're seeing.

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 basically just a single-shot video and you throw more compute at it. I think there are technical limitations if you start feeding the input back in, because we're not quite sure how to contain that. I think you can do it. 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.

Yeah, because I do think it's important that, in this transition, what I described is where we're going to be. It's just that we need to understand the interesting steps that happen along the way, because for the time being we're going to need it to stop and say, “Am I heading in the right direction or not?” Putting aside all the horror stories about taking action without consent and using accounts and data or whatever, you just don't want to waste your time on the clock while it's churning away, way off in the wrong direction.

Yeah. The question is: To what extent do they have their own agency? To me, that means they've spit something out, they've consumed it back up again, and it's still a sensible thing. 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. I think there's actually a very technical question here about to what extent we can make these things have independent agency, but we can make them run for a long time pretty easily.

Yeah. We're good at the long run.

The long-running you get back is—

Yeah, I mean, I think the interesting thing is how the ecosystem is solving, or mitigating, these issues. You're seeing this logical division of the agents. They might be long-running, but they're not actually trying to do everything. The more that you subdivide the tasks, the farther they can go on a single task without getting totally lost on what they're working on.

Unix is going to prove to be right, which is that you're going to want to break things up into much smaller granularity and tools. To the other points that you've made on X, you're going to want to divide things up so that it's an expert in a particular thing.

Yeah. Let's just say it's 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.”

It's kind of interesting. I don't know how much you've plotted this, but the conversation on AGI has evolved very clearly in the past 6 months. I think some of the view, maybe not even the consensus, let's say 2 years ago was that it's this sort of monolithic system that's just superintelligent and solves all things. Fast-forward to today, and whatever we agree the state of the art is, it's looking like that's probably not going to work, for a variety of reasons, at least in today's architecture. So what do you have? Maybe a system of many agents, and those agents have to become very, very deep experts in a particular set of tasks. Somehow, you're orchestrating those agents together. Now you have 2 different types of problems: One has to go deep, and the other has to be really good at orchestration. Maybe that's how you end up solving some of these issues over the long run.

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 has a human being in it somewhere. These things are tremendously good at increasing the productivity of humans. At some point, maybe they'll increase productivity without humans, but until then it's just very hard for me actually to talk cleanly about it.

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 then you have to dig yourself out of that hole to explain, “Wow, it's really, really good at writing a case study.” It writes a better case study than all the people that work for you, but it doesn't know who to write it about. It doesn't know what you necessarily want to emphasize. It doesn't know what the budget is, what's needed, or how many words.

But it also turns out AGI just does an awful lot of work. Someone asked me recently, “Are you worried that if we have AGI, you'll no longer be investing in software companies?” I'm like, “Well, I mean, you're AGI, right? I'm still investing in software companies.” Just because you have AGI says nothing about economic equilibrium or economic feasibility. The term AGI does basically infinite work for every kind of fear we have, and maybe every hope that we have. The more we tie it down to not only whether it solves a class of problems, but also whether the economics pencil out, we can actually have a more sensible discussion. I think that's finally entering the discourse. We're talking a lot more sensibly now than we were a year ago.

When you hear people talk about things like the AI 2027 paper, automated research, or recursive self-improvement, does that feel like fiction or fantasy? Or, even with those things, are we still nowhere near peak software, with unlimited demand?

I think you have to go first for each question. I don't want to stop.

I need you to anchor us in reality, and then we can deviate.

I would not buy into anything by year right now, 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 a ridiculous place to be. I think everything takes 10 years, but you can't predict anything in 10 years. So how do you even reconcile that? You just have to recognize that we're on an exponential curve.

So no one's predictive powers work.

Right.

And 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. If you look at the progress we've seen with storage, bandwidth, productivity through computing, and connectivity around the world, it's exponential. You can't predict it, and it's folly to sit around and try.

You could do science fiction and say, “In the future, when we all have our personal AI with all this other stuff”—and that's great. But if you say it's going to happen in 2029, you're an idiot.

Yes. That sounds totally correct, right? Three years ago, you would not have been able to conceive of Claude Code, Cursor, or any of these background agents writing code. What's the point of having some date at which you're naming something?

We've seen vastly more progress in just the past 2 years of applied AI than we would have thought. Does it matter that 1 or 2 of the predictions didn't play out? No. 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?

I mean, let me ask you, as a semi-old person—

Well, I mean, nobody thought machine translation and machine vision would become solved problems after AI collapsed and those fields failed.

Yeah, there. You couldn't find anybody who thought those would become solved problems, or that neural nets would return after they imploded.

Or expert systems.

But you were teaching—and if you tried to teach neural nets, the students would rebel because you were wasting everybody's time.

In 1999, Hinton couldn't get funded to do neural nets.

He was trying to do neural nets.

In grad school, I took this 3-volume history of artificial intelligence. Neural nets got about 8 pages.

You know, ironically, I remember when machine learning was the cool thing and neural nets were the old thing. Now machine learning is the old thing and neural nets are the cool thing.

Right, or natural language processing.

So we will return to all of these problems that couldn't be solved. Even everyone's favorite one: “It doesn't understand math.”

Right.

Okay, that is a solvable problem because math is solvable. There's just no one who put the math layer in to understand what a number was, hard-code it, and build in an expert system for math. That's actually a well-understood thing because we've had Macsyma since 1975.

You know, I think it's important for us to describe how hard it is to predict anything. Let's take recursive self-improvement. This is one of my favorite examples. The theory is that you have a graph with a box representing the system, and then an arrow goes back to the box saying “improve.” You look at that and think, “It works.”

Right.

From an intuitive, lay perspective, every time you have a box with an arrow back into it, you're like, “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 of in all the technical sciences. Does it converge? Does it diverge? Does it asymptote?

For example, you could recursively self-improve if you're doing a basic search, but you asymptote. Saying “recursive self-improvement” from a deeply technical perspective says almost nothing.

It says that—but unfortunately, because we tend to anthropomorphize AI, we say “recursive self-improvement,” and all of a sudden we're like, “Then it overcomes energy boundaries and human intelligence.”

Well, that's how it goes from being a toddler to being an 8-year-old. It just does it because it figured it out, right?

The reality is that nonlinear control systems—feedback loops that are adaptive—we don't even have the math to understand what happens in relatively simple systems. You actually have to know the distributions that come out of them and go into them.

These things are going to improve. They're going to continue to improve. Maybe they'll improve themselves, but just because they improve themselves doesn't mean they can continue to do it. That's part of this entire journey as we're learning about these systems.

Again, the good news is that I think we're talking a lot more sensibly now than we were a year ago. Hopefully that will continue. Hopefully the discourse can recursively self-improve, so we're just more sensible.

Well, the good news is that's involving humans, so we don't actually model it. But I thought—you must be seeing this even with customers. Take the conversation about hallucinations and things like that. How dramatically has that changed in just the past 2 years?

Yes, in 2 dimensions, actually. On one dimension, the problem of hallucinations has improved. As the models get better, our understanding of how to determine whether something is correct—whether it's RAG or the efficacy of the context window—has improved. So you have technical improvements across the stack.

Equally, you have a cultural understanding within the enterprise: These are nondeterministic, probabilistic systems. You're starting to see a culture shift where you can implement AI in more and more critical use cases because the employees using those systems understand that they have to do the work to verify the results.

The question is: What is the ratio of the time it took to verify versus the time it would have taken me to do it myself, and how much efficiency did I gain for that workflow?

We went from about 2 and a half years ago, when there was instant excitement—“Oh my God, this is going to be the greatest thing of all time”—to a reality check within 3 to 6 months. Everybody thought hallucinations were going to be the massive problem. Now, a couple of years later, we're seeing hallucination rates shrink, the quality of the outputs increase, and an understanding that you do have to review the work these AI agents are doing.

That takes on a different form depending on the use case. In coding, it means you just have to review the code in the same way you had to do anyway.

Which you had to do anyway. It seems people are forgetting.

You had to do it anyway, but there was probably at least a little bit of theory as to what part you should review with an extra level of detail because you knew the person you were working with.

It also implicitly limits the value of AI, which people are uncomfortable with. It basically says it helps people who know more than the AI does. As soon as it knows more than you do, it starts to bisect the utility.

Yes. Basically, it's super interesting: The productivity of an expert is now outpacing everything else.

Which was something we probably could have predicted based on historical events. I think you've got some good theories about the kinds of skills that make someone the right user for these models and these use cases.

We're seeing that with expert engineers. They're like, “I don't mind that it's a slot machine. I pull it and see what comes out, because I can still get 10x productivity. It gives me good ideas, and I get it good enough that it's worth the productivity gain.”

If you were not an expert engineer and you used this slot machine, you would probably try to deploy all the results that were also wrong.

And you don't actually know which lever to pull, which is a big thing—literally knowing what to ask for and what language to use. We'll get to a better—

I think this is an incredibly important point you're making, and 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 go to Home Depot—

And I'm like, “Dang, man. I would have a DeWalt, and I could buy it, but it's really not a particularly good idea,” right? I think that's how these platform shifts happen, and why there's so much excitement over coding. The best way for a platform shift to take hold is for the experts—the people closest to being experts in the new platform—to become the most enthusiastic.

And the biggest users overall. I've been practicing yoga at the Cubberley Community Center in Palo Alto because the studio is closed for remodeling. What's neat is that it was the original place for computer clubs in the late 1980s and early 1990s. If you ever wanted to meet the computer people, you would go there. It was like Halt and Catch Fire: a bunch of people with soldering irons and [expletive].

When something didn't work or was broken, it wasn't, “Oh, man, these things are terrible. I'm wasting all my time.”

That was the whole meeting: Who could get one of these new discrete graphics cards to actually work and debug the driver? Can anyone print? Is there anyone in this room who can print in this new thing called PostScript? I think that's what's really happening right now. First, it's obvious it should happen with development and coding, because they're the most forgiving.

Yeah, and the most understanding of what's a bug and what's something that can never get fixed. The thing to watch for is that no one is saying coding can't get fixed, right? Whatever it's been generating that's bad for a 2x coder rather than a 10x coder, no one is saying, "Well, that'll never be fixed." Right?

And then the next thing that's going to happen is what I think is just going to be the creation of work like the marketing document, the positioning document, and all of this long-form stuff. If you're really good at that job, you can ask the right questions, you know what looks good, and then you can get really domain-specific. The next level is, "Oh, I need to understand a competitor," which means using real information from the internet in real time, not just statistical information. Then you're like, "They already know what the competitor does."

Right?

And then my favorite scenario is the one that just constantly has these aha moments: "Attack this thing I just wrote."

I'm not interested in you adding em dashes and making it a little bit better. I just want to know, "What did I miss?" You said one recently, on this last one, about, "Here's my earnings statement."

Yeah. For people, that's the thing you read after the analyst. Now attack it like an analyst. There are like 6,000 hours per company of analyst questions. It knows what they're going to ask. They only ask like 3 questions anyway: expense line, you know? I feel like this is the thing that's—

Do not watch this if you're an analyst at this—

This is not any advice about being an analyst.

But this is what's really going to happen with writing. Then it's going to happen with PowerPoint and slides, and then it's going to happen with video.

But 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, you actually know all of these things; you just don't have the time to go through all of it, and you may not remember it. In a way, it's productivity-helpful, but it's not solving problems where you are a particular expert. This is maybe why, for those who are nonexperts, it's a little bit more threatening, because it can do that job.

Yeah. Well, maybe to bridge a few views and probably throw in a different tangent: Steven, you're asking, "Where is the enterprise now?" That was the coding piece. I think what you're seeing is a clear understanding that what I'm going to get out will be correlated with what I put in. How precisely I put in the prompt matters. Prompting doesn't go away anytime soon, simply because the leverage you get from the set of instructions you're going to give the AI at the start is still going to be massive.

What if prompting went away? What would you end up with?

Well, 2 years ago, I think people were saying, "You'll just tell the AGI what you want it to produce."

Oh, just one prompt. You unbox it and say, "Go do something. Be a software engineer."

No, literally, that was an open debate. It was like, "No, you're probably missing the fact that what's in my head is going to be unbelievably germane to the thing that I'm trying to produce, and 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 increasingly unhinged prompts, which are pages long. The output you're getting from that is actually way better than if you didn't give it that context. I think there's a clear understanding of that side of the enterprise use cases, and a clear understanding that you've got to go and review it.

We forget that formal languages came out of natural languages for a reason. We didn't start with formal English and say, "Oh, it's much easier to speak in English. Just speak in English." It's the opposite. We have this natural language, and we're like, "It's very tough to convey the information that I want to you. You and I are both experts, and we understand the solution space, so let's communicate more efficiently."

And that's what jargon is.

Of course.

Jargon is just a formalized way that people who have domain expertise talk to each other. The thing that's most fun to think about right now, at least—and maybe you could give us a little history lesson on this interesting parallel—is: When does the style of work change because of the tool, versus the tool being adapted to the style of work?

We're only on day 1 of this, but I'm starting to see some patterns emerge. We thought agents would learn how we work and then automate that. Agents would 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 to start with. People are saying, "I'm going to have agents and then subagents for parts of the codebase. I'm going to give them README files that the agents read, and then I'm going to optimize my codebase for the agent," as opposed to the other way around. In other forms of knowledge work, within how we use Box with our AI product, you're starting to see people tell the agent its complete job. The workflow is now starting to be almost like the agent is dictating the workflow in the future, as opposed to just mapping to the existing workflow.

I don't know what the history is on this—when the work pattern itself shifts because of what the technology is capable of. I think that's probably where this goes. It's not just that agents plop into how we currently do our work and then automate everything. I do think you start to change what the work actually is, and then agents go in and accelerate that.

Well, as important as that is, it's actually more important. What happens is that, to reuse the word in a different way, there's this anthropomorphization of work. The first tools actually anthropomorphize the work.

If you go back, this is true of 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 you look at cars, and until Elon got rid of all the controls, everybody kept all of the controls. I don't want to get into that fight. But what happened with every technology shift is that the first version of the tool always preserved the way that humans did the work.

If you look at what accounting software looked like in the 1960s, 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." 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 who have a manual process based on how far apart the desks were, right?

Everything that happened with the rise of PCs and personal productivity started off—and I always use this example because I've watched it happen like 5 times now—with the first PCs that did word processing. The biggest request was, "How do I fill in expense reports?"

A whole world grew up around tractor-fed paper that was preprinted with the expense report. Then we wrote all of this code: "Are you using an Avery 2942 expense report, or is it a New England Business Service A397?" You had adjustments in the print dialog, like 0.208 inches, and you moved little things around. Then you would print out, "Ate dinner, $22," and that was all you printed.

Then someone said, "You know, we could use the computer to actually print the whole thing," right? Fast-forward, and finally Concur said, "Why just take a picture? Why not just take a picture of the receipt, and then we could do all of it?"

The whole thing gets inverted, and every single business process ended up being like that. When email came along, it used to be that, to prepare an agenda for a meeting, somebody would open Word and type in all the things and then print it out. Everybody would show up at the meeting with this very well-formed agenda. Then email came out, and that whole use case for Word just evaporated.

Yeah. And then an email agenda became no formatting, nothing: “Here are the 8 things we’re going to talk about.” You show up, and everybody’s like, “Did you get the agenda?”

You know what’s interesting about the AI one? It’s kind of like we’re seeing the same thing, but vis-à-vis AI. Nobody really predicted the generative stuff, and we’ve had AI for a very long time. We’ve had chatbots, and so you had these AI-shaped holes in the enterprise for a long time.

A lot of the mistakes we see today are people taking the generative stuff and trying to cram it into the old models, when it’s really a new behavior that’s emerging and is much more individual. It used to be that you would centrally sell AI to some platform team, and 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 centralized.

A lot of the adoption we see is much more individual, for example. I just think there’s a bit of a mismatch as we’re seeing now, and it’s getting ironed out, too.

Well, and so I think the question is: Are we in the phase where we’re trying to graft agents and work into basically what we’ve been doing for 30 or 40 years of software?

And is this actually going to be the first real step-function shift we’ve seen in what the workflow itself should look like?

Oh, we are. I mean, if you remember, people like me tried to jam the internet into Office.

Right?

And it was fun to watch.

But you weren’t watching it. Everybody around was trying to jam the internet into their product because that was the only way you could envision it. 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. Other people would say, “Let me just try to implement Excel using these 7 HTML tags with no script.” That turned out not to be a really good idea, either.

The best was, “Let’s do PowerPoint.” Well, how do you do it? You give them 5 edit controls, tell them their bullet points, and then generate a GIF on the back end and send it back to you as the slide. Yeah, okay. That was not a good idea, and so there was that whole—

I think, actually, maybe the main point is just the durability of Office. It transcends all—

Disruptions. I like to think it pretty much rises above everything. But the thing is, that’s where we are now. Everybody—

And, you know—

Do you think—

I mean, just to dig a little bit: Do you think this is similar to the internet, in that it’s a consumption-layer change?

I always viewed the internet as very much a consumption-layer change. I go to the internet instead of going to my computer, but otherwise things kind of stayed the same. AI has this weird quirk because, for the first time I can recall, programs are abdicating logic to a third party.

We’ve always abdicated resources. We’d say, “Okay, I’ll use your disks,” or whatever, 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 these programs are also no longer being written by a human in the same way. I feel like the change is maybe a bit more sophisticated.

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 what that also does is change what you write the programs to.

Do you remember ever abdicating logic?

Oh, here’s a great example of how disruptive this can be. The first word processors in the DOS era—the character-mode 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 have a word processor; 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 WordPerfect, your ads said, “We support 1,700 printers.” You won reviews because you had 7,800 and Microsoft had 1,200.

That’s so—

And so then along comes—

That’s a great one.

And then Windows comes along. If you were trying to enter the word-processing business, step 1 was, “I need to hire a team of 17 people to build device drivers for Epson, Okidata, and Canon printers,” because you couldn’t get them anywhere.

Microsoft came along and built print drivers and a clipboard for Windows. Macintosh did it, too. All of a sudden, there was a way for 2 applications that had no a priori knowledge of each other to—

Good.

But of course, if you were WordPerfect or Lotus, that was a disadvantage. You got creamed by that because of your ability to control your information.

What happened was a bunch of developers were like, “Wow, this is cool, because now—when we did C++ for Windows, we were like, ‘Where’s the demo?’ In fact, at that Cubberley Community Center, I would go and show brand-new Windows programmers in 1990, ‘Hey, you don’t have to write print drivers and use the clipboard,’ and there was literally a standing ovation of 10 people at the thing. They were more than happy to let data interchange between products because they were like, ‘That’s nothing but opportunity for me.’

Can you imagine? They probably, from an emotional standpoint, felt exactly the same way a vibe coder does today: You’ve just given me this platform that—

The Programming Windows book was this big, but writing a device driver for an Epson printer was this big. Writing one for a Canon printer was this big.

But I’m actually trying to think of that. The paradigm shift is the same: There have been many times when we’ve reduced the amount of work a developer takes, but I just don’t remember a time when the programmer abdicates logic. For example—

SDN didn’t.

Not logic. I would always say what is correct and what’s not correct, right?

I think you undersold it, though.

No, this is the thing, by the way: Everything that Martin invented and worked on—that stuff is a big deal.

Maybe we should postmortem your pitch at the time. If you mean—well, no, let me focus on logic specifically.

I’m writing an app. My app is some vertical SaaS app for a certain customer base. The answer the app gives is based on logic that I’ve written historically, right? 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 device resources.

But if I’m saying, “Hey, large model, tell me the answer here,” I’m actually abdicating application logic. Maybe you’re right. Maybe this—

I think you’re almost playing incumbent, in the sense of trying to decide, “This is abdicating logic and this isn’t,” when it really was a huge competitive advantage for WordPerfect. They didn’t want to give it up, and they fought against it.

In Windows or on a Mac, you could rasterize anything you wanted. You wanted a button that you pushed and it spun and animated like a rainbow? You could do that in your product. But then the web came along, and you were like, “Wow, I have to use a gray button that says ‘Submit’—”

And that was like—

Yeah, because the point is, we do use a bunch of third-party things.

Well, but it took a long time for those to show up. Early on the internet, magazines in particular and the printed media were the ones who absolutely wouldn’t go to the internet because they would not give up their ability to format.

And 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.

Right?

All it is is 7,000 commands for how to do kerning, bold, and italic. It turns out AI not only doesn’t care; you could ask it to make whatever you want. You could say, “I’d like this to be a double-index pie-chart thing.” That’s not a thing, but it will figure out something that looks like that, and you’ll go, “Cool.”

So interesting. In this case, the abdication is actually figuring out the way to visually represent the data, which is absolutely—

Right.

And it turns out that 90% of people never really got to be experts at doing that task, even though 90% of the tool is about doing it. So what happens is that the programmer didn't abdicate the logic in this case; this is the user. What's the user? What's the programmer in that?

In fact, what the programmer was doing was inventing the thing called wizards, or whatever, and that would make a whole bunch of choices for you—stylesheets 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 it all. There's some Steve Jobs quote that he loves about Schopenhauer: If you've seen the conjurer, it's not a trick anymore. I really feel like this is the 3rd or 4th time that this has happened just in my lifetime of watching this.

Well, 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—code agents—and they're interfacing at the GitHub PR level, right? It's not obvious to me why you do a bunch as opposed to 1, 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, and I would love your thoughts.

Well, my read on it—and then I guess I would throw out what happens next as a result of this, because to me it's actually a little bit of an epiphany on what the future work design could look like in this world—is that engineers, back to the prior conversation, are just the first to experience this. But I think my read from talking to similar folks who are all in on this is this mix of, basically, the context-rot problem, which is that the more that we put in the context window, the more it gets confused and the lossier the answers get.

And so you have to have some kind of way to partition what an agent should work on. We see this in building agents internally. The panacea that I think we maybe would have hoped for is, “Well, you just put 1,000,000 tokens into the context window and then obviously—”

Oh, so you're saying this is almost like a counter-trend to AGI. It's almost like the opposite.

It's the opposite, but it only works because the models are so good.

Yeah. But you're giving more things more-specific tasks rather than 1 thing less-specific tasks, right?

Right. And so I think this is why it's happening. Basically, the craziest version of this is I was talking to somebody who is in startup land, and they have—to your point—all these sub-agents. But what's amazing is it maps 1-to-1 to each microservice in their codebase. They have an agent per microservice, they have effectively a README for the agent, and that agent owns the microservice.

I don't know the specific number, but let's just say you could have dozens or hundreds of these things going on. You're effectively mitigating this issue, which, if you just said, “Here's my entire codebase—go run wild” to 1 agent, it will just produce worse and worse code over time because it's going to have context rot. It's not going to know exactly what you're trying to do in that 1 area of the microservice, but the sub-agent model seems to be working for that paradigm.

I love this counter-pattern because everybody is like, “Models will get smarter and you'll give them higher-level tasks and they'll do things longer.” Yes, this is a counter-pattern. I want to tweet that, but you have more Twitter followers. We can collectively do it.

But then the question is: let's just assume this works in engineering. You have this interesting dynamic, which is, well, then that means that some of the coding practices will be pretty different in the future. We've talked about this idea that the individual engineer becomes the manager of agents, so that was already, I think, a well-understood path. This is a supercharger of that concept.

And then the question is, how does that translate to almost every form of work? If I am now the lawyer working on cases, and I can have 20 sub-agents that each do a different case and then basically come back in some kind of task queue that I'm going through, obviously, the sheer leverage you get is going to be insane. But I do think the way that you might even organize the work, and what the workflows within an organization are, are inevitably going to change as a result of that.

Oh, but I think this just gets to the fact that the flow in the workflow has been serialized or linearized, based sometimes on knowledge but other times on tooling. So what happens when the tooling changes is you just get this realignment of what's truly serial and what's not. If you're planning an event for a company—which is still going to keep happening—you might say, “I have to book the venue. I have to invite all these people. We have to create all these materials.” Well, they're actually not particularly gated on each other, right?

But if you have an events person—

They're gated—

Right. And so now an events person can start spinning up all of these different elements, and then they're going to come back: “I've gotten as far as I can on collateral until I get a logo for this event,” right? “I've gotten as far as I can on invites until I get the date and the time and the venue.”

And I think there's no reason why you can't spin up all those in parallel, because, of course, how does that happen today? Well, if you're a company and you use Box, and you've done this—this is your 58th event—you have a folder called “Event,” right? And people take the folder and go “Event 59,” and they make a copy of it and all the stuff in it.

And if you think about that workflow, that's exactly what a series of different background tasks or agents could go do. And so I think the reason that you could be doing all that in coding is, well, there's a natural way to break that up because there's a bunch of—

But there's the other side: there's also a bit of an indictment on your ability to give it a high-level direction. It kind of suggests that the human being needs to be giving them more granular orders. Otherwise, to start a company, you'd issue 1 prompt, you'd go to the beach for 6 months, come back, and you'd have a full—

Company.

Which is almost a re-anthropomorphizing effect. It turns out we did kind of figure out division of labor. We figured it out in the context of a lot of physical, analog limits that we clearly had, that agents won't have, but now there's no total free lunch. So you have this context-rot issue, which is that you do actually have to subdivide the tasks at some point. So then the question is, what are the right—

I mean, it may not be a context-rot issue. 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.

And 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.

Exactly.

And that just 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, right? It's basically a different way of thinking about templates or a different way of thinking about starting artifacts or scoping the context in a generic world.

Well, but then there's this: it might, though, be the right architecture in general if you assume that we're never going to get to a point where the model is just 100% perfect, right? It might also be the right architecture design because at some point you don't want an agent or a set of agents to go so far down a path when there was a step that it needed to check in with you on, because there's just the compounding effect of that.

So you do need to subdivide the work also because if you do have gating moments that are going to have a bunch of dependencies, the agent does need to know: at what point should I roll that back up to the user?

Yeah. Again, against the common narrative, now that I think about it, it seems that the trend is that prompts are getting more complex, not less, right? We're seeing more agents, not less, doing more narrow tasks, which is almost this kind of counter-AGI narrative. It's almost like these are much more specialized and go much deeper, working with much more specific instructions.

And there's a sort of a history of this: Wow, maybe we can actually solve it if we're specialized.

Yeah, a little bit more. If you take 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 like, “This is an expert system in deciding on what type of infectious disease, as long as you have 1 of these 7.”

No, literally, there was a paper about a digestive disorder that was actually a medical expert system.

I do, but I want to point out 1 big difference: somehow, 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, all on a computer, too. But let me try to demonstrate this with an old PC example. Early in the PC era, there were word processors, spreadsheets, graphics, and databases, and a lot of people were like, “Why are there these 4 programs? There should only be 1 program.” My answer to that, which often involved screaming, was, “Have you been to an office supply store?” Because if you go to an office supply store, there’s paper with numbers, blank rectangles of paper, and transparency paper.

This has been around for a really long time. There’s some reason that these are different human contexts.

How many minutes did it take you to know Google Wave wasn’t going to work?

Zero.

Okay. Okay.

It was instantaneous. No, I mean, this was the thing: there was an ancient Mac product, lauded by the industry, called ClarisWorks. It was like, “Oh, 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 just want as many cells as they could possibly have, and you’re sitting there saying it has to fit on an 8½ × 11 sheet of paper on a Mac.

I think that one of the things that happens is that these lenses that humans bring to specialization really matter. If you think about the medical profession and going from a GP to a radiologist to a specialist to a nurse practitioner through the whole series, they’re each going to look at and use AI in a different way.

So the only question would be: that level of specialization and division of labor emerged over a 100-year period, alongside tools, but also driven by a lot of the physical constraints and realities of how organizations emerge. In a post-agent world, 10 years from now, 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, and then is there just a new set of roles?

Clearly, there’s a role emerging in a bunch of organizations where someone says, “No, my role is that I’m the AI productivity person. I have a way of creating all new forms of productivity in the organization with AI.” Clearly, we’ll have a bunch of new roles, but is our current division of labor also going to collapse in some interesting ways because of AI?

I think that if you actually 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 in the interim.

You think AI causes more specialization over time?

Absolutely, because every human is going to be way better.

More knowledge will accumulate. I think this is something that has really happened with computing that people forget. There used to just be this morass of marketing and R&D.

There used to just be coding, and then there was coding and testing and design and product management and program management, usability and research, and all of these specialties. All of them had their own tools.

Go to a construction site. I remember growing up: our neighbors built a house. We lived in an apartment, and they built a house. There was Clem, the carpenter.

You built a house with a guy named Clem who used all the tools and everything. Now, when you build a house, it’s a 20-person list of subcontractors, all of whom have whole companies that do nothing but put in pavers.

Right.

There’s been a long disaggregation in the history of IT. Everything used to be in the same piece of sheet metal. Then you disaggregated the OS and the hardware, and then you disaggregated the apps.

In the last 15 years, we saw the app and independent functions get disaggregated. Almost everything became an API. An API would become a company: Twilio became a company, Auth0 became a company, PubNub became a company, and so on. It may very well be the case that every agent becomes a whole new vertical and a whole new specialization. Then you can actually build a company around it.

Today, just like with APIs, 1 company may have a whole bunch of agents. In the future, a third party may provide that agent as an independent business.

The opportunity, to your point, is really there for that.

The impediment to creating a company and distributing software used to be infinite.

No, exactly. It used to be ridiculous to think that a single API like Auth0 could become a company. Then, of course, it became one.

Or it used to be ridiculous to think you could build a whole company out of signing documents.

Right?

Not just a whole company. All of a sudden, you realize, “Wow, the addressable market for that is huge,” and it’s way bigger than signing because of all the things that got done that were baked into a company, causing headcount and waste and fraud and abuse. I think you can underwrite thousands of these companies emerging.

Jared Friedman had a tweet about this: go deep on a workflow, take the job of some part of the economy—a payroll specialist, for example—and then build an agent for that.

It’s not obvious that there aren’t literally 1,000 of those, for every vertical and every line of every department.

I just love this because this is literally the anti-AGI, basically following the long arc of computer science: as the market grows, the level of granularity at which you can create a company—

It’s also economic growth. Take that example. Today, just like Salesforce, which is always my favorite example, the idea of having a productive sales force used to be a consultancy.

Right.

The only way you could ever fix it was by hiring a consultancy to show up, analyze what everybody does, and then do a report that says, “This is how you need to reorganize.” It usually meant doing the opposite of whatever you had, and then they would leave.

People tried, but there was no cloud. To build a CRM, you had to do all that consulting work and then roll it out. It was static, and you couldn’t maintain it. Then all of a sudden, there was Mark Benioff and a whole way to do all this. Not only that, people actually liked it. They thought they were better at selling because they were using their phones and putting in a few notes about a client, which helped everybody.

I think that’s what’s really going to happen with all this. Suddenly, something that looks really small becomes a whole thing because there’s no problem with distribution. There’s no problem with customization. We’ll actually have ways to solve security and privacy, just like we solved reliability and things like that.

Look at the stuff that you’re a world expert in—the stack of internet networking technologies. If you had asked me 15 years ago whether CDNs would be companies, I never would have believed it. I would have said, “That doesn’t make any sense. How could you have a company that’s a cache?”

Yeah.

I think people are probably way too afraid of the model providers eating them. I think that was basically a phenomenon in the first wave. If you were just doing something basic—if you had figured out that you could do something on GPT-2 and GPT-3 where it was a text interface that produced more text—then 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 will be eaten if you’re doing AI for health care, AI for life sciences, AI for financial services, or AI for coding at the right parts of the stack. AI for coding may be the 1 asterisk area that will be hypercompetitive, simply because the model companies don’t want to use somebody else’s product to build their own models.

AI is kind of the exception. We’re in a 5-year period right now where you’re going to have to build agents for every vertical and every domain, and there’s a playbook starting to emerge for what that needs to look like.

I think there was a technical head fake that happened early on, which was pretraining. Pretraining really was a 10-out-of-10 technical innovation. I can’t tell you how many times, 2 years ago, somebody would say, “I have a friend who’s building their own imaging model, a post-training image model. We’re going to make it so good at imaging.” This is a text-to-image model, and they wanted to make it so old people looked really good in it. Then, of course, the next version of Midjourney or whatever comes out, and it does a better job of it.

The thing with pretraining was that you were just consuming all of the world’s existing data, draining all of that energy, and it generalized perfectly.

But it feels like, technically, that’s passed. Now we’re more in post-training and reinforcement learning, which is a lot more domain-specific.

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? It’s going to be about applied companies.

Yeah. If we had an infinite number of tokens, then the models would just continue to generalize, but it’s pretty clear that’s not happening. So now we’re going into what we all understand very well: companies have to choose which domains to go into, and they’ve got to solve the long-tail problems there and get access to the data, et cetera. I also think there’s the shadow—the shadow cast by large companies—

Oh, “We’re going to put you out of business and stomp you.” It’s ridiculous.

And it has never, in any technology wave, lived up to the fear that people have. Look, if you built a new word processor in 1995, you were an idiot.

That was not the thing to go build.

Yeah, there was a time, just 10 years earlier, when 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.

And so the thing that’s not being said now—which we should do a whole one on—is: what is the actual platform?

It’s all well and good to say that the large models will subsume every application. The thing is, the minute they start doing that, no one will be in their platform, right? No developer is going to sit around and say, “If you’re going to subsume me,” right?

There’s a phrase for this in the Mac and Apple world: Sherlocking. So it does have a real chilling effect, and that’s one of the things all the model people are going to learn very quickly.

There’s a chilling effect, but there’s also just—I think there really is a problem: it’s hard to go deep in 50 categories. You just can’t do that with pre-training. I think everybody is scared because pre-training was actually the one thing that was good at that, and now they have to actually choose. Yeah, I agree.

You do have to go deep. 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.

No, it’s ridiculous. In fact, it’s only good because what happens is that the big company raises awareness of a whole category, and then you just swoop in and you go—

You’re just a feature to them.

You’re just a feature, right?

But to you, this is your whole life, right?

And you’re going to win. Look, I always come back to this: there’s a whole company that just signs things, right? I can’t believe there’s a whole company that just signs things. I have so much to say about this topic.

Even minimally, if you graph the willingness to pay for an inference versus the cost to serve it, something like 20% of the inferences are 80% of the cost for most companies, for most spaces. So, actually, the problem of the application is just to choose those ones, which tend to be more domain-specific. Yeah, this is the problem of inviting the 3 of us on here, which is—

Just getting us to shut up is the trick. Yeah, guys, thank you so much for coming on. This is fantastic.