Can we find a way to actually tell the optimistic story of AI, which is yes, in some areas of the organization, we're gonna slow down headcount growth. In other areas of the organization, maybe there'll be fewer people at the end of this, but in other areas, we're actually gonna be growing and we're gonna be reinvesting and reallocating dollars.
I wanna talk about agents because you have very kind of loudly and prominently declared that we are entering the agentic era.
Yeah. I feel like that's like a well understood theme at this point in this, basically--
Well, there's a lot of talk about it.
But eras are a big thing. I mean, Paleolithic, Neolithic, Iron Age. Why do agents deserve their own era?
Yeah, I mean I think we'll figure out if in time it's at that level of Paleolithic, but I mean, I think it's pretty obvious that there's a massive transformation in the world of work. And basically it's the shift of software has always been deterministic. So it kind of does exactly what you want it to do every single time. And that has a lot of great properties, but it also has a lot of limitations. It can't adapt to new information. It can't adapt to any amount of kind of unstructured data and input that it receives. It has to be highly, highly structured. And so we kind of were able to automate 5% of a company's operations. We can automate the ERP operation. We can automate sort of managing workflows around our CRM or kind of IT workflows. But actually most of work is this sort of messy, unstructured, human, collaborative, lots of different kind of data types coming in. And we've never been able to automate any of that. And so if you just did a heat map of all the work that happens in the world, the vast, vast, vast majority, probably like 90, 95% of the work that we do is inherently unstructured. And for the first time ever, AI allows us to bring automation to that part of the work.
And so it just stands to reason that obviously that is the biggest shift that we'll ever see in kind of corporate work,
Bigger than previous ways of automation.
Easily, by orders of magnitude. And so it sort of certainly deserves its own way of kind of thinking about what that future of the information age is gonna look like. It's actually interesting, we probably called the information age too early in retrospect. We thought computers would bring about the information age, and it kind of did, but it meant that people were always the dependency for actually working with the information. And now with agents, agents can work 24/7 in parallel, on any amount of data. So that really starts to usher in the full potential of the information agent.
Can you do a quick definition? Nickle LaMoreaux was on the stage earlier. She said 95% of what people say is agents right now.
Is it agents? Would you agree with that?
You guys can fight that one out later. Yeah.
I mean, I think there's gonna be a lot of that. This is gonna be very messy terminology wise, just for a variety of reasons. But like by virtue of me having the microphone right now, I'll make the real definition of agents. But I think everybody can kind of feel it, like when you go do a chatbot and you ask a question, you get an answer back, it's sort of, 'Okay, that's kind of like an assistant chatbot thing.' I think we are collectively as an industry agreeing that an agent is probably something that can do a bit more than that. And so the general architecture that you should expect from any kind of agentic system is it's still an LLM by and large behind the scenes. But that LLM or AI model has access to a set of tools. Those tools could be your software. They could be specific, very narrow, just one-off kind of tasks that they can do. And you can run that agent, effectively LLM, for a minute, 10 minutes, an hour, 10 hours. It can take in various inputs from your systems and make decisions on its own. And so essentially there's some element of both autonomy and this ability to do kind of continuous work.
And that makes it so you're not dependent on your ability to go back and forth with the assistant. You're able to give it a task, it can go off across your systems or to some extent even into the outside world, and then go and complete that task for you. And that's probably the new definition that I think we're all converting.
And so there's something I find very confusing in this conversation because on the one hand, you just said it.
You're gonna have this agent that can go out and make decisions on its own. On the other hand, everybody says, 'Oh, but there's always gonna be a human in the loop.' Well, if the agent is making decisions on its own, there's not a human in the loop. So sort of help us through that.
Yeah, this is actually one of the, I would say unresolved issues of all of AI, which is I think we all agree that at some point you need some degree of accountability. The agent can basically never be accountable because it has no real tension in its system about making the wrong mistake. It has a reward function when it's being trained. So it wants to--
Accountability has to lie with human beings.
I mean, all of our laws are set up to require that. So at the end of the day, I can't sue Anthropic when it does something wrong. Like, the Ts and Cs of Anthropic are impenetrable. Like, you will never get a dollar out of Anthropic for it.
Now, not to digress too much but does that mean you agree with Anthropic in its dispute with the Defense Department?
That would be a digression. But why don't we come back to that? But like, that's a whole thing I'd have to unpack.
But it's basically about accountability, right? To say can you have truly autonomous weapons?
Yeah, yeah, I'm just trying to think of like, who are you assuming needs the accountability in that model? But let's just let's--
Okay, so we'll park that. I'm sorry.
More pedestrian, non-nuclear missile use cases. I can't hold an agent accountable for a hiring decision, for writing code that's buggy, for making a loan that was wrong. You can only hold either the institution or the ultimate individual that made that decision. So we do have a paradox. This is an unresolved paradox of AI, which is as an agent does more and more work in the background, that's how you get productivity gains. If it's just going back and forth and you're constrained by your ability to go back and forth to the agent, then there's limited productivity upside. It's like you'll get like 20% productivity upside or something like that. If an agent can go off and actually do like real hours of work in the equivalent of hours of work in minutes, now you're talking could be 2x, 5x, 10x productivity gain. The challenge is it's gonna make a series of decisions that are gonna be harder and harder to kind of parse all the little tiny elements that it used across that decision. And so the human in the loop element, we have greater degrees of abstraction of what the human in the loop is able to really process. And so the most extreme case that's happening in Silicon Valley right now in engineering, you can kind of look to engineering as effectively the benchmark of like where AI is going. Now, there's gonna be a longer gap as... It's actually gonna be interesting, the faster you see things change in engineering, you'll actually see a wider gap between the rest of the world and what it's gonna take to get the rest of the world to be able to use agents like engineering. So it's not like these are gonna be fast follows and the rest of knowledge work. And we can get into why if you want. But if you look at engineering, about a year ago, let's say, it was like completely the best practice, everybody agreed you have to review the code that the agent did. And then we said, 'Oh, actually, maybe what we'll do is we'll have an agent review the code that another agent wrote. And then you'll just review the final review of that agent that did the review.' And now at the real, real state of the art of like the fastest moving startups are kind of in the position of saying like, 'If the thing tests, and is verified as test complete, like we're gonna ship it.' Because a human can no longer look through 50,000 lines of code.
And do you think they've gone too far or you think that's okay?
Depends on what kind of system you're working on. And now this might get a little closer to the Pentagon thing. Like, I actually wouldn't do that type of approach for missiles. Would I do that for an internal form where you had to build a quick application for some employee experience thing? A lot probably safer. A lot riskier if it's a trading platform, if it's a healthcare system, if it's a government weapons system, lot harder to do that.
And the truth is we probably make those kinds of decisions all the time. How closely do you read that contract? How closely do you read that language?
But the big difference is I can still hold you accountable for not reading that contract.
If you didn't do it, you're accountable.
And so the fun paradox, this is where the tension's gonna get and you all are gonna be like at the center of this. It's gonna be hilarious.
They're having a great time.
So we are going to be like full throttle, we have to go AI first and everything. At the exact same time, when AI takes down your website, you're gonna be like, 'Well, like who do we blame?' And eventually you're gonna find you probably have to blame the engineer that used AI to--
That went full throttle. But then they're gonna say, 'You were the one that told me to go full throttle.' And we're gonna have this loop for a while. And this is why there's no free lunch in automation.
It's gonna be a really fun times.
Yeah, and so you're gonna say, 'Shoot, you were reckless in using agents to write that thing that took down the site, but we did technically tell you to do that.' And so is that a good performance review or a bad one?
And you're saying this is just how we learn.
I don't even know what the end state is. Nobody does.
Yeah, but it's gonna happen.
It's gonna happen because the productivity gains are simply too significant by doing so. But we do have a bunch of really interesting new problems that we as organizations and sort of best practices haven't yet resolved. I mean, there's gonna be even like new forms of insurance. And the new form of cyber insurance will be agent insurance and you'll basically say, 'I'm gonna buy a billion dollars of protection if agents do something' and so the economy will find interesting clever ways, to kind of figure it out.
So just throw it forward three, four, five years. What is work gonna look like?
Yeah, I mean, I think, again, if you look at the microcosm of engineering, it gives you the best hint again about what the future's probably going to look like with, again, massive caveats. But in general, and I think we're all kind of collectively getting the sense, we should anticipate that the knowledge worker of today is sort of the manager of agents of the future. So the individual contributor becomes the manager. What they are managing is a set of AI agents to help them sort of execute more work than they could have before. And so you can kind of just imagine and posit what that will be for every role in a business.
The brand... You think most people will be just managing agents, not so much people?
Well, I guess at some point there'll be a person managing people managing agents. But even there's an interesting tendency now in engineering where you probably want your frontline managers also to be practitioners more, partly because there's a fluency in using these tools is really important for them. And also because it's just like, once you have a really strong... Like, we all do this, which is like you take your really good engineer and you turn 'em into a manager, but then all of a sudden that really good engineer no longer is producing any code and it's like, 'That's too bad. Like, that person was actually an expert in what they were doing.' And now AI kind of gives them the ability to probably still have a player-coach type dynamic where they can be doing more and more output as well individually. So that's probably important because it keeps them fluent with the technology and they're not just in meetings with people, but they're actually using the tools themselves. So I think you're gonna see that as a practice again, commonly in engineering but that will ripple through other areas as well.
I wanna pivot here and talk about you. I'm glad you talk fast by the way 'cause we have a limited amount of time and getting a lot of information into this. But I wanna talk about you because we're using this interview for the Leadership podcast.
And I have to ask you some questions about your leadership. You started Box.
I liked the agent stuff better.
Maybe we'll come back to it if we have time. But you started your company, you were at USC, you were 19 years old. USC, I guess everybody else was partying. You were starting a company.
There were a couple of us. There were a couple people not partying, yes.
And you left. But what drove you to do that?
Well, basically, I had been trying lots of startup ideas many times throughout high school and college, and most of 'em just didn't work, obviously. And this was the first one that did work. And it kind of worked in a way that was... The first time, like there was a flywheel where we would build a feature, we'd get more customers, we'd do more, we talked to more customers and it kind of kept kind of compounding.
It's just a place to put your content and organize.
Yeah, I mean the idea was like quite simple. It just like everybody had been sitting on it and not actually executing it, which is like, the internet can store data now. You wanna be able to access your data from anywhere. So we were like, well, let's just put two and two together and like build a product that does that. So we kind of got in pretty early and we eventually pivoted in the enterprise, but we sort of started as a broad-based tool for anybody to use it for sharing and collaborating on data.
I wanna talk about the pivot to the enterprise 'cause that happened, what, 2007, 2008?
At a time when enterprise business was not cool.
I mean, I have a very clear memory of Steve Jobs coming to The Wall Street Journal with the first model of the iPhone.
And we were all sitting there with our BlackBerries in the room and he passed these iPhones around and people played with them a little bit. And somebody in the room, one of the editors said, 'This isn't gonna work for the enterprise.'
And he just stared 'em down and said, 'I don't give a.' I'm not gonna repeat the word he used but it starts with an F. 'I don't give an F about the enterprise.' And that was kind of the whole Silicon Valley jive. So what drove you to do what everybody else thought was uncool?
