The Half Life of Agent Infrastructure — Ben Kus, Box

Hi everyone, I'm Ben Kus. I'm CTO of Box, and today I'm going to be talking about building for change, and specifically around AI agents and how to continue to adapt your infrastructure as we are all in the middle of this journey.

So before I get too far, I will quickly sort of set a little bit of, like, who I am and sort of what I do. For Box, I'm CTO, and one of my jobs and my job for my whole career has been to build enterprise software.

And so if today you're from a consumer company or you're not involved in enterprise, I hope that a lot of it is still relevant, but in many cases a lot of the lessons I've learned are enterprise-specific. So I sort of will highlight when I'm thinking and talking about infrastructure, when I'm talking about the kind of challenges that we face, I'm typically talking about things that are sort of in a scale of, like, for Box we have over an exabyte of data.

So not a gigabyte, not a terabyte, not a petabyte, but an exabyte. And then oftentimes I'm thinking in the tens of millions of users, the hundreds of billions of things, in our case files or content or unstructured content.

And then the new stat that is sort of the one that we talk about is tokens. So we are now in the ballpark of trillion tokens, probably will be 10 trillion tokens sometime soon, and this, of course, is some of the new and interesting challenges that this kind of scale brings.

So in my career, and I think maybe many of us here, we've kind of lived through these technology changes. And so taking a quick step back, I started my career when the internet was sort of becoming a thing, lived through mobile and sort of this idea of, like, you know, carrying these different devices, moved to the cloud where you could kind of store and maintain all your data, and then of course we're all in the middle of this AI change.

And I think when you see these kind of technology disruptions, when you're sort of thinking about this idea of, like, all of these kind of have changed all of our lives, and then you're thinking about it from the perspective of a technology leader or a startup or an engineer, you kind of see that, like, these are where, like, major companies are born.

You see that, like, big companies adapt or die. Small companies are here to disrupt things. They're here to take bets. They're here to grow. I've had two startups. I've been acquired twice. Once in IBM, once in a Box.

And so we're in the middle of this kind of major opportunity. But for a long time, no matter what you kind of come from and what area you're at, I typically, if you were to ask my advice in about kind of what makes you successful as a company, as an engineering organization, as a technology startup, or as, like, a company who has a technology division, I would say no matter what, there's kind of three things.

The first is you need to build scalable, reliable platforms and select the technology that you care about. Meaning that, like, there's a lot of ways to do things, but get good at something. Get good at that technology, at that system, at that stack, and then keep going with that.

And then you leverage this technology so that you do more for your customers, you build your better product, you develop the capabilities, and then you optimize it. Make it better, make it faster, make it cheaper, make it more capable.

And this was sort of the generic enterprise advice, the generic engineering advice that many, many people would follow, and I think this works really well. Except now,

I don't know if this is good advice. It has been across all these major disruptive changes over time. Unclear. In fact, I don't think it's good adviceright now because there's a funny thing happeningright now, which didn't happen in those previous trends, which is that the rate of change is dramatically higher.

And you might say, look, technology always is changing. Like, you know, in those other trends, things change a lot, but not that much. The internet is still based on HTTP. The mobile devices are still iOS and Android-based and so on.

And so, but nowadays, other than the fact that, like, generative AI exists, most things that power it are changing and changing dramatically.

So last year, I was here at this at the AI Engineering World Fair, and I gave a speech, and I said, after spending a lot of time on this and thinking through this, I think there's a key to this, which is agentic, a graph based approach.

The idea was you have a large language model and these nodes, and then you sort of put them together, and you have this sort of the AI traverse this graph that you set up, you build the graph. This is the key approach that, and if you use this, this is going to really help you sort of build agents because what is anything that we do in life?

It's a workflow. It's a wave. And then if you have an intelligent agent, it can basically traverse this. This is the key approach. And I believed that at the time, and a lot of people did. And I still love this approach.

But nowadays, it's sort of a little bit out of date. In fact, I remember a guy came up to me after my speech last time, and he was like, the problems you talked about, the answer is just this exactly you were speaking to me.

Thank you so much. And I was happy. I, you know, I gave a good speech and gave somebody some good advice. And then I remember when we made a change, it was like, I wonder what happened to that guy.

I wonder if he's here. So, but the problem is, is that not that it was wrong that that was the best approach, but a new way emerged. In fact, I started to, like, look through, like, all of the last year's events and actually go to other conferences.

Like, what did people talk about a year ago? And most of them were, again, nothing was wrong. All good speeches, all good ideas, but most of them have now a better way. And so, and this is sort of the gist of the challenge.

So if you look at our journey of the technologies, the kind of things that we care about, I'll just kind of rapid-fire here. Like, so let's say that you want to utilize AI models, and let's just look at the last couple of years.

A while ago, probably distant memory now, like, people would say train your models, or maybe fine-tune them. And you're like, nah, that doesn't, that's too, why bother? Just use a frontier model. Use something from OpenAI. Use something from Anthropic.

Use something from Gemini. And then that's great, but it's kind of expensive. Okay, great. Let's just use open weight models. They're pretty close. You can use them. You can host them yourself. You can get some good GPUs. But then some companies will come to you and they'll be like, look, we just did this big deal with OpenAI or Anthropic.

Like, can we use our own key or bring our own model? Like, sure, we can do that too. But then nowadays, probably the best approach is to do an adaptive model selection where you basically are picking the big and smaller models.

What which model does well? And this is kind of the cool new thing. Maybe I could give a talk on that. Or let's say you're building agents. Like, we use a lot of things here. Like, I mean, the word agent hasn't really been around for that long, but in that time, it used to be like a single shot LLM response.

Call that an agent if you feel like it. Then you have to say, no, okay, we're going to use chain of thought reasoning. Now we're going to make a graph based agent system, like I presented last year. No, but then it turns out why are you bothered to make graphs when you could actually have an agent just figure out what to do, make a plan.

That's the new approach. That's kind of the way that Claude sort of laid the approach there. And then you're like, okay. And then now it's like, well, if you want to use dedicated sub-agents, maybe, but then maybe why not just make a generic agent and have it recursively work and then give it skills.

Skills are very generic. They're super helpful. And then maybe now it's maybe the idea is not just to do that, but to do it with an agent sandbox. So the agent can write code and execute it because that's super useful because agents are great programmers and I have them sort of just live in their own computer.

And then arguably now that's the best approach. Or maybe even we're in the world now of, like, don't even bother with any of that. Just bring your own harness. Like, let people select if they want to use one of these other systems.

And then, you know, not even just building agents, but the technology around content retrieval, things like, you know, in the old world we were like BM25 and keyword search. That's the way to do things, but that's like distant memory.

Obviously, the future is retrieval augmented generation, embeddings, approximate nearest neighbor. That's how we're going to find data. Turns out that doesn't really scale well and it kind of almost mimics randomness as you keep going. So then maybe it's about graphs.

It's difficult to get working well. It's probably not the best. So then it's about hybrid. You want a lexical and you want to do a semantic search and rank fuse those together. Arguably not. Arguably, agents are actually way better at finding data because they can find things and apply their intelligence to get to it.

So each of these things I just mentioned is arguably the leading approach for that moment over time. If you asked me to give a speechright now on any one of these, I would pick the last. I'd pick the adaptive models with dedicated RLM style agent and agentic search powered by hybrid.

But is this the end of this journey? This is not that long of that time here. And so my guess is the stuff that you're learning today likely won't last that long. Not that it's not wrong, not that it is not the best answerright now, but probably something's going to change.

The big thing that changed last year was, in my mind, Opus 4.0 to Opus 4.5. When you did that, suddenly you got to a model that could do instruction following in really high scale. This is kind of, to me, the beginning, the epoch of, like, the new agent models.

Also, hardware is getting better, faster, cheaper. Maybe we'll start to use more tokens. Like, token usage is off the charts, of course. And is that good or bad? What's going to change there? Enterprises are adopting things differently. Whether or not a company has decided to go all in on one agent to rule them all, sort of like Claude or maybe Codex style agent, or maybe they want to utilize agents from different platforms and different systems or both.

This is going to affect your lives. In addition to things like just the new techniques, new interesting approaches, new technology to power these things. So the fact that everybody is working so hard on this, trillions of dollars investment is actually leading to a lot of this change.

And again, it's happening way faster than I've ever seen, for sure.

Now, if you look back, it's not this way with everything else. Like, if you go see some of these other discussions, like I've given a speech on some of these topics, I looked at them. Some of them are a few years old.

They're pretty good. I still think they're very relevant. You want to talk about large scale databases, about identity access controls, how to scale engineering teams, how to do multi-cloud storage. Probably these are still relevant things today. These do not change as fast despite being high scale, interesting, powerful technologies.

So the previous wisdom of saying optimize for specific technologies, go deep, switch rarely,right? The reason you do that is because switching is hard. Migrations suck. Whenever you migrate, you break something every time, no matter what. It's always harder than you think, even if you know that.

And the switching cost is basically high. So basically, don't do it for most things. You're kind of just because something's better out there, that's not the answer for most infrastructure. So typically, if you say half life of an agent infrastructure or of a normal infrastructure, three to five years, reevaluate periodically, see what's out there.

We've been using databases like MySQL databases for a long time. It's still pretty good and probably need a replacement soon, but now with AI technologies, arguably the half life is measured in months, meaning a few months after you've adopted what might be the best possible thing, there's a significant chance that you're going to have to replace it coming soon.

And this is, I think, shocking. Maybe I see from some of your reactions that, like, you're kind of like have experienced this a little bit, but this is a different aspect of the way that you build technology. So if you're an engineer, this really sucks because the thing that you just learned and that you're making is now probably going to be out of date soon.

No engineer I know likes this. As a startup, you bet on something. You're like, we're going to go all in. We're going to go on this technology's approach, and then we're going to basically disrupt somebody, which probably will.

But then you see what you see now. Like, the first phase of AI companies are starting to get disrupted by the next phase. If you're a technology buyer, you're a leader of a company, you buy technology, you select open source models, you select vendors, there's a significant chance that whatever you just bought is not going to be the approach.

You're going to invest, you know, good luck doing a three-year deal, like, on things about this kind of stuff. Or if you're a VC, maybe the coolest best thing that everybody agrees is the greatest opportunity is no longer going to be the opportunity soon because everything's changing.

So here's my advice. Get good at changing.

It's almost silly to say because, you know, obviously technology changes. Obviously, it's something that is, you know, built in. Of course, we're all going to change. We've done this for a long time. It's hard. I think it's really hard.

And the faster that you do it, the harder it is. When I was going through that, like, oh yeah, we switched from the graph based agent to the more looping style, deep style agent. I remember very well the conversation with the engineer.

He just, he's like, it did it. I got agentic search working and this approach, this deep research, it does all this stuff just like you asked. Okay, we're going to switch. Rebuild it again in this new technology. And he's like, wait, what?

Like, it's working. You did what you're talking about. Yeah, but it's not as capable as we want it to be. Like, what do you mean? You didn't tell me that before. Like, and then so convince him. Like, okay, this is a new approach.

And then, you know, he does it and it's good. Two months later, we're actually shipping the product on Tuesday. And then I was like, okay, guys, on Wednesday we're going to rebuild it again on the new approach. And they're like, what are you talking about?

Like, again, then they'll say, like, it's almost hard on everybody. Like, wait, wait, give me more time. I'll make the new way, the old way, do it better. Like, and then also they're skeptical. Like, now you say that, but like, who is going to change again,right?

Like, who are you to, like, make these choices? And the answer is, yeah, I'm pretty sure it's going to change again. So this is, I think, a leadership problem. It's a technology problem. It's a morale problem. It's a team problem.

It's a company problem. And if you're not careful, it is actually can destroy you. It can destroy a lot of things because people lose faith, they lose morale. It's a problem. So if my advice is change and be ready for change, how are you going to do it?

Three things to give you. One, you just got to prepare people. This is a kind of a people challenge. So when you build your teams, when you talk to them, when you prepare them, if they're in AI world, you got to tell them, like, expect change.

It's normal. It's not a problem. It's not that you did something wrong. This is weirdly, like, helps people. Like, I have a technology review team and then they're like, we can't, like, change. We don't know. We're not sure.

We can't tell you that in two years from now this is going to be best. Like, that's okay. We're going to build these things that change. So just go with it. You have to pick something. Also, whenever possible, if you can build an abstraction so that it lets you swap out what's underneath.

We have an agent extraction in Box and you're able to go through and be like, select things underneath. And the agent still works the same for the customers, but it's better underneath. And the idea is that change is not a mistake.

And I highlight, like, that's very hard for most people. And I would sort of just tell them all the time, change is not a mistake. You wouldn't, nobody knew six months ago. Nobody today will know six months from now.

It seems very true. So at Box, we are now in the habit of reviewing every six months, no matter what. This is a great technology. We love it. Review in six months. Like, because, which is just completely crazy for everything else that we're doing.

Everything else is like three years. Also, even though change is critical, you need to define what you mean when change. If you just change all the time, there's a new paper. It's awesome. You know, our CEO, Aaron, is very active on all the newest things.

He's like, check this out. Like, don't change just because of that. Like, don't change just because it's a trend. Change because you know it matters. And how do you know it matters? Probably pitch you on eval sets. If you're building agents, if you're building AI, make sure that you know what people have.

You have the ability to give the same input, expect certain output, grade that. Cost, speed, quality, capabilities, these are the things that you probably are going to be wanting. So for us, it's easy. Does the new approach work better for our eval sets?

What the customer cares about? If the answer is yes, the stronger we consider switching. If the answer is no, don't bother. Like, or keep working on a little bit of work to see if you can make sure that you've fully explored it.

And then so the idea is build a system that lets you be able to change. And then the third and final piece of advice here is almost certainly none of us can keep up with everything. It is very hard.

I think I heard Andrej Karpathy, he was like, everything changes so fast, I can't keep up. And you're like, you're sort of quite famously good at keeping up. And so like, what's the hope for everybody else if that's the case?

And so, but then so what do you do is you rely on somebody else. You rely on a technology, you rely on a vendor, you rely on a platform when you select it. And then here, I think very useful, I mean, like whenever anybody's bought technology in the past, I would have advised them, like, look at what they do now.

Double check the roadmap, make sure it's good, make sure it's on the path you want, but just focus on what's available now. But I think something else here is, should do that, of course. That's the most important thing.

But like, look back. How have they handled change? What's their attitude towards change? What can you, when you talk to them, when you read about their stuff, like, what happened six months ago? What happened a year ago? How did they handle that transition?

Many of the vendors that I really likeright now have reinvented themselves three times in the last year. And I now trust that if something else comes along, they're very good at this. They understand agentic technologies, they understand the eval sets, they understand the observability systems.

And then you can say, ah, okay, good. I hope that they keep up. And then I now, my sort of thing I need to do is just evaluate whether or not that's a good platform. So making sure that you have this sort of platforms that do well is critical.

And if anybody's interested in unstructured content and AI associated with it, Box has a booth downstairs. Happy to talk to you about those kinds of things. And then I'll leave you with this is I actually, I fully bet and I believe that a company that's born this year, was born last year, will, or maybe even a company, a medium-sized company or a big company, they'll shoot very high.

The company that will dominate tomorrow is now born today. But I kind of bet you that the technology approach that they haveright now is probably going to change multiple times before they do that. So interestingly, it's like the challenge, the advice, the thought here is build for change.

Adaptability, arguably, that's the moat that you have until that changes. Okay, thank you, everyone.