# The Transition to Agent-First Software Architecture

**Podcast:** AI + a16z
**Published:** 2026-04-21

## Transcript

the diffusion of AI capability is going to take longer than people in Silicon Valley realize.
It's just absurd to think you're going to vibe code your way to like SAP.
All of that domain knowledge, it's not just represented in some well-orchestrated data layer.
Engineering compute budget conversation is going to be the most wild one in the next couple of years.
The biggest problem right now is everybody is trying to figure out the economics of all of this when they're off by at least an order of magnitude on how big the opportunity is.
If you have a hundred or a thousand times more agents than people, then your software has to be built for agents.
People in the abstract say things like, now you're marketing to agents and you're like an API, you've got a good idea.
I actually think that's almost exactly wrong.
Wow, this is breaking podcast news.
Every major technology wave promised to eliminate the middleman.
Marketplaces would dismantle hotels.
SaaS would replace on-premise.
But the taxi medallion was the only real casualty.
The layers persisted because they encoded organizational logic, not just software logic.
Now, agents are arriving, and the assumption is the same.
They will flatten everything.
But the first enterprise teams deploying agents at scale?
are discovering something different.
Agents do not want simpler systems.
They want better ones.
They choose backends based on durability, cost parameters, and reliability, not interface polish.
The question for every software company is no longer whether to support agents, but what it means when agents outnumber employees a thousand to one.
I speak with Aaron Levy, CEO at Box, alongside A16Z board partner, Steve Sanofsky.
and A16Z general partner, Martin Casado.
Do you start to imagine that we all have to build software for agents?
I think we're like all clear on that, right?
So like that trend is happening, which is like we spend as much time now thinking about the agent interface to our tool as we do the human interface.
Okay, and the reason we're doing that is because our hypothesis would be that if you have a hundred or a thousand times more agents than people.
Then your software has to be built for agents.
And then what is the way that those agents are going to interact with your system?
It's going to be through an API or a CLI or MCP or whatever.
And the paradigm that appears to be taking off and is quite successful so far in terms of efficacy is what if you give a coding agent access to your SaaS tools and a coding agent access to your knowledge work sort of workflows and context.
And that kind of becomes this superpower, which is the agent is not only capable of reading some data, understanding some information.
It can actually code its way or uses APIs through whatever task it's trying to achieve.
That appears to be like a paradigm that is starting to compound.
And that's the cloud co-work phenomenon.
That's the whatever open AI is kind of cooking up, you know, with the super app, perplexity computer, et cetera.
And I actually think it kind of makes sense as like the ultimate manifestation of this stuff.
I mean, I think you're right.
It makes...
sense in a theoretical way, but in a practical way, we have to be really careful in that the way to say it is algorithmic thinking is really, really, really hard for the vast majority of people who have jobs.
And so the easiest way to think about it is if you were to go into any person and ask them to create a flow chart for a particular thing that they have to go do, they would probably fail at producing that flow chart.
So within any organization, say, doing a marketing plan and there's 50 marketing people working on a giant product line, one person probably understands and could document the flow chart.
So if you put one of these agents or you put this tool, this co-working tool in front of people to create these things, their ability to explain to it what to do is really, really limited.
But what if that becomes the new...
this is the new way you have to interface with computers.
And you just have to cycle that through.
Well, then you're basically just developing the next abstraction layer for how people interact.
And developing an abstraction layer has historically, at each level of the abstraction layer, been a highly skilled, very specific individual within an organization developing that.
And then the little parts that they build just become little toolets in the world of people doing particular tasks.
And some people are able to stitch them together and some can't.
But that happened with paperclips and thumbtacks before.
And it's going to happen with whatever we do next.
I think basically the timeless part is the job just moves up a rung and you learn a new set of skills.
And that's why I actually don't think anything about this is any different.
It's just now the leverage you get is obviously fantastic.
There was this viral kind of tweet that went around, which was the Anthropic Growth Marketer.
Did you guys see this?
Basically one person, and he was using Cloud Code at the time to basically more or less automate what maybe five or 10 people would have done in various kind of silo jobs.
And I think the reason why it's interesting is you had to have been a systems thinker to be able to accomplish that.
So like clearly he already was technical enough to be able to pull that off.
But it did kind of represent what would each of these jobs look like if you had, like imagine you had, you know, X job in the economy and right next to that person was an infinite pool of engineers that could automate whatever that person wanted.
And what would that job look like in the future as a result of that automation that now is possible?
Yes, I agree that you'd have to find a way to think through your job as a system to be able to pull that off.
Maybe the agent gets better and better over time at being able to like nudge you in that direction.
But it does sort of stand to reason that like you will start to try and automate a lot of that kind of work of like, well, why don't I take like the keywords that are working in this in Google AdWords and then port them over to Facebook and make sure that those are replicated and then taking the new signal from what's happening in the market.
That's a big leap.
Yeah.
So one thing first.
I almost had you.
You were nodding a little bit.
I said something that went too far.
Using the anthropic growth person as an example.
That's a job.
That is the rest of work.
Yeah, I could do that job.
When demand is infinite and you've got the best thing.
Like when demand is infinite and frankly supply is infinite, this is not a difficult job.
And so let's.
The guy that runs the petrol pump in Australia right now is amazing.
Right, right.
So like be instead be the $600 PC marketing person.
see how you can do against the Neo.
That's a real job.
All right, fine.
We need a better example.
But there is, I mean, it is really interesting.
Like I, here, let me do an old example, an old person example.
Like my cousin, MBA, elite school, joined her first job.
She's a little older than me.
Joined right on the cusp of computing.
Like she actually didn't use a spreadsheet in grad school.
And then a spreadsheet showed up, but she wasn't a spreadsheet person.
So instead, they told her, hire as many interns as you want.
And so her first year on the job, she, like, supervised, like, essentially a whole room of agents.
And the kids, who was me, not literally, but they were in college, came and just did all the spreadsheeting.
But then what happened sort of magically over the next couple years was she and her cohort all became the spreadsheet people.
And then this idea that you being a manager in a bank or just a...
two years in meant you had a cadre of people doing this.
No, the whole abstraction layer moved up.
And the old job before those interns was you just sat there with basically calculators and an HP calculator figuring out the model for some M&A deal or whatever, and you only got to do like two iterations before you had to put out the pitch deck or just go to the customer or the client or whatever.
And then all of a sudden, they're doing 30 iterations themselves.
But...
And so I think where we are with agents is just at this step where you think you need 50 and the abstraction layer is such that we're dividing up in these really small pieces with one super smart person coordinating them all.
And pretty soon, the whole thing is just going to, they're all going to collapse on each other.
And there is just going to be like a skill set amount of code, call it an agent, that is like...
marketing-ish.
And you'll be able to ask it marketing stuff.
And then the next step will be and have it go do things.
I'm a little skeptical of the, until the whole, like, non-reproducible, non-random element of this AI stuff goes away, the doing stuff is going to get...
very costly.
And so then you get into the human in a loop discussion and all of that.
But I feel like when I talk to people trying to do stuff that were right, I feel like I'm at Thanksgiving dinner talking to my cousin six months in her job.
When I'm using a spreadsheet already and I'm like, I don't know why this is so hard.
You should just use one.
And then two years later, she's doing it.
And I think this right now, you have to be a rocket scientist and the growth marketing person to create 42 agents and spin them all up and do all of this stuff.
But the rocket science part of it just is going to evaporate in very short order.
And then you're talking about, wow, there's a giant chunk of domain expertise.
Yeah, it goes back to the domain expert.
So I actually think something that you said, I'll take the other side of, which is, I think it's very tempting to be like, These agents are going to code and do X.
But I think we're going the opposite way.
So I think actually where we started was we'd take like a piece of SaaS software and we'd add AI.
And then that's like the new kind of like AI enabled.
So that's like the extreme version of using code for these types of things.
But now what are we actually doing?
We're like, okay, the SaaS software is still SaaS software and the agent uses it as a computer because it's actually very good at that.
So I'd say like we started with code, then we went to the terminal.
which is actually less code.
And now this year is going to be the year of computer use.
So it's almost like they're much more like humans using computers than them generating code.
And that feels like very much like this mezzanine step.
And I actually come from like the generating code.
type of the world.
Like, I would argue that that's happening less, not more.
Yeah, I think, so, to me, whether it's computer use, API use, or writing code on the fly, I kind of maybe erroneously put that all in one bank category.
Well, they're very different.
They're very different, but we have an agent that we're working on where it just makes a determination whether it should use an existing skill.
It should be using an existing tool from Box or it should write code to solve that problem.
And its ability to do any one of those three at any moment ends up being incredibly useful because sometimes there's just some specific operation you want to be able to do where writing code to be able to do that operation is just faster.
And we can't possibly pre-plan for everything that anybody would ever want to do on their documents.
And so the fact that the model is good enough to also write code on the fly for that use case.
ends up just being like an amazing property, even though maybe 90% of the things that it's going to do should just be using an existing API.
And over time, Predo takes over.
And over time, there's literally like seven apps on her iPhone.
There's seven SaaS apps we end up...
Like over time, these things tend to consolidate.
But the seven apps on the iPhone is an issue of humans don't want to learn these things over and over again.
And so I, as a human, I don't have the mental bandwidth to learn that many apps.
But an agent that is going to use tools and APIs and be able to code...
doesn't have any of the same constraints that we have.
So I don't know.
Like, I don't mind.
Well, you could argue that there's just so many things to do and you can make interfaces sufficiently general.
Yes, fine.
Well, fair.
Let me say, I think I like what you said then because...
Oh, I'm back.
We're back.
We're aligned.
We're aligned.
No, but I think there's something super interesting here, which I do really, really like, which is that where software has evolved, you know, like...
I use SAP all day.
I work in finance.
I have to go and generate all these reports.
And then somebody shows up and says, I want a report that does this view sliced this way.
And I'm like, oh God, I don't know how to make that.
And like, now let me go wade through the SAP help system and try to find it.
One thing that, let's just say AI could be very good at is it actually can navigate that surface area.
much, much better.
You know, the help is all there.
And so it's a matter of finding it, mapping language.
And humans have been a bottleneck in tapping the past 25 years of software capabilities.
I mean, like, I spent my life...
My life with sitting next to people on airplanes saying, how can I make PowerPoint do X?
Just go to the ribbon.
And, you know, it was because it hurt, physically hurt to watch somebody suffering with bullets and numbering in Word or trying to figure out, you know, like, oh, let me just make a two-sided, a two-axis graph in Excel.
Right.
Which, like, is rocket science.
Like, almost no one can do that.
But yet it's super common.
And so people are like...
And so that impedance mismatch was a human user interface design.
I totally buy it.
On the consumption layer, I totally buy it, which is like the perfectly fluid UI or consumption layer.
I just feel the back end, like the systems of record, it'll probably converge into some database, some generic set of APIs that they'll connect to.
And that seems to be the direction it's going.
I agree.
So I spent all weekend...
implementing my Nanoclaw bot.
And when you first start out, it's like you're building an integration for everything.
Nanoclaw is very, like OpenClaw has all of the integrations, Nanoclaw has a few of them.
And so you have it build all of its own tools.
But after, you know, two or three days of these, like, you know, you kind of have the tool integrations that you need and, you know, like.
Yeah, but back to the, I mean, we're talking about personal productivity probably, like you're like organizing your life or something.
Well, it's work productivity.
Okay, fine, work productivity and then an SAP system and like, and like, and so there's like an, like there's an infinite amount of complexity when you get to, okay, some company that has a global supply chain and they're dealing with 75 pieces of information across, you know, 30 different systems.
that does require a certain amount of horsepower from the agent that is just, I mean, we just haven't been able to get from any architecture up until now.
But what you just described is literally what IT has been doing for 50 years and will continue to do, which is, yeah, I have a friend who was the CIO of the VA, and he spent...
All he spent his time on was gluing the 75 VA systems together.
And it's all just integration, redundancy.
Perfect.
For integration, these things are the best, but it's integration.
It's literally, how do I stitch these two systems together?
But now the thing that I think is happening...
is it's kind of like integration on demand.
It's my new query in the system that the IT team didn't pre-wire.
Now I need it to happen at runtime.
Let me get off my lawn.
Okay, so the reason I just was in a room filled with a bunch of CFOs and CIOs and...
They all looked at me when I said something along these lines, although not as optimistic as you can imagine.
More realism was in there.
No, it caused like six of them to come running up afterwards and say, you're insane.
You've lost all credibility with me.
Wait, what specifically?
That the agents are going to do integration?
The integration is a problem that will get a lot easier.
Yes.
They were against that?
No, no one's against it.
I think it's practical.
But their fear is unleashing not just the agents themselves, but humans to do integration.
Because you put people creating new integrations, and you just say, please break my system of record.
Oh, yeah.
And so this idea that you just create a new API between system 27 and system 38.
Yeah.
And then you're – that might be fine for a report because if that person wants to be wrong, that's their business.
But you're not – Consumption layer.
I think we have a read-only version of this for a number of years before – Where N is very large.
Yeah.
And a lot of it's just a consumption layer where the consumer is a human being.
Right, right.
It really feels right now a lot of the AI stuff is consumption layer.
But, yeah, I mean, it's – you know, we actually have – so we just rolled out the official box CLI.
Thank you for liking the tweet on that.
I used it.
I have some feedback.
We'll talk about it.
I'll take all the feedback.
But it's a really interesting thing.
So we had all these debates internally of like, okay, you give Cloud Code the box CLI, and you can now interact with your entire box system via natural language, and you get the horsepower of Opus 4-6 being the orchestrator of doing a bunch of operations.
And it's like...
It's like, you know, blows your mind.
I guess I'll get some feedback, but it blows your mind in some ways because you can just be like, upload this entire folder from my desktop into Box and it'll work or process all these documents in this folder and it'll work.
And it's amazing.
And then we started thinking through like, well, let's say you were a company with, you know, 5,000 employees and everybody had access to some shared repository, like, you know, engineering documentation and, you know, marketing assets or whatever.
And everybody had cloud code or codex.
you know, running with the CLI, wow, we now have some really interesting new challenges, which is like, like, how do you coordinate, you know, possibly the fact that you might be hitting the system, like, you know, 10,000 times an hour or something, not from a, like a performance standpoint, but just like, how do you make sure that people didn't move?
like a file from one thing accidentally from one folder to another folder while the other person was trying to do a right operation and somebody else was trying to delete something because you have these agents running wild this is this is going to be like the new big question that every CFO, CIO, et cetera, is running around with their hair on fire trying to figure out.
That's exactly what I ran into, which is I played around with your example, which is create, the video example, which is create like a marketing plan directory or something.
And like all of a sudden, I'm like in some loop creating directories.
And it's going to go on as long as it can.
Right.
And I was like, I wonder what the limit is on Vox for nested directories because I'm about to hit it.
Actually, we're going to find out too.
Yeah, yeah, yeah.
But it does feel...
to me that a lot of the intuition is to build a new layer of controls and whatever, but what's actually happening on the ground is the opposite.
So I'll give you an example.
When we all picked up a lot of these personal agents, we would give them our API keys, we would give them our email addresses, and then they would kind of access those things.
They're like, oh, but how can I stop it from whatever?
And so what everybody's doing now is you give it its own phone number.
I actually gave my NanoClaw its own credit card.
Hopefully just a Visa debit card that you bought at CVS.
It's got all the money.
No, but then I gave it its own Gmail account, which you can log into.
And then Gmail actually has all of these RBAC permissions.
So you could make an argument that like...
you know, we've actually built in a lot of these permission systems.
You have to treat it like a human, as a separate human, and then instead of like building another auth layer, building another...
Okay, now can I instantly do a takedown of this element that we're going to run into?
Please.
Okay, so that is fantastic for personal productivity.
And the question that we're going to run into is in an enterprise, let's say I have, let's just make a simple example.
I have a 50-person team of something.
Should...
Everybody also, basically, will we have 100 people now collaborate?
I mean, basically 50 humans.
And then 50 credit cards.
And then 50 agents in that same shared space.
And do I have, I obviously have complete oversight over my agent, but what if my agent collaborates with somebody else and then accidentally gets access to some resource because they were sharing with that other person and I'm not supposed to have access to that resource.
And now this autonomous sort of stateful.
you know, agent is running around working on somebody else's information.
The default end-to-end argument is you treat them like human beings.
It doesn't work.
So you can't fully treat them like humans because here's the thing.
And with regular humans, you don't get to look at the Slack channel of the person that is working with you or working for you.
You don't get to log in as them.
You don't get to oversee them.
They are accountable for their own set of execution in the real world.
You don't get penalized for how they screw up.
The agent, you have all the liability of whatever they're doing.
You do have complete oversight and you're probably going to need to have that complete oversight.
They have no right to privacy.
So there's going to be some of these breakdowns that aren't as clean as just treat them like a person because I need to be able to kind of, I need to be able to give access to something to them.
But I also need to be able to like log in as them at some point and be like, no, no, you fucked up the whole thing.
Right, right.
And I need to undo it all.
But if I can log in as them, how could they have operated in the real world working with other people and keeping anything, you know, confidential or secure or whatever?
So it really is still an extension of you.
It's like almost impossible to get around them being an extension of you.
So now the thing that we're thinking through that we're not going to be able to do anytime soon.
This doesn't logically follow.
Yeah, maybe.
But for example, for my employees, I can log in as them.
You don't though.
I can get access to their email.
Yeah, no, if you get, like, sued.
You're not logging in as them on a regular basis because they sent one email.
But isn't the right operating model with an agent the same thing?
The risk is, like, a thousand times greater.
Like, these people, like, they will just leak your information whenever they want.
Like they will happily just go and send some email to somebody because they got prompts injected.
You think the terminal state is that these things are still these sloppy computers and therefore they will always...
I don't like the word sloppy unless we're saying it very in a colloquial sense.
But like...
They'll never be able to contain information.
They'll never...
So like I think the ability for you to keep something in the context window a secret, like as in like you tell it do not reveal X thing in the context window, I think that's a very hard problem to solve.
And so then thus, if anything can ever enter that context window because they have access to a resource, then in theory, you should assume it can be, you know, prompt ejected out of the context window.
And I don't know that we know of a way to solve that at the moment.
Like, that's like, and so if I know your new agent's, you know, email address and I email it like it's an assistant, but like I can social engineer it 10 times easier than a human.
like it'll be hard for you to pull off that that agent is now also has access to your like M&A documents and stuff.
But isn't this like literally all of AI right now?
Which part?
I mean, the fact that we've got these shared systems that we use the intelligence for that have shared context.
But what do you mean by it's all of AI?
Well, I'm just saying like right now, when we use AI internally and agents internally, this is exactly how we use them.
But this is why they're working as you effectively right now.
And we don't yet know how to make them not work as you.
Let me offer an example.
Let me offer an example.
And then solving this problem, though.
Like, the issue will be, like, you will just be able to trick the agent to reveal information.
So then that's why, like, having them have access to their own resources where they can fully make their own decisions is not yet something that we've been able to pull off.
There's a perfect example for solving your problem, which is we already lived through this with open source.
Yes.
The model for open source was it's all there and you just use it and you pick and choose.
And then like...
Nobody debated it because the world was much smaller then and we weren't all on X doing podcasts when this was all happening.
But then quickly everybody realized all the problems you were just talking about.
Like if you're running a big company, you can't have some person just go copy in a bunch of source code from open source into your commercial product like that.
There was a whole licensing problem, a whole quality, a whole bunch of stuff.
And so all these norms got developed.
The debate that's happening right now is this really interesting modern artifact of how new technologies develop, which is this is all happening in real time.
During open source, we met in a conference room this big and debated how much open source we could use in Windows or Office.
And nobody on the internet knew we were having this debate.
And I think it's just so interesting that not just the debate about specifics, but this whole notion of where is this heading.
is happening in writ large.
And everybody is just trying to get to the end state like way, way more, like in a sense, more quickly than we can actually reach the end state.
And so what really needs to happen is people just need to go build.
We need standards.
What?
We just need some answers.
No, I think we've got different intuitions on the end statement.
No, no.
You don't want my intuition.
One could make an end-to-end argument that these things actually converge on the same type of reliability as a human being, which is exactly how we view, like, self-driving.
And in that case...
you use the exact same mechanisms that we use to protect with human beings.
You consider insider threat.
You consider the fact that people can be bought off.
You consider the fact that people make mistakes.
And that's just a risk.
And that's operational processes.
So one intuition is like, that will be the end state.
There's another intuition.
Well, don't point at me.
I'm just saying, I'm talking about where we're at now.
I actually, I don't know that we disagree on the end state.
Okay.
And by the way, like strategically, we're hedging because we're going to build, we're going to build agent users and we're like, so we're like, I love the idea of OpenClaw having a box account and it operates and you share with it.
Yeah, you just like twice as many accounts.
Yeah, exactly.
It's great.
Double the seats.
No, no.
I love it.
I'm just saying on the ground right now.
We don't yet know how to give it an M&A data room to fully, securely be able to...
But that, but it's actually, it is harder than that, though, because the threat...
He's the skeptic, I'm good.
The threat vectors are going to be way more sophisticated.
So we do have a cat and a mouse game going on where you can't just assume that the agent acts like a human does today because it's going to be the fastest...
most thoughtful, craziest-ass human that ever existed trying to actually leak the information because it got injected in some way.
And so...
Part of what's going to happen is we're going to go through this phase where like the enterprise customers are just going to like close everything off until there's some sense of sanity in all of this.
And then, but in the meantime, the individual and specifically the developers are going to, and that's going to be, that I think is the most exciting tension that's going to happen is that the enterprises are going to be, are going to get left behind by these.
sort of advanced individuals, which will then start to look like the startups.
And the startups will start to move much, much faster than enterprises because they just don't have any of these problems.
And, you know, you could end up with, like, the agent going rogue in a startup and doing bad stuff.
But you have employees that go rogue routinely in startups.
Yeah, yeah.
Well, it'll just be an episode of Silicon Valley.
And so, you know, big deal.
I agree with you on, like, the, okay, it's people, et cetera, the same risk.
I think there's a couple, you know, differences though, uh, in the sense that, that I can't really threaten, you know, the like cloud code that it's just, I'm going to pull the plug on it.
In the same way that you do have that threat as a regular employee is like you, at least like 95% of people are not, you know, trying to do bad stuff, you know, within an order.
Yeah, but they, they, they're not trying, but the ability to inadvertently do bad stuff.
Yeah.
To your point about it still not having that stuff.
I would argue that, that.
It's a lot easier to have people not share, let's say, files with somebody outside the company in a wrong way more than it is for an agent right now to have that same set of instructions.
And also you have the tools so that you can basically stop that at a whole different level of abstraction.
Which is why you have to build this into software.
But I do think actually if you were to like put a bow around your last point.
A lot of this is actually why the diffusion of AI capability is going to take longer than people in Silicon Valley realize.
Because what's happening is like we see startups that can start from the ground up without any of the risks that we're talking about because they have nothing to blow up.
And so we look at that as the trajectory that we're on.
And then you go to like JP Morgan and you're like, how are you going to set up NanoClaw?
to be able to actually like, you know, automate your business anytime soon.
And it's like, oh, okay, there's going to be like a little bit of a gap there.
Well, what do you guys think?
I think that that opens up a pretty interesting problem, which is this split between big and small startup and enterprise, which is just that the...
The current SaaS vendors, who are all struggling in this SaaS-pocalypse weirdness that I don't really agree with, but they are struggling with this problem that they don't really sell the line of business data.
They actually sell this intelligence and domain expertise in this whole system.
And the agent side of things wants to only buy the data now.
And they only want to license the data and they want to have unlimited access to the data.
But they've actually never really enabled that.
Like, that's never been their business.
And it's been a longstanding tension point with the likes of Workday and SAP and stuff.
Like, how much API access to have.
I mean, Salesforce went through three different massive platform redesigns.
You know, I think that that's a particularly interesting problem.
Not for the same reason that Wall Street does.
Wall Street's all wrong about.
the economics and the problem and all that stuff.
But from a technology perspective, what does system of record mean in the face of people wanting to access the data?
For training or for?
Are you talking about for actual day-to-day operations?
I think of it as executing the day-to-day operations.
Their concern is that somebody, that they want to do the training layer on your data.
Like, I'm a big customer.
They want to do the.
My vendor wants to build a training...
Actually, even if you don't even get into training, they're concerned because...
Oh, yeah.
Because, like, monetizing, you know, sending a little bit...
over the internet versus, like, you're in my UI.
Oh, all that stuff.
It's a very different level of monetization initially that you perceive.
But that's sort of that monetization part is the Wall Street point.
Because I think, like, look, there is so much domain stuff in an SAP, just to pick an example, not to pick on them or anything, but, like, they're not going anywhere.
Like, it's ridiculous.
It's just absurd to think you're going to vibe code your way to, like, SAP.
But also, all of that domain knowledge, it's not just represented in some...
well-orchestrated data layer as much as they tried.
There's like a whole bunch in the UI.
There's a whole bunch in middle tiers.
There's a whole bunch in just how you use it.
And so I'm really unsure how this thing evolves because SAP isn't going anywhere.
So then that's going to slow the diffusion of AI on that particular data source, independent of whether or not it's...
agentified AI that's doing stuff or just read-only reporting on stuff?
So where do you come down on it?
Where do you think that's going to go?
I'm afraid of saying something that happens.
I want you to say something.
Otherwise, you're not going to get invited back.
Say something good.
I think I've drunk the Kool-Aid on build something agents want.
So this kind of the Paul Graham term kind of like emerged on, you know, the past year on this topic, which is just like, like, I think we would actually then fully agree on this, which is at some point, you do enough sort of iterations of this.
And at some point, the agent is largely in charge of what tools it wants to implement and use and whatnot.
And yes, it can't, the agent is not going to be able to change out an enterprise system.
But like, again enough generations later the agent might might just run into so many walls with your software that it's just going to say you need to finally rip out your legacy hr system or i'm not going to be able to automate this workflow for you yeah so i do think you have this really interesting you know dynamic which is back to this whole whole point of imagine that there's a hundred or a thousand times more agent volume on software than people you do that enough times and eventually the software stack that agents talk to has to be built for them and And maybe there'll be a couple holdouts.
Maybe a couple ERP systems are like the final holdouts that don't do that.
But everything else, you basically, like your business will be, your business performance will correlate to how well your agents can get access to the information they need to do their work.
And so thus your enterprise IT stack has to be set up in such a way to support that.
And so agents are kind of in charge because basically your software has to support those agents being effective.
And that's going to mean Everybody that built a SaaS business or a software business is like, the game is, can you build really, really high quality APIs?
Can you have a way of monetizing that?
You know, do you have a way of handling the identities and all of the access controls for agents?
And like, that becomes the new problem you have to solve if you're building a software company.
And so, yeah, like, and then how you monetize it, like, do you monetize it?
Like, does Workday charge a penny for every HR record it pulls?
Like, we'll figure that out.
I do think that in some businesses, it could mean less revenue.
And then in other businesses, it could mean a lot more revenue.
Like the thing we get excited by is like every agent really loves working with files.
So there'll probably be more files in the future than there was going to be before.
And so, you know, can we build a platform that like makes it really easy for agents to work with that data?
You know, we're betting that that's actually a really optimistic outcome for, you know, our kind of business model.
There might be some business models that are like more constrained because like the agent is doing more of the value than the software is in that kind of future scenario.
And then there'll be everything in between.
Can I quibble with one thing?
You're going to quibble with that?
I thought that was like so not controversial.
No, no, I generally...
We're here to quibble.
No, no, no, no.
But there's one thing I think like Paul Graham and many actually gloss over, which is they focus on the interface.
They'll say things like you build something for the agent.
Yeah.
And I actually think that's exactly wrong.
Okay.
In the sense that...
And to be fair to Paul Graham...
uh he didn't he had been extrapolated yeah yeah yeah yeah i have brought i brought paul graham into this okay so okay let me talk about something people in the abstract say things like now you're marketing to agents the most important thing is to being like whatever you're like an api you've got a good idea i actually think that's almost exactly wrong which is wow that's This is breaking podcast news.
That's the one thing agents are really good at.
Oh, okay.
It's finding their way through.
And at the end of the day, like it's the semantics that end up mattering a lot more.
Right.
And so like the agents in my recollection or in my experience are very, very good at picking the right.
back end for whatever they're doing.
So they're not like, oh, the interface for this is very good.
It's none of that.
They're like, the cost parameters of this, the durability of that.
And so they actually have the collective wisdom of our experience using these platforms.
Let's take cloud platforms.
There's a bunch of cloud platforms out there.
And whenever I ask an agent to choose a platform, it's actually using meaningful stuff, not interface stuff.
So I think as an industry, we're so focused on these interfaces.
Like, oh, you need to market to agents, this and that.
But really, I think that we're going to be pushed to actually build better systems.
And that's what's going to be chosen.
Okay, actually, so then there's probably no quibbling.
I think we're actually fully aligned.
I'm sorry to ruin the quibble thing.
I don't treat this as like a, you know, kind of a marketing-esque thing.
I more mean like if your tool is closed off to the agent.
the agent eventually will find a better tool for that company to go use.
And so what will happen is it used to be that you would go to Gartner to be like, tell me what system to use or whatnot.
At some point with enough iterations, the agent is going to say, you should probably use this kind of database for this type of operation.
And if you're not in there, then it's your DOA.
And I think we should actually be celebrating this because agents are actually pretty smart at choosing the right technology.
In the past, I really think it was a lot of the other things that caused people to buy it.
But don't worry, we will, in Silicon Valley, we will ruin the meritocracy of this very quickly.
Because you'll just be like, I'm going to outspend.
Well, the agent, they'll bring an API to incent the agent.
The marketing agent at Workday will have the ability to purchase.
We will find a way to replicate steak dinners for agents.
Here's a thing that, again, that happened with the web sort of internally.
Like, internal.
Like, just pick internal sites.
Like, every company had file shares with, like, the best documentation, the best slideshows, the best financial models for any department or working area.
And people sort of got familiar with that.
And then when they didn't find the one they wanted, they created a new one.
And many organizations...
sort of operated like that was essentially a free market.
In fact, because before the world of Box, like IT didn't, if it was in a file, they just didn't care.
They only cared about if it was in SQL.
And so one of the risks with the model you're describing is that the agents themselves will spin up what becomes like a de facto new system of record.
Oh, they're going to fragment the heck out of it.
In what the IT people think of as some middleware end user BS area.
And I think that that is a real risk.
100%.
Is that like in a sense, like the macros end up running the corporation.
Yes.
And so I think that they've seen this movie and they've seen what happens when you let marketing just go buy a website on the internet to do an event.
And then it's like a huge security vulnerability and the mailing list is leaked and the whole company gets sued.
Totally.
And so I think there's a lot more real world tension in this dynamic than...
than we just let on.
But I also think it's one of these ones where organizations are going to run at different paces, and JPMorgan is going to be the slowest at doing this, and the startups are going to be the fastest.
But the delta is huge, but even the startup one is a little far off, because even startups do need some systems of record.
at some point.
And they are going to all start with some SaaS and they're not going to replace it very quickly.
So I think it's a little bit trickier.
So it feels like there's two very competing viewpoints on this one.
And like Elon said, it was like, okay, we're going to like issue a prompt and it's going to like spit out machine code.
And that's basically the collapsing of layer view.
Like whatever existing interfaces and layers that we've created in the past are all going to go away and it's literally like prompt to machine code.
The other argument, like, the history of systems is layers never go away.
They just get layered, right?
Because a lot of the layers are actually more of, like, organizational boundaries or, like, state boundaries.
Or compatibility.
They're just, they stay for compatibility.
Right.
So the other argument is, is, like, we've actually evolved these layers very specifically because of, like, more human and organizational needs.
And they're not going to change.
And the agents are going to go ahead and map to those.
And I tend to be in that latter camp.
Like, I don't think that we're, I think, like, systems are going to continue to be used in fairness.
similar ways.
Maybe there's more agents using them.
I don't think they're going to evolve as much.
Elon might be back in the, like, anthropic category of the anthropic growth marketer.
Yeah, yeah.
Which is, like, he, like, you know, over the years when you kind of, like, study the various IT, you know, departments of his companies, like...
They are the most...
I mean, I'm not going to say that.
Well, he could do that.
He can do it.
He's the most homegrown...
Like, everything is...
This is first principle.
Elon AI would do that.
Elon AI would like...
But also, it's first...
And then from your mortals, you're like, yeah, we kind of just want a CRM system that kind of works the same way every time.
I mean, this is not...
It also hasn't been not tried before.
Like, if you were to look at an ERP system from first principles, you know, well...
In 1970-whatever, when SAP started, there were a bunch of different assumptions, and today you would start from a different set of assumptions about what's important, and you would architect a thing completely differently, but then it would still only last like 10 years until you thought, wow, that was a broken decision.
And so I think that there's intentionality in layers, but there's also this first principles thing.
And, you know, that always will exist because the decisions you can make at first principles at any given time mandate a whole bunch of different stuff.
And so even if you don't go with LiDAR, which made total sense 10 years ago, you still need 10 or 15 years to get to where LiDAR...
not having LIDAR worked.
And then now there's going to be a whole bunch of other things that you're like, wow, we could have done that completely different.
And so I feel like this is, again, like this discussion about trying to race to an endpoint.
But let's see a first example of what you described happening.
And I think that that's going to be the real tell, because I think that there were just...
Companies will figure all this out.
And I think that they will just fall back on layers and architectural models because it's the only way.
We don't have to think about it for policy.
We don't have to think about it for security.
But it's also the only way to build a system.
Otherwise, you're just building an app.
And if you're building an app to do one thing, we don't need all of this.
Like, there's a whole different way to do it.
The thing that I'm pretty fascinated by is, and I don't even have any amazing data points or anecdotes, but at least the notion of these sort of companies that are emerging in these kind of services categories from the ground up, from the pure first principles approach, which is like, okay, well, if I could start a marketing agency or consultant, you know, engineering consulting company, or I don't know, maybe somebody's doing this for law firms.
Construction work or anything, yeah.
Well, maybe construction.
Because it's design, construction design, architecture.
Oh, yeah, exactly.
Architecture, design, anything that would be like a knowledge worker kind of services company.
Because you could kind of build your company pretty differently.
If you had no constraints of, I have no information barriers and boundaries of what people should have access to.
I can give the agent just all the context it needs to do its work.
I can write software on the fly for particular things.
Like, I do think that will be relatively disruptive, you know, for some time until the bigger incumbents can kind of, you know, get out of the way on this.
And that will at least create, you know, some precedent or case studies of what this new sort of corporation could look like.
Over time, they'll still run into the same exact problems of every other corporation.
Well, they'll run into geography or market segments or distribution challenges.
Those things, anything outside your little walls, you will run into the physical world.
Right.
I do kind of like the idea that there are some new business models that open up now.
Oh, of course.
Yeah, yeah, yeah.
Because there's so much either information or software that basically...
goes underutilized by like 100x relative to like what its economic value is, simply because like nobody wants to pay five cents for accessing a piece of data or use a tool for $1 once.
But like you do give these agents, you know, a budget and a protocol to work with.
And all of a sudden you're like, oh, like on the fly, they can go get medical research for some deep research tasks they're doing.
And I'll pay like $3 for that.
And the agent is able to go and transact.
It kind of opens up a whole new world of business models for the internet.
Let me, oh, I'm going to, that was too nice.
Oh, okay.
No, no, that one is one.
You're going to go farther.
No, that one is one where that's actually the biggest, I think that the biggest sort of in the air problem right now is everybody is trying to figure out the economics of all of this.
When they're off by at least an order of magnitude on how big the opportunity is.
Because the new models.
that people will come up with that nobody knows what they are right now, but they will absolutely come out with new models because that's what happens with every new technology.
And the thing that holds back sort of the discussion now is you basically have a bunch of finance and Wall Street people trying to justify GPUs and tokens and things like as if we're in some old world.
So they're viewing the world of revenue as sort of this...
linear growth curve and trying to justify all the expenses.
When people are going to create, like this was the problem with PCs.
People viewed PCs as a finite market because they just viewed the consumption of MIPS as some finite thing.
And they didn't think what would happen if we put all those MIPS on every desktop.
And in particular, people thought software just came with the MIPS.
And nobody thought, oh, well, they'll just sell the software.
One guy did.
And it turns out that was like a really good idea.
Was it Bill or somebody?
Yeah, Bill and Paul.
And the same thing happened, but the same thing happened with the cloud, which was people looked at the cloud and they said, oh, we're going to take all of the server business, which was like literally like 60,000 units a year.
and we're just going to move it to someone else's data center.
And that would be the business, and then we'll divide up the price.
And nobody went, oh, people are going to use a thousand times as much.
of the resource leveling if we move it there.
And that's exactly, I mean, that's the thing that I, it just drives me absolutely bonkers that the Wall Street models have this fixed revenue pie.
Zero-sum thinking.
And it's this weird zero-sum where they just think that the amount of money that a company is going to spend, and like this was the problem with Salesforce that they faced when you were starting too, but like Mark was just blazing the trail, which was like the CRM business was two billion a year, and it was two billion in like...
you had to go buy all these servers and these Oracle licenses and this huge headache and years of deployment and consulting, when if you could just get salespeople to sign up individually, they all will sign up with no friction.
And that is exact, there is no, no doubt that that is what's going to happen with AI.
Let me give you an example of this.
So I, you know, I've been in for investing for 10 years now.
I probably have a portfolio of 240 companies that work with.
So visibility, let's say 50 of them, these are all infrastructure companies.
historically have done well, some not so well.
Every single one of them has gone asymptotic in the last six months.
And you're like, okay, why is this?
It just turns out there's so much more software being written now than ever has been before.
And so it's like, and it's not because they've got enterprise customers, you know, it's just because there's just so much consumption of the infrastructure layer right now.
And so with more software, with more agents, there's going to be a lot more consumption of computer resources.
So certainly in the case of the computer side of things, we're going to see a massive...
Well, we haven't even gotten to the point yet where everyone's phone is a huge consumer of AI.
So once everybody's phone and on device, like once your phone on device is consuming AI, the amount of it is going to go up by a billion.
So do you like the micropayment piece?
All of them.
The micropayments, there's a little bit of micropayments that has come with every technology where they always think that like you'll be able to get like a fraction of a penny.
But in the end, especially in the enterprise, like people are just going to consume things.
It's just cheaper and easier to buy like a bulk license for a bunch of stuff.
You want some predictability in that.
Well, you want predictability and you just want like to not have to think about it.
I just, I like the idea that it is the first time that you could, like there's just, the agent doesn't care about the friction of a small transaction.
And so it's the first time that you can have resources behind a paywall that something will actually be willing to pay for that resource.
The world has built up the infrastructure to aggregate those payments into something efficient for a customer or a service.
And because tokens are such a significant part of COGS, I know it is pushing the industry to do usage base in a way that we have.
Like I remember when we went from like perpetual to recurring and that required like a bunch of huge changes.
Like, we're going through the exact same change right now towards usage base, and usage base is pretty granular, and it actually allows.
I mean, again, you will have a contract with, like, AWS or Google.
We went through this with AWS.
Yeah, yeah.
Like, people learned to do usage credit.
And we went through the phase where, like, people were, like, so terrified of cloud compute that they were like, we need companies in the middle to help us find the cheapest and to arbitrate it all.
Okay, well, now you write tokens into this, and I don't see how we possibly have time in this conversation.
Please, I'm here as long as you guys can say.
Oh, okay.
But, like, the engineering compute budget conversation, to me, is going to be just, like, the most wild one in the next couple years.
It's just like, how much should you allocate of your engineering expense to tokens?
And it's like, you know, depending on who you read on Twitter, it could be 1%, and the other side could be 100%.
Yeah, but this stuff.
No, no.
CFOs have to literally, they actually have to know the answer to that.
I understand they have to know.
CFOs always want to know the answers to things that don't have answers.
No, Wall Street is going to make them know the answer.
No, Wall Street is going to make them come up with some number and hold them to it.
Then they'll get fired.
I hear you.
R&D is somewhere between 14% to 30% of revenue of any public technology company, let's just say.
The difference between compute being 2x the cost of your engineering team or, you know, 3% more is like, that's all your EPS.
I get it.
We will have to know the answer.
I'm perfectly willing to sacrifice a few CFOs at the altar of this.
I want that.
That's a good clip, by the way.
But the reason is because, again, this is trying to know what we just don't know right now.
And this has happened with internet bandwidth.
No, this is not even close to internet bandwidth.
Oh, no, no, no.
I beg to differ.
Like people were free.
It happened with vacuum tubes.
It happened with transistors.
It has happened with every technology.
There was this, oh my God.
It happened with programmers.
There was a time when programmers were going to swallow every company.
Yeah.
And that's not, it was in my lifetime, not some made up weird thing.
Yeah, but I don't think we've ever had a point where every end user in an organization has sort of a completely elastic ability to spin up.
a resource on their behalf.
That actually is actually in many cases very valid for them to go spin it up.
But it certainly rhymes with what happened in the early 2000s with the cloud.
I remember very similar discussions when we went from CapEx to OpEx and then unlimited spend.
Oh, no, and remember there were companies who the CFOs would sit in our briefing center here and say, you don't understand, we are like, we are an agriculture company.
I can see the rhyming.
We are an agriculture company.
We only know CapEx.
We have no...
I sold through this.
Right, no, we both did.
Or like, oh, no, we are an OpEx-based company.
So if you tell us, we love the cloud because we just shifted everything to OpEx.
And so all of the stuff, like the rules of accounting work out.
Also, don't, I keep thinking, do not discount the local compute engine as being a release valve for all of this.
When's that going to happen?
Well, the question is, it's not when does it just happen with today's view of the technology, but how all of a sudden, wow, there's a whole...
Has that historically ever gone that direction?
Yeah, exactly.
It goes the opposite, right?
No, it went all to the client.
Well, okay, you can go back to the 80s, yes.
No, that's most of the examples that we're hearing so far.
Whoa, that was uncalled for.
Vacuum tubes, he's talking about vacuum tubes.
But I do those examples because you can't argue with them, and it's much easier that way.
You're right, I can't prosecute.
But it's only been, you know, 10 or 15 years that it's all, that it moved back to all cloud, and then what has happened recently with that?
A lot of people wake up in the morning and they say, oh, we're moving back.
to doing some critical but stationary workflows on-prem.
With AI, that's true.
Dude, you wrote the blog post, man.
Don't make me go through the archives.
I had to deal with so many Wall Street questions on that one, by the way.
Well, also because you're a competitor.
What?
Went back to...
Oh, yeah, yeah.
We're talking about two very different...
I agree with building your own data center.
I'm talking about this notion of edge computing where things go to devices.
That seems to be...
I'm more in the cloud maximalist camp.
Sorry, so you just don't think...
You don't even think for one second that it matters how you're supposed to be an engineering leader right now managing the compute budget of the engineering team?
No, of course it matters.
This thing in the long term, this thing will get.
Oh, sure.
Oh, what do we do?
Who cares?
We don't even need a podcast in the long term.
Here's what I think.
But here's a rule of thumb.
First, like the startups are going to burn through available capital.
pretending like it's not a problem.
And they are going to do that.
Yeah, but they do that anyway.
Right, right.
And a lot of big companies are going to be so terrified, they're just going to freeze and not do anything.
And then people are going to actually start buying it on their own, and they're going to do all the things that companies do when they're big, have a lot of money, but don't want to spend it.
And in the middle, we are going to see, like, if you pick a category of product or go to market or something, there are going to be people who are willing to make the bet for whatever reasons, that they can because of their financials.
And they are going to go ahead.
And they are going to become the people who lead in the space, so long as they can maintain the financials.
Now, they might do it in, they might say, oh, we're going to just do it here in this particular application space or here in this particular usage space.
But this idea that nobody is going to go in because they're so terrified that the CFO is going to get fired or something is just crazy.
Yeah.
But then there are going to be CFOs who make a mistake and like, okay, everybody gets a little.
Yes.
Well, if they do that, that's a complete fail.
Yes.
But also, or like you get, there is a really interesting like, you know, finesse here, which is like.
You don't really want your engineers right now having to think about compute budget because we're still developing the, oh, okay, so that sent you over.
I just feel like we've been having the discussion for 15 years when it comes to cloud.
This is totally new.
Only like 10% of your engineering had to think about cloud infrastructure spend.
In 2016 to 2018 timeframe, there was a whole set of companies that was basically like the dashboard.
for, what was it called?
FinOps.
FinOps is very cool right now because of AI.
Developers would have access because cloud spends are getting out of control and API spend are getting out of control.
And so it was like, you know, here's your Twilio spend.
But, you know, it's pretty different.
And I'm going to wait for all the comments to come in on YouTube to call you out on this.
Like, it's like you can get into a conference room and just be like, hey, can you make that one, you know, kind of algorithm a little bit more efficient so you don't use as much?
you know, of our cluster at this time of night or whatever.
And then you get out of the meeting, somebody go improves it and you're good.
This is like every single prompt that every engineer is doing.
Like, do you, like you have to decide, like, do you want that to be a long running prompt?
Do you want that?
Do you want to be a long running agent?
Do you want to parallelize that?
Like, like, do you want, like, what is your comfort level of wasted tokens?
Like for me right now, I'm like, yeah, we should probably waste a lot of tokens because that means that we're like trying new things.
And like, like, should your head of engineering be happy?
if you run 10 experiments in parallel, and thus you're obviously going to be wasting 90% of the tokens, but you're going to choose one of the successful paths?
Or do you want to tell the team, no, before you go do that, make sure to really go and design the perfect system?
We actually have a whole bunch of open questions that are going to start to happen.
Literally, as of this recording time, people are freaking out right now on the new Cloud Code Max plan because they're getting blocked after three prompts.
So this is going to be like a very like real topic until we can actually find a way to build data center capacity.
Oh, that's a different problem.
Okay.
No, because no, one is, well, assuming, well, wait, no, you can assume that if we build more capacity, the price will drop because there is more capacity and we're priced now based on limited capacity, whatever.
But like.
this is just going to get worked out, and I feel bad for those that have to make a decision immediately about which 17 people get no more tokens this week or whatever, and that the whole company is walking around with, like, a token card, and the person in the lunch line is punching their card every time they do something.
But, like, you know, I...
I don't know, like, somebody we were talking about today about performance and how, like, you know, we used to write command line tools that spit out the time it took after you ran a command line just so you knew.
And if you knew you were getting better or worse.
And, you know, but the thing is, this is all going to go away.
There's absolutely no doubt that this just goes away.
I think on a 10-year timeframe, 100%.
And the biggest reason it does...
is because you have to do the Benioff kind of math, which is if you're paying an enterprise salesperson, you know, a million dollars a year, you have to ask, how much is their tool worth?
And if you're paying an engineer X dollars a year, well, at some point, their tooling is worth...
It's absolutely worth it.
And it's not going to even be an issue.
And...
Yeah, yeah, I don't think it's...
And so if there's a capacity thing in the short term, that's a different...
That is true.
a different problem driving the price than this just we're going to forever have to be in some budgeting exercise.
I think law of large numbers solves this because eventually you have enough engineers using this much compute.
But, like, we're in a transition phase where, like, most people thought, you know, the two-year-ago level of spend on AI, which is like, ah, it's a chatbot.
Yeah, yeah, but they were wrong.
Yeah, right, okay.
But they were wrong.
We tried to warn them.
No, but they were wrong because they saw it as this...
particular use case yes and but again like you you know like the vacuum tube thing you made fun of yeah but like there was a time when they thought that the that like whole like all of the dakotas would be covered in vacuum tube warehouses and people on roller skates would be running up and down the aisles replacing vacuum tubes just so we could fight world war ii i mean like that was how that was the and they thought that and then someone said hey how about a transistor And, like, we are going to have a transistor moment with all of this.
It might just be more supply the way we think of it, but it also might be an actual algorithmic fundamental change.
It could be a change in the hardware.
There's a lot of stuff that can happen that changes this particular moment in time.
It's just this, I think it's particularly weird that everybody has just gotten to token, which is the same thing that happened with IBM and mainframes.
People were on MIPS, and then one day...
The reality was IBM was selling more MIPS for fewer dollars every year and didn't even realize it.
And they were still pricing their mainframes by MIPS until it got pointed out to them that they were on a decreasing curve because they were making MIPS faster than they can charge for.
And that's what's going to happen.
Guaranteed.
I just said that in a hardcore way.
I think that was great.
It sounds really great to sound like I know what I'm making.
Guaranteed.
Guaranteed.
I should probably believe it.
Thanks for listening to this episode of the A16Z podcast.
If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family.
For more episodes, go to YouTube, Apple Podcasts and Spotify.
Follow us on X at A16Z and subscribe to our sub stack at A16Z.substack.com.
Thanks again for listening and I'll see you in the next episode.
As a reminder, the content here is for informational purposes only.
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.