Microsoft's Kevin Scott on the Agentic Web

Source material

Full source text

you're someone who I think cares a lot about the craft of things.

One of the knocks on, you know, using agents for coding is it gets rid of some of that feeling or something like that.

How do you feel about that?

I love the fact that my people, you know, makers, uh, writ large, so software engineers or, you know, mechanical engineers or woodworkers or potters.

If you are really passionate about what you do, you're going to have very strong opinions about how you do it.

I've been a woodworker for almost as long as I've been a programmer.

This is not the first moment in the past four decades where the nature of software development has changed in a non-trivial way.

Like, if agents are going to be useful, they have to take action on your behalf.

They have to be able to use tools and make changes in systems and consult information sources that are, you know, diverse and rich.

And in order for that to, like, really be great.

You need an ecosystem that looks a lot like the internet.

If you have a source of information, you already have a website, you already have an API that's doing something for people out there.

Like you've got to figure out how to plumb things through where agents can like talk to those things.

Like one of the things, like one of the things, the CTO that I've been pushing for at Microsoft is like, I want all of our systems internally to speak a standard protocol to all of the agents that we're writing inside of Microsoft.

Be curious, try stuff.

And if it works for you, use it.

And if it doesn't, don't.

Kevin, welcome to the show.

Thanks for having me.

One of the things that's interesting is I was here last year.

And you said two things.

There are two big themes.

One was agents are going to be everywhere.

That's one of the things you said, which I think really came true.

That was very prescient.

Another thing is I noticed a big emphasis last year on scaling laws.

There are a lot of graphs of like, we're building big infrastructure, we're building these bigger models.

And every two years, we're going to get these big performance improvements.

This year, the emphasis is really on the agentic web.

So what has changed?

What have we learned from last year to this year?

Yeah, I think there's a bunch of things that have changed.

Like one of the things is that I think last year, people were like really in this state of mind that they were doubting that the scaling laws were going to continue to work really well.

Whereas like I think, you know, we've demonstrated year after year after year that they are intact and working quite well, that like that's not a thing that people need to be reminded of anymore.

And I think the other thing too, that's happened honestly, is that the reasoning capabilities of the models has actually gotten a little bit ahead of what we're using the models to do in products.

So I've been talking a lot lately about this thing called the capability overhang.

And, you know, I think we actually have some work to do collectively across the whole industry to close the gap between what the models are actually capable of and like what we're delivering to users of that capability.

So that's one of the like big thematic things why, you know, scaling laws might not be as interesting to talk about it this year's build as last.

And then the other thing too is that we've just discovered is all of these agents have emerged over the past year.

And so like both the number of agents and like the amount of time that people are spending doing stuff inside of these agents or with these agents is that there's a bunch of other stuff other than reasoning that has to get sorted out in order to make them as useful as they should be.

So, you know, we need better agentic memory like our agents right now because memory is constrained in a bunch of interesting ways.

They're a little bit transactional.

So like you use them for like one thing and like, you know, memory is coherent across the course of that task.

But, you know, then it may or may not completely go away.

And like you're starting from scratch the next time, which really inhibits your ability to delegate increasingly complicated tasks to these things.

And then like there's just this real issue that like if agents are going to be useful, they have to take action on your behalf.

Like they have to be able to use tools and like make changes in systems and consult information sources that are, you know, diverse and rich.

And in order for that to like really be great, you need like an ecosystem that looks a lot like the Internet, where if you have a source of information, you already have a website, you already have an API that's doing something for people out there.

Like you've got to figure out how to plumb things through where agents can like talk to those things and where, you know, all of the incentives are aligned for everyone to like go have all of this stuff participating in, you know, ways that make sense to them in this agentic web.

And so I think that is the big story this year.

It's like you've seen the first glimmers of like real progress with like super awesome, simple open protocols like MCP that are serving the same purpose in this agentic web as HTTP does on the Internet and where you have things like NL web that are like serving, you know, the same moral equivalent purpose as HTML does on the Internet.

And so I think you're going to see these things like simple things that are composable and layering and like just lots of activity in the open community and like a bunch of things hopefully getting to ubiquity so that agents can actually do stuff.

Right.

So I think to play that back, like one of the things I hear is like we have agents, agents are starting to work.

And in order to make them powerful, agents need access.

They need access to like whatever is out on the Internet, whatever is on your computer, all that kind of stuff.

Yep.

And you need basically protocols and processes for agents to be able to access that stuff.

And so you're looking at different parts of the stack.

So like the runtime layer where you're, you know, you're building like memory and all this kind of stuff.

And then MCP, which allows you to connect into the wider Internet to get more information into agents.

Yep.

I guess why is that important to Microsoft and like what role do you want to play in that kind of an ecosystem?

Well, look, I think there are, you know, two, maybe three things that are super important.

So one is like we make agents and in order for our agents that we're building for folks to be useful, like we need to solve these problems inside of these agents.

And like even if you sort of scope it down narrowly to enterprise agents, like one of the things, the CTO that I've been pushing for at Microsoft is like I want all of our systems internally to speak a standard protocol to all of the agents that we're writing inside of Microsoft.

So that we're not exposing the entire world to like Conway's law, which is, you know, organization.

So Conway's law is a really funky thing in compilers where, you know, this guy Conway said that the number of stages or passes in your compiler is going to be dictated by the number of teams you have working on the compiler.

You ship your org chart.

Correct.

You ship your org chart.

And so, you know, you certainly inside of the confines of a company like Microsoft, like you don't want to be shipping your org chart when you are building your agents.

And like it's just kind of a horror show to watch like as an engineer, like all of this inefficient building, like when you don't have like those standard protocols and services that everybody's using.

But I think, you know, if you sort of imagine.

If you really imagine like what agents could do and like what you, you know, what.

If you're using users, like not me, but like, you know, just people who are hoping these things can be more useful than they are right now.

Like you need things to start happening the same way that they were happening with the web.

And like I kind of see it right now.

Like MCP is like a really great example.

So it is a really simple protocol that solves a really important problem, not just for people who are making agents and building platform infrastructure, but for, you know, users of these systems who want them to be more useful.

And for people who are providers who are like, hey, like I want to be participating in this new organic web is like people are doing less of one thing, which I knew how to connect to like how.

And they're, you know, they're sort of sitting here using these agents.

Like how do I get my stuff wired up into this and like how does it make sense for me, like even from a business model perspective to do that.

And so like that's the two things that's like make our own agents more useful.

And then like we're a platform company, like even more important than the agents that we're going to write ourselves, like that platform layer that Microsoft has been building in technology for 50 years.

Like we just want to make sure that we are helping solve the problems that are emerging as this agentic web is happening.

Yeah, it's been it's really cool to see you guys like leaning really hard into MCP and like integrating it into all of Windows and all that kind of stuff.

That's really awesome.

And it brings up for me, I've been starting to hear rumblings from people who are thinking a lot about MCP that the security model needs a lot of work.

And I'm curious like for your take on that, because you're you're you're making a lot of comparisons between this stack and the Internet stack.

And the Internet has a bunch of like things in its security model, like the same origin policy that makes sure that like if a website is serving you, serving you code, like it's only able to execute on its own data.

And MCP doesn't really have that.

So like how do you what do you think the right security model is for this?

Well, look, I don't know that I know what exactly the right security model is.

Like the interesting thing about MCP is like it is so coherently simple that it's going to be relatively easy for the community to decide what the security model is.

It's like we have a bunch of like enterprise things that we care a lot about and that like we're doing really good work with the with the MCP team to get done.

And so like we need agents to have identities so that you can build entitlement systems.

So you can say this agent is acting on behalf of this person and like they are entitled to see these resources in this system.

And like even having a way for an agent to sort of go query a bunch of systems and say like these here's the thing that I would like to do.

These are the systems I need to touch in order to do the thing.

Like what entitlements do I need to ask for in order to do this?

And so I can request permission like from the user, like the person who's delegated this task to me.

Like can I present to them like can I have permission to these things in order to do this thing you asked me to do?

Yes or no.

And then for the administrators of all of these systems, you know, like, hey, is it OK for all of this to be happening?

And so like there's all of that's going to be like relatively straightforward to do, like not easy, but like relatively straightforward to do on top of MCP.

And like the important thing is like let's do it in an open way.

Like we don't need it to be proprietary to like our agents or our systems.

Like we've just got to figure out how to get this dime where things kind of work like the web works.

Yeah.

Well, it's an interesting question for me because I feel like there's there's maybe two potential models or potential go to markets for AI stuff that you guys have been talking about.

One is this sort of like verticalized model where you own the model and you own the UI layer.

You do all the applications and everything in between.

And the thing about that model, which maybe it's sort of you could say that the App Store, Apple iPhone model is a good example of that, is you can guarantee security in a lot of ways.

And an open model, it's harder to do security stuff, but you get much more like innovation basically because there's no central authority.

So how did you guys think about making that decision?

Yeah.

So look, I know that that's the argument that a lot of people make.

Yeah.

I think it might be a false dichotomy.

There is, you know, like a thing that you have in these open systems where they are permissionless.

Yeah.

And like there's a real advantage in having permissionless innovation.

So like the thing that is an individual that excites me most about what's happening right now is the extent to which you can go innovate and build things without having to seek someone's permission.

Like where you have to have them grant you permission for distributing your things to other people and, you know, like having all of these like complicated gatekeeping things that are sitting in between you who are the person who had the idea and like the people who might benefit from it.

Like I think some of these like middle layers that we have that have emerged over the past handful of years, like just aren't contributing much value, honestly, to the two parties in the transaction that matter, which is the person who did the hard work to make a thing.

And then the people who are going to either spend their attention or their money or some other currency that's valuable to like access the thing.

So like I get it kind of excited about that.

And so like that's one of the reasons why we've made the decisions that we want to.

But like I also think that there are ways that you can get real robust security in these systems, like leveraging some of the AI capability that you have now that, you know, where you may be able to have better security.

Like if you have an agent that you are running that is attending to your personal security requirements, like these are things I'm willing to share.

These are things I'm not willing to share, like, you know, and that has some kind of knowledge of risk assessment.

And like, yeah, like for instance, like my wife this morning while I was while I was getting ready to jump on stage, like I got this flurry of emails that because I'm the like backup security account for my wife, where like somebody was like fiddling around with two factor authentication on her account.

And like the first thing I did was I texted her like I didn't want to email her because somebody might have gained unauthorized access to her email account.

I texted her.

I was like, hey, are you screwing around with the configuration?

And she's like, yes.

And so like you could imagine like having an agent that is like privy to like a whole bunch of your communication modalities, like being able to like notice that something funky is going on and then like, you know, using a bunch of resources to triangulate like whether that's legit activity or illegitimate activity.

So like I think there's just a bunch of stuff like that that like you can have both, right?

Like I don't think it needs to be like, you know, one or the other as you framed it.

Yeah, that makes sense.

One of the one other thing that I'm I'm curious about is it just seems pretty clear that software engineering is changing.

Yeah.

And you're someone who's been involved in software engineering for a long time.

You're someone who also I think cares a lot about like the craft of things, like the craft of how things are made.

We were talking earlier about you're making you do a lot of ceramics.

You make your own bags like you're you love having your hands and things.

And I think one of the knocks on, you know, using agents for coding is like it gets rid of some of that feeling or something like that, which I don't necessarily agree with.

But I'm sort of curious for you, like as someone who cares about the craft of code, looking into this future of coding with agents, like how does that how do you feel about that?

Well, so let me start by saying that I.

I.

I.

I love the fact that my people and like when I say my people, I mean, you know, makers writ large, so software engineers or mechanical engineers or woodworkers or potters or, you know, just sort of pick your thing where people are trying to create things from raw materials or nothing.

If you are really passionate about what you do, you're going to have very strong opinions about how you do it, the tools that you use, the materials that you use, like how things get put together.

It is a necessary thing for you to be great at your job.

The interesting thing is people have lots and lots of different opinions.

And like one of the event you said, like, I've been doing it for a long time.

Like I am like I'm an old fart.

Like I wrote my first program when I was 12, which means I've been programming for 41 years.

And so the thing that you get to see when you've been doing a thing for a very long time is like this is not the first moment in the past four decades where the nature of software development has changed in a non-trivial way and where people have very strong opinions about the change and what it means.

And so, you know, I think the reality is that people will have choice.

You know, like I still when I go into a text editor, I probably shouldn't say this because like we make, you know, Visual Studio Code.

Like I'm such an old recalcitrant fart that I still use VI.

I use VIM at least.

But like my text editor of choice is like this extremely antiquated thing.

And like I just refuse to go use something different.

Even though I know for sure that that is suboptimizing a part of what I'm doing.

But I make the decision anyway because like I get to choose.

And then in other aspects of my making like either software or something else that I'm doing, like I will be like, OK, the important thing here is not the way that I'm doing this particular part of it.

It's the outcome that I'm trying to get to.

And like I'm going to use the most powerful or most convenient way to go get to the outcome.

And like, you know, I don't care who's going to throw rocks at me for doing it.

And like it's literally everywhere.

Like I've been a woodworker for almost as long as I've been a programmer.

And when I was a teenager, like the big debate was like, oh, are you a real woodworker if you use power tools?

Like, you know, real woodworkers only use hand tools.

Like there's still a little bit of that debate today.

But like the real debate is like, are you a real woodworker if you use CNC tools like computer controlled power tools versus just power tools?

And I understand it.

Like I actually think it's interesting, the debate itself.

But like sometimes like people are going to make, you know, different choices because like they value something different than you do.

If you value the process more than you value the outcome, like sometimes you'll make different decisions than people who value the outcome more than the process.

And I think the like, is it or are you a real woodworker, a real programmer?

You're sort of saying like you're only real if you do it the way I grew up doing it.

Yeah.

When you ask that question in a lot of ways, you know.

But it's just so varied, right?

And like so the thing that I will say is like I would never in a million years tell anyone not to have strong opinions about their craft.

Like have them.

Like it's great.

Like the advice that I would give people and like this is not me telling what it's just advice, like things that I found useful for myself is just have an open mind.

You know, when the tools are changing, I can't even tell you like the number of times where I have looked at a new technology in some other like non-software dimension of making that came along where I'm like, like I don't want to learn.

Like 3D printers, like I waited forever to learn how to use 3D printers.

And like I regret it.

Like I should have started earlier.

Like they're so damn useful for almost everything that I do.

And for a whole variety of complicated reasons, like I didn't let myself be curious about that, which is odd.

And so, you know, just like be curious.

Try stuff.

Like and if it works for you, use it.

And if it doesn't, don't.

Yeah.

What do you think is the future of software engineering agents?

Like is there going to be like one agent to rule them all or is it you're just going to use many different agents with different tastes or like how do you see that ecosystem shaping up?

Yeah, look, I think it's going to be a lot of different agents.

I mean, like and it's good to have a lot of agents.

And like we certainly with GitHub Copilot and the, you know, GitHub agent stuff that we're working on.

Like I think we will compete very hard to like be a tool that lots of people will choose because it's very useful to them.

But I think it's unrealistic in the universe of developers to like think that like every developer on the planet is going to snap to using like one tool for like an important part of their job.

Like part of the joy of being a developer is you actually have that choice and you can sort of choose and play around with a bunch of different things and like do irrational things and like do rational things.

And like it's just, you know, it is one of the very consistent things that I've seen over the past four decades of my programming life is like people choose to change their tools all the time.

What are the dimensions?

Do you have an opinion on the dimensions along which the different agents might differ?

Well, I think the most important thing about agents is like probably the product making part of them.

And so the most interesting startups that I'm seeing right now are not trying to innovate by building some kind of differentiated infrastructure.

They're innovating because they think they have an understanding of a problem that someone has that is better than anyone else.

And they think that they can pick up infrastructure or modify infrastructure or tune infrastructure to go solve their understanding of that problem in a world-class way.

So I think that's what we need a lot of right now.

And like that's what's going to dictate the diversity of agents and like how, like which things get used for what.

Yeah.

And I think honestly, because it's so much easier now to like have that nuanced understanding of what someone's problem is and to pick up these tools to go take a swing at solving it, that you're just going to have a lot of companies like building a lot of things trying to, you can even see it with the software development tools.

It's like crazy how many things have come out over the past year.

And like they're interesting, like all of them.

Yeah.

Yeah.

It's a lot to respond to when you're a company building software development tools yourself.

But like it's super, super interesting.

And it's like what we've seen is if you've got some kind of nuanced understanding of what someone needs, like people have like high tolerance and like high interest in like giving things a try.

Yeah.

So we're almost out of time, but I'm curious, like let's say it's a year from now.

We're back at build.

Yep.

What are some things that are a hot topic right now or big questions that people have right now that are not going to matter in a year?

And what is going to matter in a year?

And what are your predictions for what we're going to be talking about?

Yeah.

I think people who are still like hanging on to these ideas that, oh, like the technology is not ready yet because like I, you know, I tried to do something and it was like marginally too expensive or like it was marginally capable of, you know, doing the thing that I wanted to do.

Like I think anyone who is using those as excuses to wait to get started are going to be super behind because everything's going to get cheaper and everything's going to get more capable every year.

I think this is actually not a hard sell in 2025.

Yeah.

It was this loud chorus of like, oh, you know, the progress is going to, you know, it's about to end and like everything's going to stop and like everybody's going to be super disappointed.

Like I, I mean, there's still some people out there saying that, but like, I don't think, you know, folks are paying much attention to them anymore.

Mostly because like, you know, what, what, what do you win by paying attention to like some crank who's saying like, you know, the, the thing's about to stop.

Like you're sort of betting on failure and the cost of betting on failure versus betting on optimism is like a real big difference there.

So, yeah, like I, I think we're, we're gonna, uh, we're gonna see a ton of progress on, uh, like the level of ambition of problems that people are tackling with agents.

And then I think, you know, modalities that are going to be really different is like, is this agentic web starts to get more complete, more plumbed out, uh, and the models reasoning and planning capability get better and better.

Like you're going to start to get to the point where you're able to go from this synchronous mode of interaction with agents to asynchronous.

Like right now, most of what people do is like, they sort of sit down, like they got a thing they want to do, like the issue of the prompts, uh, like, and they wait until, you know, the thing comes back, uh, and like, you know, do something with that response.

And so I think, you know, by next year, you're going to see people using these agents to like, Hey, go sort this out.

And like the agent is going to take a lot of time.

It's going to go make a lot of calls out to systems, like the things that it's sort of, uh, like taking action on or going to like take a while to come back.

Then they're going to like integrate all of those responses and like do something.

And like that whole thing may iterate a bunch of times.

And then, you know, at some non-trivial amount of time later, you're going to sort of say, okay, like, here's the, here's the, here's as far as I got.

Like now it's your turn.

Like go take some action now.

Sounds like a future I want to be in.

Yeah.

Yeah.

Me too.

Right.

Um, well, Kevin, thank you so much.

Uh, it was really, really great to talk to you.

Yeah.

Good to talk to you as well.

Thank you for having me on course.

Oh my gosh, folks.

You absolutely positively have to smash that like button and subscribe to AI and I.

Why?

Because this show is the epitome of awesomeness.

It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure unadulterated knowledge bombs about chat GPT.

Every episode is a roller coaster of emotions, insights, and laughter that will leave you on the edge of your seat craving for more.

It's not just a show.

It's a journey into the future with Dan Shipper as the captain of the spaceship.

So do yourself a favor, hit like, smash subscribe, and strap in for the ride of your life.

And now without any further ado, let me just say, Dan, I'm absolutely hopelessly in love with you.

Fancy.

That's how you know it's the real deal.

Yeah, excellent.