Pragmatic AI with Matt Stauffer

The Pragmatic AI Podcast is sponsored by Tighten.

T-I-G-H-T-E-N.

We will take your AI ideas, prototypes, or even vibe-coded apps,

and we'll take them to production.

Scalable and secure. Check us out at tighten.com.

Hey, and welcome to Pragmatic AI, where we talk about using AI in the real world.

What works, how to use it well, and when it causes more harm than good.

Practical tools and real trade-offs for builders and business leaders.

My guest today is my friend Kent C.

Dodds, the Kent C.

Dodds trademark, registered, whatever, an independent software development educator.

And it's funny, because the phrase independent software development educator sounds like
you're just some dude, but you're not just some dude.

You're one of the preeminent teachers of technological things on the internet and also
very respected in the world of AI.

So before I kind of blab about you anymore, Kent, would you mind just saying hi to the
people and talking about who you are?

I'd love to.

Thank you.

You're so kind, man.

Yeah, so for the last over a decade, I've been trying to help make the world better by
teaching people how to make their corner of the world better through software.

And so I live in Utah with my wife and six kids.

I, yeah, it's a great time.

We love it here.

I have taught all kinds of things.

I focus primarily on the JavaScript and TypeScript ecosystem.

I'm very much a web guy.

I also Node.js on the back end primarily.

And yeah, I started out teaching on other platforms like Egghead.io and Frontend Masters.

Eventually decided to go full time with that, so I was working on the side as well.

decided to go full-time teacher, made my own platforms with actually the same team that
makes Egghead.

ah And so we made testing, JavaScript.com, and Epic React.dev, Epic Web.dev, Epic AI.Pro,
and now we're doing Epic Product.Engineer.

And so, yeah, that's all about me.

Yeah, and so, you know, for listeners who are not programmers, one of the things I've told
you all a lot is I want to make sure that while I live in the programming world, and I

think the programmers have a lot to say about this, we want to make sure that these things
are accessible to non-programmers.

And I don't know why, but I just feel motivated.

Like, Kent, I feel like when I follow you and I follow your life, I do learn a lot about
the program that you're doing, but you were always talking about, I mean, you're sing in

choirs, I think.

You're a one-wheel aficionado.

You're obviously super involved with your family or whatever.

So what...

What, like if I were to just see you on a random Saturday, like out in your town, what
would you be doing?

You know, are you an outside, outdoors guy?

Like I feel like that's the vibe I get from you.

Yeah, well, my kids are entering teenage years and so uh most of the time I'll be driving
them around to different things.

ah Yeah, yeah, but like we spend a lot of time uh outside picking weeds in the yard and
like just spending time together.

We'll go out and do fun things.

We'll stay in and do fun things.

um

And yeah, pretty much like I do occasionally go out on my own to ride one wheel in the
winter time.

I'll go snowboarding and I'll take my kids with me.

There's a theme park, like an hour drive north of me that I like every year I take my kids
one at a time.

So it's just one on one time with dad and we go ride roller coasters and stuff.

So yeah, it's pretty much like I'm either developing software, teaching people how to
develop software or hanging out with my family.

And that's, that's pretty much all that I do.

I love it.

And so you know, like one of the first things I asked at the beginning podcast is like,
what do you do and how is what you do interacting with AI?

And obviously you literally mentioned Epic AI as one of the projects that you're working
on.

But I am curious to hear kind of like, we all have our story of what it looked like for us
to go from not using AI to using AI.

And it usually is a multi-stage story.

So would you mind sharing with us as a developer and as an educator, like what was the
process for you of starting to adopt AI?

Were you a late bloomer?

Were you early on it?

Were you critical at first?

What's that process been like for you?

Yeah, uh I would definitely categorize myself as an early bloomer or early adopter.

em so ChatGPT comes out in 2022 of November, I think.

And of course, that was the inflection point.

uh AI has been a thing since, I don't know, was it the 70s or something?

They've been doing artificial intelligence things for a long time.

And I'm very much interested in Tesla full self-driving and Waymo and what they're

doing, so like way back with Google.

In fact, when I was in high school back in the early 2000s, I wrote a blog uh about Google
and what they were doing just for the fun of it.

um so I remember early, early, like there were discussions about self-driving cars.

so like this AI has kind of always been in the back of my mind as like,

I don't really understand what that is or what machine learning is.

And I don't know math well enough to really get it, but like it's as a consumer, it's a
really exciting thing.

And I'm very much looking forward to this.

Not to get, I guess, too personal, but my brother actually died in a car accident when I
was in elementary school.

And so ever since then, I'm just like, how quick can we get to removing steering wheels
from cars?

Like I just, I don't.

It's insane to me that we're driving down the road with hundreds of other people and we're
just hoping that one of them doesn't get distracted or like have a death wish or

something.

um And so I'm very much looking forward to that.

So I'm always very bullish about what AI can do to improve the safety of our roads, but
also just improve life for people in general.

And we've seen that like...

um

Being able to index the entire human genome and all of the advancements in medicine that
have come from that, there are obviously dark sides of AI and everything, just like

there's the dark side of the internet.

There's always that side of things.

But I choose to be optimistic.

And I see a lot of really exciting things.

So when Chat GPT came out, I was very much an early user, playing around with it, doing
interesting things with it.

And then,

GitHub comes out with their copilot.

OK, now I've got AI right in my editor.

I don't have to copy paste my code from Chat GPT anymore.

I can jump right into my code editor, and it's right there, and it has all the context.

And then it just went from there to getting to where it is now, where I don't look at the
code as much, and I certainly don't write the code anymore.

And it's been a pretty exciting journey as a software developer.

And you're one of the more kind of pro-AI folks.

And I've said this before, but like I'm trying to get anti-AI, pro-AI and everything down
the middle.

And I'm not saying you're just like an unthinking maximalist, but you are a little bit
more pro-AI.

But there's nothing wrong with that.

And I'm actually fascinated.

My wife and I were driving down the road, down the interstate.

So I live...

right off the interstate.

So even though I'm in Atlanta, I have to be on the interstate to get almost anywhere else
in Atlanta, just based on where I am.

So I spend all this time and I'm thinking about when my son who's 13 will eventually be
driving on the interstate to get to school every day.

And that terrifies me.

And so then I'm on the interstate looking around and I'm like, we're all driving 80 miles
an hour on five or six lane traffic for hours every day.

It's one of the most dangerous highway segments in the entire United States.

And that kid over there could be 16 driving that F450 or whatever.

You know what I mean?

I'm just like, this is not

This is not ideal and I for some reason I hadn't paired that together with the benefits
that come from something that is not also on the phone and also listening to music and

also tired and also their brain hasn't fully developed a weatherbird.

Like your full time job is just driving cars safely and that's all you're supposed to do.

I'm like that's, I'm not that, you have made the most compelling pitch for self driving
cars that I've ever heard in my entire life.

Oh, well, yeah.

So like, I often will tell my kids, years ago I said, you will probably get a driver's
license, but you probably won't ever use it.

And I still feel that way.

My oldest is 13, so a couple of years.

um I do feel like we'll get to the, uh we're very close.

I have a Tesla and it drives me everywhere.

Like I rarely am not using self-driving.

um

And like the compelling thing for me, yeah, sure, it'll be nice to be able to look at my
phone while I'm driving, but the really compelling thing is that everybody else is not

driving.

Wow.

Yeah, I appreciate that.

um OK, so you kind of talked a little bit about what your journey to AI was looking like.

So I'm always interested in what everybody's day to day interaction with AI looks like.

And obviously, you're educating people about AI and you're educating them about using AI
in their development.

In the development world, there's a lot of folks who are pretty critical about our
ability, actually in every world, our ability to trust the output of AI.

And so one of the things you mentioned was your interaction with the code that you're
writing is changing as you use AI with it.

Are you at a point where you would ever commit a line of code that was AI generated
without you having reviewed it and made sure you fully understood it, or are you using AI

in languages you don't understand yet?

Or is it all like, I don't maybe necessarily write it, but I'm gonna review all of it.

Yeah, so to add some context to this, I am independent.

I don't have an employer.

any um subscriptions that I have or whatever are paid by me, my company.

um That said, I am uh pretty influential and I feel like I have a pretty good opinion on
things.

so companies will reach out to me and not want to sponsor me necessarily.

Actually they do and I say no, I don't do sponsorships.

I will take free credits to try out their stuff and give them feedback.

And so one of those is VS Code, like from GitHub.

uh Also, I have same thing from uh Cursor, and that's the editor that I prefer most.

And so they give me a bunch of credits, and I can use all of their features.

And so um this is important framing, because lots of what I do is actually quite
expensive.

And so a lot of people can't reasonably do this.

Their job maybe has enough

that they'll just say, use whatever number of tokens you want.

But yeah, a lot of people can't.

With that framing, I'm max, I'm like, uh I use all of the features.

And so what that means, well, one other important piece of context is a lot of how careful
I am with my software since before AI depends largely on the ultimate outcome that I'm

looking for.

Am I building this?

medical device that doctors are going to use on their patients.

Like, okay, I'm going to be really careful about that.

Am I building a personal side project?

I'm not even going to test it.

Like whatever, who cares?

You know?

And so we, that same framing applies with AI.

um If I am building something that's really critical for people, then I'm going to be a
lot more careful.

As an independent, I don't really build software for, you know, medical professionals or
airplanes or whatever.

And so in, in general,

uh I do build software that people use.

I've got lots of users of my personal website is actually a lot more than you might think
when you say personal website.

Like I've got a podcasting platform on there and uh I've got, there are user accounts that
can do all sorts of things.

It's like a whole big thing.

So I do have uh tens of thousands of people who use my personal website.

that like that does.

I do build software that actual people use.

A lot of people see Educator and they're like, he just builds demos all day.

That is false.

um But then I also have other things where I'm the only user.

So anyway, with all that context, I definitely ship stuff all the time, every day.

Thousands of lines of code changed that I don't look at any of it.

And the way that I can manage that is because um I use two tools to do my reviewing.

One is CodeRabbit.

And the other is uh Cursor Bugbot.

in fact, Cursor is building the software.

So often, the code that I'm shipping doesn't even touch my computer because I'm using
Cursor's Claude Agents to, like, I'll plan something if it needs to be planned, if it's

like an obvious, like, just add a checkbox for this feature or whatever.

can just, one sentence sometimes is enough and it can go figure it out.

And so I'm doing that, like, often from my phone.

when I'm waiting to pick up my kid at gymnastics.

I am shipping an insane amount of stuff.

And then it creates a pull request, the AI agents will review it for me, they'll go
through a bunch of iterations, and depending on how much I care about this feature

actually working the first time, that will determine whether I'm gonna look at it myself.

If I'm okay with it not working, if it's just an inconvenience, if it doesn't work, then
I'll just merge it and then I'll fix it forward.

If it really matters, like I've got my workshop application that runs locally on people's
machines, and that's how they go through my workshop material and do some intentional

practice and stuff.

ah That I do have users, and they paid for use of this.

That matters more to me than my website, which is just a free thing.

So ah that I'm a lot more careful.

I'm going to look at the code.

ah And then there's something in between where it's like, I kind of care about this.

It would be kind of annoying if I got this wrong.

So often the most important thing is the data migration and like, did they get the data
structure right?

Cause if that's wrong, then lots of the other things are wrong.

So that's kind of where I'm at.

um It is a bit of a spectrum.

And then also like for some really big things, I'll go back and forth with the agent a lot
and talk about how we want this feature to fit into our product.

Cause we're product engineering now.

Like we got to think about how this fits into the whole picture.

And so I'll go back and forth with the agent a lot and then

We turn that into an actionable plan and I'll send that off to a cursor cloud agent.

So that's where I'm at.

Well, I love that.

mean, and I'm I am a owner of a consultancy where people pay us to do expert level work
that doesn't break.

And so we're treating the code, writing their very not a line of code would hit production
that we have not seen and understood.

And usually if there was an agent involved at all, probably fixed.

And then I also am doing stuff on the side.

I'm making little toys and platforms, trying new technologies, sometimes for my own
personal use and sometimes for promotion of Tighten

And it's exactly what you said.

just like, this if I launch this and it breaks, I just fix it, you know, and I have that
exact same attitude and I had not actually named that being the case until you just said

this.

I'm like, yeah, very, very.

It's not even just medical devices and planes, but it's also like somebody else paid money
for this and expects it to work.

And they're paying us to be the experts who ensure it's good.

I'm like, that's a very different perspective versus I want my wife's blog to be on the
Internet.

Right.

And like it's just and I want her blog to be excellent.

But the code architecture of that is not the big deal.

It's it's still in the early stages phases.

So nobody's even looking at it.

But the two of us, it just doesn't matter.

And there's so much more we can do if I'm just like until she's got a million people
looking at it every minute.

It's it's just fine if I just push, push, push, push, push, push, push, push, fix the
things as they come up.

So that's helpful.

and the other thing that I think is important here too is that we are experienced software
engineers.

And so, like, even without looking at the code, just looking at the summary that the AI
gives us, like, here's what I did.

Oh, okay, I know that's going to be a problem.

And so, like, let's iterate on that a little bit.

So speaking of being experienced software engineers, one of the things that I've talked to
a lot of people about is that uh we are able to prompt and review in a way that someone

who has never worked with technologies before.

is not going be able to.

One of the questions I asked you was, you using in technologies you're not familiar with?

uh You're an educator.

What concerns, if any, do you have about the future of people learning things when they
have AI to help them?

Are there things you're putting in place so you're saying, I'm making sure that my people
are learning this particular way?

What do you think about the future of education and people onboarding into things as
juniors or apprentices given AI?

Yeah, this is an enormous question that I've been thinking about uh extensively.

And there are a lot of different ways that I can go on this.

So the first thing I'm going to say is,

We as experienced software engineers have a lot of past experience that has faded into
irrelevance.

Let's ignore AI for now.

um I learned AngularJS that is completely irrelevant now, like as a specific technology.

um Webpack, same sort of thing.

And features of JavaScript that I don't use anymore because we have new features of
JavaScript that I use instead.

So we're constantly um sifting through all of the experience that we have.

and new software developers join up having never learned that stuff and being no, worst
for it.

um The things that really matter from an experience perspective is like that underlying
stuff that you never intentionally learned, it just kind of happened as you were going

through this, like, oh, I shipped this broken thing and this is a broken migration and now
I gotta go fix it.

And then like, just kind of in the back of your mind goes this a little bit more of
experience.

And that's actually something that...

My Epic product engineer thing that I'm working on is trying to uh help people acquire
that level of experience through intentional practice.

So a lot of people ask me, well, Kent, should new software developers just turn off the AI
assistant when they're learning how to build software?

I think that's the worst idea.

um Think about early on um when we were starting to get code editors and everything, let's
say, oh, we've got this new code editor.

it has this cool feature called syntax highlighting.

So you can easily tell what's a variable declaration versus a function or whatever.

But you new developers, you should not be using syntax highlighting.

You should go with notepad.exe and just build the way we used to build because then I
can't even finish that sentence.

Like, why would you do that?

It's ridiculous.

Yeah, yeah, exactly.

And so I see it as exactly the same thing.

you are not gonna not use AI to develop software.

It's just, you're not gonna do that.

You're just not.

And um certainly like having the understanding of what is let versus const versus var in
JavaScript helped me at some point, but there's like zero help that that offers me now.

And so um I do think that a new developer getting some foundational understanding of what
programming is and all of that is useful.

And so I actually shipped

a practical TypeScript course on Epic Web that takes people from foundations to fluency in
TypeScript.

And you can do that in any language, just like a fundamental understanding of how
programming works.

I can't make a really solid argument for exactly why, but it just feels to me like having
a foundational understanding of programming is probably a good idea will help you use the

agents better, even as they get increasingly better.

um But beyond that, I really don't.

don't think that you need to worry about that.

So that's on the beginner side of things.

um about like as far as the education space and how that changes, I mean, obviously it
changes an enormous amount.

So my whole business has always been accelerating the acquisition of knowledge for
experienced software engineers who just happen to not know this.

So you don't know how to test, let me teach you how to test.

You don't know React, let me teach you React.

really need that anymore?

If you're an experienced software engineer, it's that lower level, behind the scenes stuff
that is what is really going to help you.

And not only that, maybe agents coding uh AI assistants um will get good enough that you
don't even really need that as much.

They can slowly erode at the necessity of that past experience.

I'm not predicting the future or anything, but I kind of see a future where that could
possibly be the case.

And so what's left, to me, what's left is, um like, let's use a metaphor of a quiver of
arrows.

Like, you've got your agents and your quiver of arrows, and now you're pointing it at the
target.

Yes, your ability to aim at the target properly and hit the bullseye is useful, but let's
say that these arrows are getting better and better, and now they're more like homing

devices, and they can just go right to the target.

So now it's more a matter of uh are you shooting the right target?

And uh that is what I'm calling product engineering.

So you understand what the um users actually need.

You understand the problem.

You have some clarity on that problem.

You have a good feedback loop on making sure that you're solving that problem well.

That is the skill that I think was useful 50 years ago and will continue to be

useful. I'm not sure that agents will be able to replace

that. um And so that's why

my shift because in the education space, it's impacted everybody.

Like, Laracasts and everybody has been impacted by this.

And so where a lot of my peers are kind of going into, all right, I'm going to start
teaching uh AI workflows, development workflows.

Let me teach you how to use Claude Code.

Let me teach you how to do these.

Five months ago, I was using these tools completely different from the way I use them now.

And it's because the tools are getting better.

And I want to teach people really durable skills, like things that will still be useful
beyond five months.

That's always what I've tried to do.

And so I don't think that that's the right direction for me.

um Instead, I want to teach you something that I think even if agents get so, so good, um
let's assume that they just can't quite knock our industry off of its feet.

What's that last skill that they won't be able to replace or the last skill that they will
be able to replace, but it's the last one.

And I think that is product engineering, knowing which target to aim at.

And so that's, yeah, that's how it's impacted education.

Yeah, that's a lot, but that's really helpful.

uh Recently, Justin Jackson, a shared friend of ours, put out an article where he was
talking about how he's seeing a lot of teams where you've got product, design, and

engineering, and all of them are encroaching on each other's spaces because of AI.

Your product person now says, I can do design and engineering.

Engineering says, no, can do design and project.

everybody's kind of moving into the middle.

um Are you worried about that?

Do you have a fear that there's just going to be less slots for everybody or are you more
optimistic?

You're just, ah, we'll just build more stuff.

Like, how do you feel about what this means for the future of our industry?

Yeah, so um I do believe that we'll build more stuff.

There's so much software that still needs to be made.

Like you might not think that in the tech world, because you're like, there's software for
everything and I can just build my own software for, but like you look in the non-tech

world, uh everybody hates their software that they use.

Like they just hate it so much.

So there's so much software yet to be created.

So I actually feel pretty optimistic that there's still jobs to be done.

The thing that we are losing though, is if you loved coding because of the craft of
coding, rather than the problems to be solved, then you're gonna be disappointed.

Because like, yeah, that's not all that useful anymore.

um The salary that I pay you is a lot less than the salary that I pay these tokens um from
the AI.

And uh it might not be quite as good or whatever, but maybe it's good enough and it's
getting better.

Yeah.

you're very replaceable if that was the thing that you cared about.

So people who really care about solving problems and exploring the problem space,
understanding and trying out different solutions, they're going to be just fine.

And for the foreseeable future, I do see a difference between a product manager who does
that on the people side and a product engineer who does the technical solution

Maybe eventually that technical side gets eroded, but we're not there yet.

And I can't really say with any confidence that we will or will not get there eventually
or how soon that would happen.

I hope that we don't get there because then I don't know what we're going to do.

But for now, I'm optimistic.

yeah, so like designers, engineers, managers, I actually do think that there is still a
distinct

difference between what they're doing.

But there are aspects of all of those jobs that are being eaten by AI.

And honestly, think for the most part, for most people, those parts of the job were the
part they hated most anyway.

Like who liked arguing over syntax in engineering?

The worst people liked arguing over syntax in engineering.

Yeah.

like, yeah.

Yeah.

great stuff,

And then, I'm not a designer, I'm not a product manager, but I imagine in design it was
probably really annoying to go back and forth with people on a couple of pixels here and

there, or like, you know, make it pop, or whatever.

Now your client can come to you and say, hey, I talked to Chat GPT and iterated on a bunch
of ideas.

Here's an idea, I did this for my logo for Epic Product Engineer, actually.

I iterated with bunch of ideas, handed that over to my illustrator, and then like,

We didn't go back and forth.

He delivered, his first idea was exactly what I wanted.

so I don't know if he loved that part of his job or if that was kind of an annoying part
of his job.

Like hopefully he didn't hate that I did that.

But I liked it as a customer because I don't really like going back and forth a lot.

know, with the timeframes and everything.

Okay, he sends me the email.

Now I send him a response and now I'm waiting.

then with Chat GPT, just threw it really quick.

And then I hand it off to him.

And he uses his, not only his skills at developing those things in the first place, but
also like at understanding what my vision is and taking this concept and turning it into

something that I love.

And I really think it's great.

So I think that the parts that AI is currently eating in our jobs are probably the parts
that the very best and skilled people in those domains really don't really enjoy doing

that much anyway.

At least that's the cope from me.

Well, that reminds me of three things.

Hopefully I'll remember all three of them as I go.

The first one is you have not heard it yet because it hasn't come out yet, but we had
Steve Schoger on the podcast as I think the episode right before yours.

He's a prominent designer and he has talked.

publicly about how he's using AI as a part of the design process.

And I was asking the same question.

I'm like, do you feel like you're missing out on things?

He's like, I don't miss laying things out and turning on auto layout and Figma and all
these very implementation detailed things.

He's like, I don't miss those at all.

He's like, it's not as if it's taking my vision as a designer away.

I'm still the creative.

I'm still the art director.

I'm just no longer the implementer.

And he's like, I don't miss that at all.

So there's definitely a case to your point.

uh Also at Tighten we recently had a client who went back and forth with Lovable, hundreds
of dollars of credits with Lovable to get an MVP in Lovable that would never stand up to

the light of day for his client base, but allowed him to do that very rapid iterative
process without paying us to sit there with big whiteboards for him.

And he comes to me with that, and then we whiteboarded in front of that.

And I helped him translate that into an actual true MVP.

took away the things that didn't make sense, helped him think through the data model, and
then we built an app from there.

And the work we were able to do from there was no less fun for us.

We're still creating, we're still bringing our expertise to bear, but he was able to have
a much more affordable process by doing that initial feature definition.

So I think that's very similar to the process you're talking about, and I'm in the case of
your illustrator there, and we had a great time.

Like it was wonderful project, really happy to ship something for him in a budget he
wouldn't have been able to do otherwise.

So.

Yeah, yeah, it's way cheaper.

So we talk about how the AI is really expensive.

Yeah, sure, it's expensive.

But relative to the outcome, it's way cheap.

It's so cheap.

It's crazy.

Yeah, well, my.

My expectation is actually, like there's a lot of talk about the data centers and
everything and there's electricity problems.

Yeah, yeah, sure.

Like these are problems.

Humans solve problems.

Like these problems will be solved.

And not always are the solutions to the benefit of people.

Like I definitely feel for people whose electricity is going up and everything, keep
pushing back.

Like that is fine.

We will find solutions to these problems.

But as these data centers come up and more investment is in the space,

I think that the cost will come down.

And um luckily, there is competition in this space.

Because if it was just OpenAI or just Anthropic, then yeah, I might be a little bit
worried.

But no, there are compelling reasons to improve the models while bringing costs down.

So I do expect that uh costs will be more approachable in the future.

Yeah, and to your point, in a lot of our world, I know this isn't, I just had um an
educator on it, she said perplexity is what they're using for everything, but in a lot of

the programming world, it has felt like Claude has won, and I know that you have been a uh
fan of cursor and Codex uh

And I'm delighted by that for exactly the reason you said, I don't want one big winner.

I already wish we had more competitors in the space because the more competitors there
are, the more that the humans are able to vote with, you who they support.

And the more there's just one big player, the more we're all stuck with whatever they
want.

So every time somebody gets really excited about something that's not whoever the leader
is at that moment, I'm like, great, let's let's continue rooting for the underdog.

So it stays a little bit more of an even playing field, you know.

Yeah, yeah, because Cursor sponsors my tokens, I don't really have to care quite as much
about the cost.

But when I saw how much more expensive Opus was over ChatGPT 5.2 at the time, I was like,
oh, goodness, I'm going to try ChatGPT or GPT 5.2.

And I was like, I'm actually pretty happy with this.

And then they came out with 5.3 and now 5.4.

And I actually feel like it's better.

um And it's a lot cheaper.

Cursor also came out with Composer as another model that they built on top of Kimi.

Or Kimi, I don't know how you say it.

But it's even faster, even cheaper, not quite as good, but like...

So I used to really hate having to decide which model to use, and I still kind of do.

playing around with the models and saying, this task is definitely something Composer can
do, no problem, and I want it to be fast, so we're going to use Composer, we're going to

be cheaper.

I do try to be a good steward even though my tokens are paid for.

So I do use Opus sometimes, but pretty rarely, mostly because it's so expensive.

Yeah, that makes sense.

I have so many other things I want to go to for you, but we're literally already half an
hour in and I have to at least make sure we take some time with MCP because you know for

those who don't know Kent is one of the the premier teachers about a concept called MCP
that I think is very misunderstood on the internet if an understood at all I would say

there's a really good chance that half the people listening have never even heard those
three letters put together before so Assuming that a listener has never heard of MCP

before can you give us the introduction to what is MCP?

And why are you a fan of it?

Yeah, so MCP stands for the Model Context Protocol.

the short version is it lets your AI agent do things.

So back in like 2022 when ChatGPT came out, we were entered what I call phase one.

So now you can send a text to a large language model and it will send text back.

And at the time, like it's hallucinating, it's making things up, like, you know, we're all
laughing out loud about some things.

But it actually was pretty useful, especially if this knowledge was in its training data.

Then you could say, who was the president of the United States in 1942?

And it would be able to answer that correctly.

I could not answer that correctly.

But yeah, so it was actually useful.

And then people started saying, hey, what if I take this snippet of code and paste it in
and say, find the bug?

And like, yeah, it could.

So then I take the code that it generated and put it back in my project.

And now there's a new bug, but the one that it fixed was fixed.

um And over time, this got better, though.

m But that process of copying things over was kind of annoying.

And so now we put the models into our editors.

And now the editor has some special contract with the large language model to say, hey,
let me know when you want to change a file.

And I will go and change it for you.

And you just tell me what you want to change it to.

And this evolved into something called tool calling, where now the large language model,
it's still the only thing that a large language model does is generate text.

That's it.

um And so it takes your input and it generates what's the most natural next thing to
appear.

um And so we add the special syntax for that language model that says, uh if you want to
do something here, like here's a list of things you can do.

If you want to do it, like use these special characters or whatever, generate these
special characters so then I can...

parse that out programmatically, and then make some sort of tool call to do something in
the world.

And they use this also to do search queries.

And so now it doesn't have to be in the training data.

You could say, or what's the weather like today?

And it can go and make a request and come back to get that answer.

um this also evolved into chat GPT plugins like a year and a half ago.

And so you can make your own ChatGPT that was connected to your own data.

You could upload documents and now you can ask questions about those documents, stuff like
that.

um There were a lot of problems with this.

uh ChatGPT plugins didn't really take off and they're also, and like the biggest reason is
that they're very proprietary.

And so if you build one for ChatGPT, well, now you got to go build one for Claude and now
you got to, you know, if you want it to be generally useful.

um so Anthropic sees this and also like to be frank,

Anthropics behind OpenAI, especially in the consumer space, by like an enormous margin.

OpenAI has like 80 % of um the consumer market.

So um from Anthropics perspective, they're like, well, I sure hope that uh chat GPT
plugins don't become like the way it's done because that's like the iPhone I store or the

app store, right?

Like we're a completely different thing.

So they're like, what if we make a standard way to do this?

that is implemented, that's like open standard.

Like OpenAI can participate, Google, Amazon, Microsoft, they can all participate.

We can build this standard.

And what that allows is a common language.

It's similar to the browser, how like I can make a website and I can put it up on a server
and that website sends back HTML.

Your computer makes a request, sends back HTML.

And I don't have to go to Google or to Apple or to Mozilla and say, hey,

Could you integrate my website into your browser?

Because I've written to a standard HTML, and all of those other vendors have also written
to uh that standard to render that HTML, to show it on the page.

And so as a result, um I can now publish any number of websites without having to talk to
anybody, and that integration just happens automatically.

That's just how it works.

And so MCP was the same solution for LLMs to say,

Hey, we've got oh this capability that is emerging in large language models to be able to
call out to tools to gather more context or perform some change in the world.

if we have a standard mechanism for doing that, then developers can create a server that
talks that protocol.

And then any other agent can connect to that server, and that tool calling can happen.

And the other really good thing about this is that there can be a standard authentication
mechanism.

So now I can do things that require a particular user and with particular permissions and
whatever.

And by having that be a standard, you no longer have to um cobble something together where
it's like running on your machine and it's like the uh agent has full control over

everything that your machine can do.

Instead it can be on like a very specific protocol.

that has a specific permissions mechanism.

And the mechanism that was chosen was perfect.

It's uh OAuth, which is a standard protocol for uh managing authentication on the web.

And it works really well.

uh So that is the whole idea of what MCP is all about.

It's uh a standard mechanism for uh a large language model to actually gather information
from the world and put information out into the world.

Yeah, that was incredible.

It's almost like you teach about this every day.

uh So let's talk about some practical use cases.

m Are people using MCP servers in things like chat, in chat?

Because I'm used to MCP servers in development environments.

Are people using MCP servers just to add functionality to their chat GPT?

Yes, they are.

So there's a whole history that I can give you.

I don't think we need to go too deep into how this evolved.

But where we are now is that if you go to claw.ai or chatgpt.com, whether on your phone
with the mobile app or in the browser, uh you can connect to MCP servers.

And they've got marketplaces to do that.

You can also connect to custom MCP server.

And um what is...

uh

Developers are actually really used to MCP servers because our editors also have support
for MCP servers.

And uh also like Claude Code and these agentic uh editors as well.

um And what's interesting about that is developers that were developing on a computer.

And that computer has a capability of doing just about anything.

And so uh if you are permissive enough, your uh agent

can also do all those things.

It can use the command line.

It can send commands to different programs.

And so a lot of developers are looking at MCP.

And they're like, I don't get what the big deal is.

You just added a bunch of steps to what is simply like a shell script that can be run.

And I can do anything.

I'm logged into GitHub with the GH uh CLI.

And I've got my other CLIs for other things.

So I don't get what the deal is.

And for those people, I say, yeah, I'm with you.

That is not why this is interesting.

What makes it interesting is that it works uh in the browser, it works in these
consumer-based apps, it works from my phone, it works anywhere.

And I don't have to have a CLI installed, like I can install a CLI on my phone.

uh And what makes that possible to work in all those places is the authentication model
for this, that it's a standard mechanism for giving permission for the agent to do these

things.

So in answer to your question, yes, and I am um going way deep on this.

And my own AI assistant, named Cody, which is the koala, it's my mascot for my business.

um yeah, Cody is an MCP server that exposes just three tools for um doing just about
anything.

can do, like my shades in my house are controlled by Cody.

And Cody...

decides at what time of day, based on the position of the sun, should shades be open.

oh and it's a hot day, so we're gonna keep these shades closed.

it's a cold day, I'm gonna keep these ones open.

Oh, at this time of day, there's a glare on the countertop, so I'm not gonna open that
shade.

now it's overcast, great, I will open up the shades.

It's doing all of this, and it's actually doing that through...

through deterministic code that Cody wrote.

And so now I'm not actually paying inference for all of that either.

It it wrote the code once and now I could go so deep on Cody, but Cody is only possible
because of MCP.

And I can use Cody through ChatGPT, through Claude, through cursor, through like any of
these because of the standard protocol.

It's the same, like, maybe that sounds interesting, but, or maybe not, but like it's the
same sort of thing of.

I shift a website, it's kentcdodds.com and you can visit it in Safari, in Chrome, in
Firefox, in Edge, like, okay, yeah, of course you can, dummy.

No, it's really important that it's a standard because that's what enables that sort of
thing.

So that's why when I first saw MCP just over a year ago, I looked at that and I'm like,
that is the future.

That's how we build Jarvis from Iron Man.

I'm like.

This is how you make the LLM do a thing.

It's MCP and that's why I dove so much into it.

Yeah, so for those who aren't programmers in the programming world, have historically had
things called there's two main ways where we consistently programmatically meaning like a

computer program calls it.

interact with other data and they are CLIs so that is a command line tool if you've ever
seen people in the movies where they're like typing and it's like green text on a black

screen like that kind of interaction that's not just for nerds that type of stuff is very
predictable in your interactions and so computers can run against those things very easily

because it's you know just run this command with these flags and it's a single command and
it's going to do the same thing every time whereas if you imagine going to website click

around that website it's very hard to teach a computer how to click around a website and
then the other thing is APIs which is

Similar to just sending a command, but you're sending the command over the internet and
setting certain that command in the little terminal But again, it is a defined spec.

It's it's uh

predictable, it runs the same way every single time, and somewhere there's a specification
saying, here's how to do these things.

So the simplest way I've been able to understand MCP as it is just another standard, but
this standard is structured around the best ways for AIs to communicate, LLMs to

communicate, rather than APIs, which were for web applications to communicate, or CLIs,
which are...

humans and you know local computers to communicate.

What does MCP add that a OAuth API doesn't already give?

Or if they add the same thing, what's the benefit of choosing to build one versus building
the other?

um What it adds is the standard.

um That's like primarily the thing.

If you look at the APIs that are exposed by servers all over the world, they're all going
to be so different in how you communicate with them.

Some of them are going to be stateful.

Some of them are going to, um you're going to need to do this like other auth process to
get through that.

If we say, okay, no, it's going to be standard rest with OAuth.

Well, okay, you're kind of building MCP now.

um

And then there are specific affordances from MCP that make it um communicate with the
agent in an expected user interaction model.

So it's not just tool calls.

There's also resources and prompts and sampling and a variety of other features, uh tasks
and different things too.

um So the primary thing is just that it is a standard.

The other thing to consider too is um if you just give like,

One thing in the tech world is this idea of an open API spec.

It's like a standard way for making APIs.

um And so because of that standard, you can generate basically documentation for an entire
API.

You can have 3,000 different URLs that you can call and get responses back.

So uh if you just handed that over to a larger language model,

uh it might call some of the right things, it might not.

This is actually part of why we have websites in the first place.

Behind every button that you click on a website, when you say send email, there might be
like six different APIs that get called.

And that is kind of a hard thing to communicate to a large language model.

But with MCP, you can turn all of those different things into tools, and those tools can
handle calling those APIs.

And so you can think of an MCP server as largely similar to a website, it's just

for a large language model.

that's fun.

That's a cool way of thinking about it.

And I really appreciate the point you made because the first time I heard MCPs, the first
time as a programmer, said, why not just APIs?

And you had already answered this, but I'm glad you answered it here.

The answer is not all APIs are the same.

All APIs expose functionality over the internet.

Well, internet APIs.

That's it.

They're not all using the same auth.

They're not all using the same way of talking to each other.

uh

If every API on the internet was using OAuth for auth, if every API on the internet was
using uh JSON API, like one of the very stricter standards, then maybe we wouldn't need

MCPs, but there isn't that same level of expectations and predictability.

So AIs have to do more work.

LLMs have to do more work to interact with all AIs or APIs around the world than they do
with MCPs, which is just a, it's a stricter standard, right?

It's like a stricter standard.

It was built with AI in mind, but the most important part, what I'm hearing you say is
that it is a stricter standard, which means there is

overhead and cost for an LLM to go from initial exposure to this MCP to actually being
able to use it practically.

Okay.

So if I'm a normie, if I'm a programmer, I just heard that and I think my brain started
running and saying, maybe I can build an MCP for this or whatever.

But if I'm a normie and I'm like, look, man, I've got a job where they're having us use AI
and they've talked about MCPs a little bit.

I use chat GPT every day and maybe I'm a little pro, so I'm using Gemini a little bit.

What does my exposure and interaction with MCPs look like?

What is like a next step action for a normal human being who's not a programmer, but who's
trying to be more advanced in their AI usage?

look like and what MCPs look like in their day-to-day life.

Yeah, so for non-technical people who aren't going to be building MCP servers, you're
probably not actually going to see the word MCP or the acronym MCP anywhere.

Because everywhere that implements MCP support um doesn't really use that word.

They call it connectors or integrations or that sort of thing.

But that is what you're doing under the hood when you go to ChatGPT or Claude and you add
a connector is the way that it works is MCP.

uh And so some of the

are going to connect to your canvas and or canva or your Figma and you can once you
connected that not like approved it you know authorized it to act on your behalf then you

can start talking to chat GPT and say hey could you pull up

a list of my FIGMAs, yeah, want that one, show that to me.

And it can show it to you visually.

And all of that is working through MCP and a substandard called MCP apps that make
interactive experiences.

And then you can just keep chatting, say like move this over, I'm actually not totally
sure on the capabilities of the FIGMA MCP specifically, but on Zillow, you can say, I

wanna find a home that is near a park with a good school.

Really?

city and yeah, and it will update all the filters for you and then show you that.

So our user interaction model is completely changing now.

We no longer have these complicated user interfaces and like a special website you go to,
everybody's just gonna use their AI agent and they'll talk to the agent and say, this is

what I want.

Like I'm telling you, this is Ironman and Jarvis where Ironman like.

just gets a bespoke interface and doesn't really uh use it unless it actually makes sense.

Like for example, in the Zillow case, I'm not gonna control all the filters and stuff.

I don't know.

let me find where the max budget amount is.

I don't wanna find that.

Just tell the LLM it will take care of it.

And then you're like, okay, here are three houses.

I'm gonna go click on that one.

Cause it's easier for me to click on it than to say, you know.

pick the third from the right, like, no, you're not going to that.

So there's a merge that's happening.

It's not just natural language.

It's not all just a check box.

But it's also not just the UI that we knew from five or 10 years ago.

Now it's a merge of these things where you use the LLM to control the UI until it makes
more sense for you to use the UI directly.

I love that because one of the things I've been telling people

I've been this this every podcast I ever do is therapy in some point for me and I'm just
like, you know, and one of the things that I'm figuring out through talking to people is

that I have some initial concerns and biases.

The majority of my concerns, biases about AI are justice and environment related.

So and we've already kind of acknowledged those in this podcast.

You know, we said like, those are things we want to see people work on.

But I think the secondary one is fearing um the maybe there's also a copyright.

But the secondary one is uh fearing it taking

away our humanity and our actual real interaction with things and just allowing AI to do
things that we want to do as creative people.

And one of the things I found myself often saying as well, as long as this is something
that I might have previously been using Google for, but it's doing a better job, I've

already allowed this part of my life to be delegated to computers.

But for some reason, I have a stronger bias against AI than I do against Google.

And that's that ups me experiment with it.

And so like to the degree that you just said, my first response was because I had Adam
Wathan on the podcast earlier and he was talking about how if you

use AI to get your documentation for Tailwind.

Nobody goes to the Tailwind website.

I can't show ads to them.

I can't make money from open source.

when you were saying about Zillow, I was like, is this a bad thing?

Right?

Is us interacting with all these things through LLMs actually detrimental to business
models that we have?

On the other hand, I'm like, well, if I am looking for those things, I might be just as
likely to use Google to look up that data and then click on the search result from Google.

But I can't tell Google, find me a house, whatever, whatever.

So basically that we've we've had a limitation of Google in my

ability to find houses and eventually I do have to land on the Zillow website and do some
searching there.

I'm still buying the site having found it on Zillow, right?

I'm still clicking on Zillow to get to the thing.

Zillow is still getting my data.

Zillow listing was still of value to those people because I found it through Zillow, but
I'm using a tool that might be easier and better for me and also potentially cheaper for

Zillow because they don't have to build this really robust search tool.

They just build the MCP and then rely on the robust search interaction happening at the AI
level.

I'm fine with that, right?

Like I'm not replacing a human thing.

with this AI, I'm replacing a search engine which we've sort of overloaded to do
everything and we're getting better search and I think we've all already said like AI is

better search at some point in our lives, right?

yeah.

You know, I have to push back a little bit on the hesitation to um replace a human thing
with an AI thing.

Lots of things that AI is replacing, as far as jobs are concerned.

I want to be sensitive to the fact that these are people's livelihoods and they are making
money to provide for their families and things.

But like...

Let's take call center workers, for example.

AI is replacing lots of support, um largely because, it's cheaper.

In some cases, it's actually better.

It's going to be if it's not already.

There are some services that are better.

And the experience for the user is often better because the AI is never tired, it's never
annoyed, it's not like it's always going to be the perfect interaction.

um But yeah, we're displacing the workers.

like, what about that?

The workers hate that job.

Like, for the most part, that is not a fun job to have.

And so...

Yes, AI is taking some things, but like we were talking about before, the parts that AI
has taken from my job, I'm thrilled.

I hated that part of the job.

Maybe I didn't realize I hated it, but no, it's not the fun part.

so AI is definitely gonna take jobs that people really do enjoy.

There are probably people who like trucking and they love being out on the road eight
hours at a time oh or for days on end.

And yeah, the AI is eventually gonna take that job.

That's gonna happen.

um So I do wanna be sensitive to that.

uh One thing that my parents bring up occasionally too is like, well, if AI and robots do
everything, then where does purpose come from?

And my answer to that is like, do you need a job to have purpose?

do you need a...

um I uh don't know, I feel like if you...

if you like what you do, like if you like making chairs, then the invention of a chair
making factory didn't prevent you from making chairs.

It just meant that, okay, well now you gotta find another way to make money.

But like if we're talking about, you know, pushes out forward into utopia period and
everything is good, like you can work if you want to.

You can work if it doesn't fulfill you.

And already we have people um who just decide whether they're gonna do things that
actually fulfill them versus people who don't.

do that, right?

And so that will continue to be a problem.

And there will be people who are like, I have no job, I have no purpose, and I'm not going
to put in the work to like find that purpose for myself.

That's always the way it's been.

It's the way it is now.

It's the way it will be in the future.

But um for the vast majority of people, I think they will find purpose.

And um let's say like we're not to that utopia yet.

um It is sad that the way you make money is different.

But one thing that I love about what Aaron Francis says sometimes is that you need to
operate in the world the way that it is now, not in the way that you wish it would be.

Yeah.

Yeah.

Um, I appreciate you pushing back.

um I it's, it's, it's fun because every time I talk to somebody in the podcast, I get to
learn about how a new industry is experiencing that.

And since some industries, there's no threat whatsoever.

And some industries it's, it's existential threat, right?

Like I had a writer Jordan Keller on earlier and he's like, yeah, the jobs are not the
same for writers as they were before.

And that's, that's a bummer, right?

Like, I don't think a lot of us are thinking, wow, writers just are being largely replaced
in these ways.

And a lot of developers, I mean, I think most developers I know have

had some power in the last six months where you feel a real existential crisis.

Like, will I even have a job potential?

Yeah, right?

Mine was definitely the beginning of January, 2026, when everybody came back from
Christmas break having really experienced the capability of Opus 4.6 and a lot of the more

advanced level things.

And I'm like, I'm a professional developer who runs a company of professional developers.

And I don't know what the future is of needing us.

And I feel a lot better today than I did then.

um But yeah.

um

It's.

I think the closest thing that I can kind of really come to that I think lands on kind of
what you were saying is the people have said every new set of technologies is going to

disrupt the jobs, right?

Like there's going to be some new jobs that didn't exist before, some old jobs are going
to go away.

And to me, as long as the end result is not more people in poverty, more people working
jobs that they hate, uh more people with billions of dollars having now trillions of

dollars, I'm okay with it, right?

Like job shifting is fine as long as the end result is not

more consolidated power, more misery, those things.

And so as long as we land there, then I'm with you, man.

I'm just like, yeah, we might have to shift.

I might have to change the way an agency works.

Developers might have to think about targeting a different part of development.

Like you said, developers who do the rote work are in the most risk right now.

Developers who are experts and deep thinkers are in the least risk right now.

So you might have to say, oh, I was just uh a cog in a system.

AI can be cogs a lot cheaper than humans can.

So I hear you and I appreciate that note.

Okay.

And um my hope is that um I would like us all to be trillionaires.

that would be cool.

I'm less into like, get rid of the billionaires and more into let's make us all
billionaires.

Like why can't we all uh benefit from this?

I can definitely see futures where uh the people who own the robots and have the AI

are the winners here.

It's very comforting to me to know that there are multiple companies working on solving
this.

at least the capitalistic competition will hopefully regulate that a little bit.

But if we play things back to the 1500s, my quality of life is way better than people in
the 1500s.

so even if...

it doesn't, know, rise and tide lifts all boats to the same level.

If it can at least raise all boats, then I think that would be good.

And there is going to be disruption in here and I'm hoping that the people most disrupted
are paying the most attention and that they're just not like, you know, wading through

life and just waiting for things to happen to them.

They are, that's like, that's the fact.

There are people in our industry, in highly disruptible industries, that they are not
paying attention at all.

And

I would like to help them, but there's only so much that I can do.

They need to be paying attention to this.

So I appreciate your podcast helping do that.

Yeah.

So speaking of that, one of the next questions I wanted to ask you was if there are the
breadth of types of people that we know listening, some of them are programmers and some

of them are not, is there something that you wish more people understood about AI that
would help prepare them for the future, that would help us all have a better experience

with it?

Is there something that people just don't get?

You know, I'm actually not sure.

I think there are people who are completely ignoring it, and then there are people who
think that it's like the end of everything and why are you even trying?

um I guess I would like uh people on both ends of that spectrum to come here to the middle
a little bit.

um I do not think, I have religious reasons to believe that AI is not the end of the
world.

Like Jesus is coming, like everything's gonna be okay.

um

And so like that perspective helps me uh a bit and happy to share with people on uh what
that perspective really means for me.

But uh at the same time, it is causing a huge disruption and I'm really concerned for
people who are ignoring that.

Have you found in your interactions with people you know, and I'm not asking you to name
them, in their interactions with people you know who are sort of like head in the sand,

just unwilling to listen, have you found there's a consistent reason for that?

Why people just don't want to interact with it?

Is it primarily criticism?

Is there any element of just not wanting to be made uncomfortable?

I think I remember talking to some people who are non-technical and are just kind of doing
it in the name of boycotting, right?

Like they just feel like it's wrong.

Not a lot of people.

Most of the non-technical people that I talk with about AI, um it kind of, we'll get into
a conversation, I'll say some of the things that I use it for, and it blows their mind.

They just have no idea.

Like there's no...

ah They just, I mean, they're living their lives.

It's not like everybody can be so tuned in like we are.

um And so I don't blame them.

But I think that that is a dangerous place to be in right now, to be completely oblivious
to what it is capable of.

um And so I encourage my barber and people that I know, you gotta spend a little bit of
time with this.

It's already at the...

the capability to automate some of the parts of your job that you hate.

And so you should just get familiar with it.

uh The other day, like a year ago, um I was at my dentist's office and he and um his
assistant were complaining about this insurance company that was making them fill out all

this paperwork to explain why this certain procedure needed to be done.

And they said it needs to be five pages and it's like this whole thing.

And he was so frustrated by it.

And he was like, it's so obvious why this needs to be this way.

And I said, you could just have ChatGPT write the whole thing.

You just outline the things, hit these points, and he was so blown away by that.

And this is years after ChatGPT shows up.

So just completely oblivious.

It's not part of their thought process of how can I use AI to solve these problems.

uh And that's what I would say the primary reason that people aren't getting into it.

is because they just don't have a concept of what it's capable of doing.

Yeah.

And I mean, to to to your credit, most people don't have a concept of what's capable of
doing, but you're you're like several levels past the average, even very capable,

competent.

Like, I wish we could have done an entire podcast just about Cody.

I'm just fascinated.

many things to say about Cody.

Okay, so let's let's go there a little bit before we wrap.

you said Cody, I've just seen your tweets about it.

I have not had any time to dig into it.

So you said it's an MCP server, but it also is an AI and it's also some software.

What's the actual architecture of Cody?

Yeah, so Cody probably the biggest unlock for Cody is that it's built on top of a platform
called Cloudflare, which runs 20 % of the internet.

It's a huge platform.

Started out as a content delivery network uh all over the world, and then they built a
couple primitives that just made it so much more.

So uh workers, so it allows you to write code that runs really close to users wherever
they have over 300 locations all over the world.

And then durable objects allows for those bits of code to have state or data that's saved
with that code.

And so between those two things, they've been able to build a lot of things.

And through that, they were able to build um a special mechanism for MCP support.

And so I've been using them to host my MCP servers for a while.

One of the problems with MCP is that

If you've got this large language model, there's only so much it has in its brain, it
doesn't have a brain, but only so much room it's got up here to be able to do something

useful.

If you're talking to somebody and you tell them all of the rules of Cricket and then you
immediately ask them, okay, what happens when this?

They're gonna lose it unless they have really ingrained their whole lives into Cricket So,
LLMs are the exact same.

If you fill them with too much information, they won't be able to answer your questions
very well.

reliably.

And one problem with model context protocol is if you add this MCP server and this one and
this one and all of these, um the naive way to implement this is to just load all of the

tools and descriptions for those into the context.

And now there's just this tiny little bit left for you to like send a message.

And so this was a big criticism of MCP um and it never concerned me.

I don't know why everybody like

took off with this is like, MCP is a failure.

It's impossible to make it good because of this context bloat problem.

And so there were a lot of solutions to

this. um And one of them, a really simple one, is don't load them in the

context. Put them over

here. And before you answer the user's question, you do a search, and you grab just the
ones that are

relevant. Very easy,

obvious. And that's what I've been putting on Epic

AI.Pro. Listen,

people. It's not that big of a

problem.

Yeah.

But another solution that is really interesting is Cloudflare decided, you know what?

These large language models have been really trained to write code.

Because software development has been such hugely impacted by large language models that
they are specifically training it to write code really well.

And it's done super great.

um And so Cloudflare was like, what if instead of taking our 3,000 endpoints and saying,
here's all the things you can do,

And actually what they did was they chopped it up into 13 MCP servers.

So you just enable the ones that you want.

They said, that's stupid.

We don't like that.

So what if we instead just um tell the LLM with three tools that it has the capability or
no, just tool tools, has a capability of doing stuff with Cloudflare.

If you want to know, use the search tool.

And so now it does a search and it comes back.

Here are the TypeScript definitions for accomplishing.

these tasks, the tasks that you want to do on Cloudflare.

And then the second tool is execute.

So now the LLM, has the TypeScript definitions for uh what it needs to do.

It's gonna write code and give it to the execute tool.

And then Cloudflare will take that, make it an isolated environment that's safe.

like it's not gonna do anything bad in there.

And then it's going to uh do, who knows, all kinds of things.

Because not only is context bloat a bad problem when you're loading those MCP tools, but
also when you run them, if it gives back more information than you needed, then that's a

lot of stuff too.

So if you put that in a code block, then you can do all this stuff inside the code block
and just return the pieces that you need.

And so that's called code mode.

And Cloudflare put this together in a really nice packageable way.

And that is what makes Cody so capable.

Cody actually exposes three tools.

The third is Open Generated UI.

And so that tool allows uh the agent, whatever agent I'm connected to, to write HTML and
then hand it over to uh that tool, and it will display that HTML.

Now, lots of uh agents actually have this built in as a built in capability, so it's a
little less impressive now.

um But the concept of being able to generate an app whenever you want is pretty powerful.

So as I was building this, was like, you know what?

Every single thing I wanted to integrate with, I would add code to Cody to integrate with
that.

OK, now you can control my Spotify, and you can control my developer tools, GitHub and
Cursor and all these things.

And it was really cool.

I could say, hey, spin up a cursor.

I would actually have a conversation in Claude with Cody.

Claude would have a little bit of hard time.

And so I'd say, could you spin up a Cursor Claude agent to fix that?

Make it not hard next time.

And so it's like self-improving as well, which is very, very cool.

So um after a while, I realized like adding, changing Cody every time I want to integrate
with something is kind of dumb.

I don't want to have to do that.

And also there's something in the back of my mind that's like, how can I make this useful
for more than just me?

And so uh I decided, what if I um come up with this, like an OAuth flow or like a
mechanism for handling secrets?

So I don't have to have all of that handled in the deterministic code.

but the agent can actually just write the fetch call.

uh And so I came up with a system for secrets where I don't want the agent to know what my
private key is for accessing my bank.

I wouldn't want it to do that.

And so uh instead, I have a special syntax that the agent can write.

So it can write a fetch request that makes a call out to a third party API or something
like that.

And it requires this token.

And I intercept that fetch request.

So it's running inside this isolated environment.

It pops out and says, hey, I need to go and make a request for data or uh change some data
or something.

And so I intercept that.

And I look for any authentication stuff that's going on.

And it says, OK, here's that special syntax.

You need the secret to be replaced in this fetch request.

So before I send it out, let me replace that and then send it out.

But I also say, where this request is going, is it supposed to get that secret?

Yeah, okay.

Yeah.

I have the ability to integrate with anything that has an API, anything at all.

um And what's cool about that too is that in my closet over here, I've got this server.

It's a storage device, a NAS.

I forget what it Network Attached Device, or uh storage, yeah.

And so on that, I have a uh special service that's running that connects to the Cody MCP
server.

and says, here are the things that I can do.

And Cody's like, great, I'll add that to the things that can be done.

So my local network is now accessible to Cody through these capabilities as well.

And so I can say, OK, you control my lights, you control my sprinklers, you control my
blinds, you control whatever.

And so now I can connect to literally anything.

I don't have to change Cody.

And um now I'm just focused on what are the primitives that Cody needs for these agents to
be able to write code.

that can go do those things.

So I've got jobs and so like I scheduled a certain blind to go down every night at 9.30 um
and actually jobs are what control all the blinds.

um And I've got skills, so they're like reusable bits of functions.

um So I have one that's kind of a combination of a job and a skill that will manage my
journal for me.

And every day at a random time of the day, it looks through my journal.

and it uh identifies, this actually, this part takes inference, because I call out to a
model, and I look at the journal and say, what's an uncovered part of Kent's life that we

can ask him about?

And it'll send me a prompt and say, you told me about this time that you planted 15 trees
in this park for scouting, uh tell me more about other experiences you had with scouting,

or whatever.

So over time, I can build out this personal history.

I have a personal autobiographer who is asking me interview questions.

over time.

they're just, the possibilities really are endless.

um And it's in a similar vein, some people might have heard of OpenClaw, it's kind of a
similar thing there.

Except what makes it really awesome is that, and no, it's not running on my uh Macbook or
Mac Pro or whatever, my Mac Studio, uh it's running in Cloudflare.

And the majority of it, unless a particular skill needs to do some inference with an LLM,

m It's not using a large language model.

I am jumping into Claude and ChatGPT and Cursor where I already have a subscription and I
tell it to use Cody and then from the point that it gets to Cody to everything else,

that's all whatever it costs to host, which Cloud Flare has a really generous free tier
and so it's free.

And I'm giving a talk next week at um AI Engineer in Miami and the title of my talk is
Build a Free Agent.

uh

Fantastic.

yeah, it's perfect.

It's free as in puppies, like you gotta manage it yourself, but like it's, you got it for
free.

Free as in cookies, so like they're free.

You just consume them, they're yours.

And free as in freedom.

Like the agent can really just do anything.

Do you know if that video is going to be put online?

Because if so, we'll link it in the show notes.

I believe it will.

It will be recorded,

All right.

So what we're going to, so this will come out maybe, well, maybe after you give the talk,
but I'm sure before the talk is online, but we will make sure that when it comes off

online, we will come back.

We'll put it in the show notes.

So it might not be there when you all are listening, but just make sure you check back in.

Cause that's, I literally was just like, man, I wish we weren't at the end of this.

Cause I, I have so many more questions for the non-technical listeners.

Thank you for sitting through over the last 15 minutes.

Cause that was exactly what I wanted.

So even if it didn't make sense to y'all, that was, that was so good.

This is amazing.

And honestly, I've, I've seen so many people do

open claws and all kind of stuff and never been super compelled but man you just you made
it really sexy you made it really interesting ah my dad and brothers have been using I

think it's home assistant which is you know runs on a raspberry Pi and it's but it's a
it's much more human controlled and much less you know human language or whatever so I'm

very curious to kind of hear the overlaps there um

the cool thing there is I don't use Home Assistant, I've been wanting to, but yeah, just
being able to control things.

Like I have an app for controlling the color of these lights and stuff, but it's just
easier to pull up Claude and say, change my lights red.

Yeah, yeah, that's wild.

um OK, so because we're getting towards the end, I want to ask for you, is there anything
when you kind of saw the agenda for the podcast, when you heard about it or even as we

were kind of like going through conversations where you're like, I wish we had gotten that
today.

I want people to hear someone talking about this.

You know, we covered a lot of ground.

I'm not feeling like we're missing anything here.

I guess the one thing that I just wanna call people's attention back to, especially the
technical people there, but even non-technical people, this is applicable to you.

When implementing an idea is so easy and almost free,

uh What is left for humans to do?

I think that there is still something.

think humans are still valuable in this future.

It's not like the AI machines are gonna just completely take over humanity and they are
gonna be the ones to go out into the stars and explore.

I just don't think that is uh in their nature and we can't really code that into their
nature.

em What humans bring into this table is that empathy for other people.

and solving problems and solving real problems and exploring and creating.

And so you need to develop that capability of connecting with other humans and being
really genuine to try and find solutions to their problems um and learning how to bring

that clarity to the problem enough that uh an AI agent can take over the implementation
from there.

um And so...

Yeah, that is why I am building EpicProduct.Engineer.

I welcome anybody to go sign up for updates on that to give you intentional practice on
developing this skill.

That's good.

I mean, that's lines up with what you were talking about before, which is just sort of
like, what are they we're losing the need for some people, it's the opportunity, but the

need to do the individual hands on thing.

But the brains aren't going away.

The empathy, the understanding, the humanity, the soul is not going away.

Yeah, I love that.

um OK, so last thing we're to do before we wrap for the day is well, actually, first of
all, if somebody says this Kent C.

Dodds man is brilliant, I want to follow him and pay him money and learn from him.

Where's the best place for people to go?

Yeah, kentcdodds.com is kind of the place that'll link out to everything else.

Like I said, it's a pretty big thing.

So if you want to ask me a specific question, there is the Call Kent podcast, where you
can record your question or even type it and choose an AI voice to read it for you.

And then I will listen to it.

I will uh record my voice to respond to it.

And then our conversation turns into a podcast episode.

And so if you have a specific question for me, jump on that.

um I'm also very active on X, the everything app.

So Kent C.

Dodds over there.

And then I've got a podcast going on, the Call Kent Podcast, or the Chats with Kent
Podcast as well.

um So yeah, there's also like all my courses and stuff, but kentcdodds.com is like the
central hub for everything.

yeah, thank you.

So ah I forget half the time, but I remembered this time to share a community member who
shared with us a trick about how they are using AI in their day-to-day life.

My friend Bogdan Karchenko said, I used Claude code to help me start and manage an
aquarium for my kids.

And it also built a website and he linked to it.

We'll put that link in the show notes.

And when I saw that, I assumed what he meant was like an online aquarium or something like
that.

He said, no, like he went and he did all the research.

He learned a particular method for creating aquariums, which is called the Wallstead
method.

ever had fish but apparently there's a lot that goes into making sure the nitrogen
balances are right and stuff like that I don't know any about this but apparently chat GPT

set it up for him and they've got a little like old-school webcam on there so like you can
go look at and see pictures over time so ah that's it's it's very fun I'm surprised by how

many people are using using it to do fun creative science things with their kids is that
something you've done at all Kent are you using it at home with your family at all

Yeah, yeah, like the other day I had Cody build a little color and shape matching app for
my kid.

So like my two and a half year old.

So that was kind of fun.

There's also this woman named uh Jessie Garrent, think.

I can't remember her last name, but she has been using OpenClaw a lot and she is a
non-technical co-founder, like sort of technical, but not very.

And she just decided, okay, I'm gonna be a mom now.

Had four kids in five years and is now

homeschooling them all and uses OpenClaw to do all the administrative work.

She has like 11 agents, uh OpenClaw agents that are running, doing different things for
her.

They order her groceries, they plan her meals, they um plan and keep track of the lessons
and the curriculum for homeschool and then just enabling her to be a much more present

parent.

So she doesn't have to do that, she can just be.

the one doing the actual instruction and with the kids.

uh that is one person that I have just, and she shares the whole process and the whole
experience in like two minute videos on X occasionally.

She's just a joy to follow.

So I definitely recommend that.

We will hunt her down and we'll put it down.

I think I saw a Primeagen comment on one of her things being like, so you're not doing
your work anymore.

And then later being like, I was wrong.

She's actually letting it do the work.

So now she can be present with her kids.

And so really enjoyed that interaction, but I never clicked through to it.

So obviously I need to, we'll also put that in the show notes.

So.

Yeah, she actually would, I don't know if you can get her, but if you can, she would be a
really great one on the pod.

Alright, I'm gonna put her on a, see if we can try to hunt her down on it.

I appreciate that.

Well, Kent, thank you so, so much for spending your time with us today.

Thank you for teaching and sharing and for being a great friend.

I really appreciate you being here.

Well, thank you, Matt.

Thanks for all the good that you do.

It takes good people like you to keep the world going in a good direction.

So I appreciate you.

God thank you and for the rest of you we will see you all next time.

Bye, everybody.