Jack Herrington explains why frameworks still matter in the AI era: a test showed Claude rebuilding an app without frameworks produced 2.5x complexity and 2x code size. He introduces TanStack Start an
Key Takeaways
- Frameworks reduce accidental complexity and save developer time and LLM tokens, which translate into real cost savings.
- AI-generated code built from scratch without frameworks tends to be larger, more complex, and harder to maintain.
- TanStack Start offers Next.js-like SSR capabilities with a client-first approach and flexible, opt-in server features.
- Choosing a framework with widely known patterns and training-data familiarity benefits both humans and LLMs working on the codebase.
What the video covers
- Jack Herrington, a full-time TanStack contributor, opens by noting how AI is prompting developers to rethink coding, testing, design, and workflows.
- He questions whether frameworks are still needed, distinguishing the app's bespoke functionality from the underlying infrastructure (routing, data loading, etc.) that frameworks provide.
- Herrington argues that SSR applications need code that runs isomorphically on server, web, and potentially desktop (Electron) or React Native, which the raw web platform alone doesn't fully handle.
- As a 'taste test,' he asked Claude to build a two-page app (routing, tables, AI features, an actor/Marvel Cinematic Universe lookup) with a framework, then to rebuild it with no frameworks, including a custom React.
- The framework-free version was roughly 2.5x more complex (by cyclomatic-style complexity measures) and about 2x larger in code size, implying higher maintenance costs, undocumented bespoke code, and more tokens consumed.
- Citing Dan Abramov, he distinguishes accidental complexity (unnecessary tooling) from purposeful complexity (the app's real requirements), and argues frameworks reduce the former.
- He presents TanStack Start as a Next.js-like framework that starts client-first and adds server capabilities as needed, mapping well to the Next.js Pages Router model.
- TanStack Start supports three data-loading modes: client-side, server-side (like getServerSideProps), and a data-only server approach that improves security for dashboards.
- Features mentioned include preloading on link hover, streaming, custom server functions, and deployment flexibility via Nitro (e.g., Docker).
- TanStack Query provides cross-component data caching; TanStack Form handles validation; and TanStack AI connects applications to LLMs, including tool calls and approval workflows.
Chapters
- 00:00Introduction and Speaker Background
- 01:43Rethinking Frameworks in the AI Era
- 02:53Why the Web Platform Isn't Enough
- 04:03Taste Test: Claude Builds an App With and Without Frameworks
- 04:54Complexity and Maintenance Cost Results
- 06:20Accidental vs. Purposeful Complexity
- 06:26Introducing the TanStack Ecosystem
- 07:20TanStack Start Features and Data Loading
- 10:14TanStack Query, Form, and AI Libraries
- 20:00Code Mode and TanStack Intent
Full Transcript — Download SRT & Markdown
Speaker A
I'm Jack Herrington and I run a YouTube channel, but I also do a podcast with my friends TJ and Paige and we talk about all things friend, but we're actually thinking about changing that over to a more of an AI focused
Speaker A
podcast cuz that's the world we live in nowadays. But I also am a full-time contributor on TanStack and that's really what I'm going to talk about today.
Speaker A
Um Interestingly, my co-colleague Allen is over there in the harness talk. So that if you want to get really in the weeds, I don't mean to discourage you from hanging out in this talk, but if that would be really fascinating if you
Speaker A
want to talk about actually building your own agent harness. All right. Uh okay, so I did a a kind of a talk around TanStack at CascadiaJS where I talked about all of our amazing libraries and frameworks.
Speaker A
And then like everything happens. The whole agent thing happens and we're starting to like just rethink like everything that we're doing. We're rethinking coding and we're rethinking how we test stuff and design, cloud design and our workflows and
Speaker A
we're going to chat GPT for about our relationships and try to rethink our society and does it What does it mean for frameworks, right? At the end of the day do you need frameworks anymore? And really what are frameworks? I mean,
Speaker A
there's your app, right? The thing that people are actually showing up to use that's got all the bespoke functionality and then there's kind of like all the infra stuff and that's really where the platform sits.
Speaker A
And when you think about it, like people have been saying for a while like the Remix V3 guys are like the web platform is pretty much all you need, but or is it really not so much? I mean, there is
Speaker A
like routing, and there are all kinds of, you know, extra stuff that you need to make the web platform actually work. So, in the case of TanStack, our options in that case are start, which is our variant of Next.js,
Speaker A
and another example in this space. And obviously, I've compressed this to just show three things.
Speaker A
There's so many more. Form, because we all use forms, and AI, which is the library that I work on, which gives your application access to LLMs and all that.
Speaker A
And why is the web platform well in and not enough? Well, because for a lot of us, we're building like SSR applications, which means that that exact same code needs to isothermically work in both of these environments. Needs to work on
Speaker A
web, needs to work on the server. And if it even goes beyond that, it could work on desktop with Electron and React Native, if you're going to use that.
Speaker A
Let's get my dry mouth out of the way. Now, if you ask for a you ask Claude, "Make me an app." You're going to get a V plus React app.
Speaker A
So, should you actually ask for a framework? Should you ask for TanStack? I think you should, but let's do a taste test. Let's actually go and take this app, which is actually a two-page app.
Speaker A
So, it's going to do routing and all that. Uh it's going to have tables. It's going to have some AI in it. You can actually like ask this thing to like give you this It was a kind of cool app,
Speaker A
actually. It was trying to figure out like given a particular actor, are they going to show up in the Marvel Cinematic Universe if they're in like everywhere all at once. It was actually a pretty cool app.
Speaker A
So, I basically took this and I said, "Okay, great. Okay, cool." So, Claude, build us just this app, and then rebuild it, but don't use any frameworks all.
Speaker A
Just use the Yeah, just start with V. So, in the case of building this application, we have V and React, but then we all we have all the cool TanStack stuff. So, that'd be start and query and form.
Speaker A
Dominik Dorfmeister is here if you want to talk to him about TanStack Query. A lot of great TanStack folks here.
Speaker A
And then on uh the the build everything side, it literally builds its own custom version of React.
Speaker A
Builds everything else. The only thing that it actually reuses is of course just like V hanging out the bottom there.
Speaker A
So, what happened? Well, complexity was 2.5 times the use complexity scale, one of the old CS complexity scales.
Speaker A
Which means that the app is so much bigger for the LLM to actually walk around and try and figure out.
Speaker A
That means that in terms of maintenance, you're going to be paying 2.5 times more because now you're maintaining all this bespoke code. And if you actually want to jump into the code and look at code anymore like the old school way,
Speaker A
like who knows what's that's going to be. There's no documentation on it. It's just completely unique to this app. And then of course the code size itself is way bigger, 2x bigger. So, frameworks, they save everything. They save your
Speaker A
time, they save your tokens. And of course tokens burn into money. And at the end of the day, because of all of the bug fixes and everything else, they save your what? So, which one should you pick? If you're
Speaker A
going to pick a framework, which one? Because Dan Abramov, I don't know if you know him from JavaScript Jabber, said that choosing the right framework reduces accidental complexity. There's two types of complexity. There's accidental complexity, which is you know, I'm just building this app and
Speaker A
out of nowhere I decide to use this weird thing. I don't really need to use that. And then there is kind of purposeful complexity, the complexity of the app that you actually want to build.
Speaker A
It's a complex topic. You're doing really cool, interesting stuff. And you want to reduce the accidental complexity. You don't want to bring in more stuff just because, you know, you wanted the app to be complex for no particular reason.
Speaker A
So, TanStack with our awesome new logo. Everybody familiar with TanStack or is that Yeah?
Speaker A
Mhm. Okay, great. Actually, less is more because then you guys get to hear about all the cool stuff that you're going to get with TanStack.
Speaker A
In particular, TanStack start. So, who uses Next.js? There you go. Okay. Who uses Next.js pages router?
Speaker A
All right. App router? All right, cool. Well, kind of mixing. I can't really see cuz I'm getting blinded up here, but it looks like there's some folks doing both of those things.
Speaker A
TanStack start, if you do uh so, it handled everything in this app in terms of the routing and everything else. And we can kind of go into the kind of feature by feature breakdown of it.
Speaker A
Uh it will give you everything that Next.js does and more. So, in in particular, it's a really good route map uh map to the Next.js pages router. So, if you're familiar with that model, it's going to work really well for you.
Speaker A
But, it takes a different tack on it. It actually starts with the client first and then gives you the ability to just kind of add on more server support as you want.
Speaker A
So, it does It's It's built on Beat. It's got all the SPA and MPA routing that you'd ever really want. I mean, it's got actually three different types of ways to load data. You can load it client-side, or you can load it on uh
Speaker A
the server-side like you used to do with with like get server-side props. Or you can actually do this really cool data-only thing. Where you've got like a dashboard and you know all the rendering's going to be up on the front
Speaker A
end anyway, and you don't really want to go and cache it. You can actually have it just get the data on the server and then send that up to the client. And that actually solves a lot of security
Speaker A
problems. It's a really nice model. Uh again, so you're going to move that heavy loading to the server.
Speaker A
And it's going to do all of the really cool like you you glide over a link and it's going to do all that preload.
Speaker A
It's got really great streaming support. Streaming wherever you want it. Use custom server functions for that.
Speaker A
And then if you really want to do the RSC thing, it takes a different tact than the Next.js App Router does because in our world, what we in the TanStack Start world, we actually just kind of give you RSCs as a generic tool that you
Speaker A
get to decide when and where you want to use. You're not the the platform hasn't decided for you that every single page is going to be RSC'd and every single and all of that data is going to come
Speaker A
through on every single page. And the really cool thing is that it just hit 1.0. So, we're really excited about that. Just hit the 1.0 like last week. So, we're cooking on that and we're really excited about it.
Speaker A
So, if you want to use that, you can go and just ask your agent, "Hey, use TanStack Start." You can also do PNPM create start app if you want to get started with it. That is really great.
Speaker A
You can actually go and it's so cool. You can go and pick from all kinds of templates where you can get some of our partners.
Speaker A
You can get Stripe and you can get all that. You can go and just say say where you want to deploy it so you can get Nitro if you just want to deploy it like in a Docker for instance or something
Speaker A
like that or you can say, "I want this on Vercel." and it'll automatically create a Vercel adapter for you.
Speaker A
It's fantastic. And also, we have all of the skills and I'll talk about that in a second for your AI. So, that's TanStack Start. TanStack Table is fantastic.
Speaker A
This is actually the OG TanStack library. So, most people think that it was React Query. Uh Tanner Linsley back in the old days created React Query to go make queries really easy out of React and that became TanStack Query, but it was really the
Speaker A
table that was before that and it is fantastic. In fact, it gives you really, really high-performance virtualization of tables in with millions of records, not a problem. You can get sorting and filtering and aggregation and even better, it's a really lightweight.
Speaker A
Meaning that if you don't want sorting, you don't want filtering, you don't want aggregation, it doesn't go in the bundle. And in fact, this is kind of a theme across all of the TanStack libraries where one, we want to make
Speaker A
them platform agnostic. So, if you want to use it, most of the libraries are like you can use it on React, you can use it on vanilla JavaScript, you can use it on uh Solid or even What was the one that just
Speaker A
came out recently? Uh I don't know, whatever. Um You can just keep on using it like anywhere you want it, view, whatever.
Speaker A
The The way that we do that is that we create the agnostic layer first and then we build the adapters on top of it and that means that those adapter layers are really lightweight and all that they do
Speaker A
is just connect to the underlying framework. It also means that your code is really, really portable.
Speaker A
Again, yeah, so you just use Just tell it to use TanStack Table and virtual for those tables. If you want, you can just install it directly.
Speaker A
There's also TanStack Form used in this example as well. So, it's not a huge example. It's just the one form field, but this is something we do all the time. You know, we do things like uh really small forms or up to complex
Speaker A
forms that have every kind of validation that you'd want. Is it a string? Is it a date? Is it an email? Blah, blah, blah, blah, blah, blah. All that.
Speaker A
One of my favorite features is it does async validation. So, imagine if you've got like a zip code and you want to know like is this thing going to, you know, send off to uh Can I deliver to this particular address? You
Speaker A
can actually have it go and make a server call, go and get the data and then come back and it's all in the same exact uh uh validation process.
Speaker A
It's got a single source of truth for all its data structures. So, if you want to go and take that exact same data structure and validate it on the client and on the server, which you absolutely should every single time, not a problem.
Speaker A
It handles that for you. And this is a really cool one. So, like who works at a larger company, like a thousand-person company? Big Yeah? Okay.
Speaker A
All right. So, if you work at a big company like that, you're probably going to have like a design system. Because every you've got multiple different teams building multiple different apps and you want to have like fields and buttons that all look the
Speaker A
same and all look, you know, consistent as your your customers are navigating through the interface.
Speaker A
And the really And the the problem with that is when you start using other forms managers like React Form or whatever, uh the problem is that it like you'll get these massive declarations of buttons inside of the structure of the
Speaker A
validator. And it's just a real hassle to try and figure out. So, what we did with TanStack Form is we create these compositional components, meaning that the context actually holds the form and then you can have like a an editable form uh a form edit
Speaker A
component that you just drop into any form and it hooks in and you get pure type safety, but it also means that you don't it's a it's a much tighter code code path.
Speaker A
And again, of course, you know, it's really easy to just use forms. TanStack Query I Who's used React Query?
Speaker A
Like every person who's used React has probably used React Query. React Query is in 30% of all deployed React applications, which is kind of absurd. Like the only larger React library's going to be React itself.
Speaker A
And so what do they give you? Well, it gives you all this amazing stuff. And this is why people use it. Gives you caching data across components. So, if you want to go and have one one query that goes off and fetches data from one
Speaker A
spot and use it anywhere else as long as you use the same query key, you get the same data. And so you're just moving data around and using it kind of like an implicit state manager. Super cool.
Speaker A
It integrates super nicely with TanStack Start. And that's one of the great things about TanStack is because we have so many properties now. We've got a Next.js equivalent and we've got form and we've got AI and we've got all the
Speaker A
rest of it. We're constantly looking for ways that if you were to use more of it in combination with each with each other, it just gets better.
Speaker A
And so we really do well with server-side rendering not just on our TanStack Start, but also on Next.js.
Speaker A
Of course, if you want to do like polling and retries and streaming, this has all that stuff built in, no problem.
Speaker A
And of course, like, you know, because it's been around for so long, like literally every model was actually trained on this thing, so it's not really a problem in that space.
Speaker A
And then of course, because we wrote it agnostically, you can use it on view. It's not just TanStack React Form, it's TanStack View Form and all the rest. All right, sorry.
Speaker A
TanStack React Query, it's also TanStack View Query and all the rest of it. And so of course, fully fully typed. And of course, all you need to do is just say use TanStack Query and that's it.
Speaker A
That's really all you're going to need to go and make sure that everything coming out of your endpoints is fully type safe.
Speaker A
And then of course, there's my favorite one. This is the one I work on all the time, TanStack AI. So, TanStack AI is an analog to Vercel's AI SDK.
Speaker A
And we kind of thought that we we didn't want necessarily one company to own not just a an AI gateway, but also the SDK itself because then you're uh by default kind of integrated into their whole ecosystem. And we think that
Speaker A
having an external set of libraries that connect you with all your models is a real there's a benefit to us as customers.
Speaker A
So, what's the deal? Well, I can tell you from a personal experience that LLM tool calls, which is how an LLM actually becomes an agent, you give an LLM an an a tool a set of tools like, "Hey, add this product to
Speaker A
this cart." And then your LLM becomes something that can actually act on your behalf, has agency, and is thus an agent. And so, tool calls are amazing, but they're also incredibly hairy. They're even they and they're consistently getting more complex as
Speaker A
Anthropic and OpenAI add more functionality to the LLMs. And that includes things like all of the approval flows and all the rest.
Speaker A
Even asynchronous stuff. So, like for example, we just as Allen was talking about in there, we're just going and bringing out our first harness system.
Speaker A
So, you can actually use Ten Sec AI to build a full harness, your own custom harness. And we do all of the flows for tool approvals, including even asynchronous ones. So, if you think about it like you got some
Speaker A
agent out there that's figuring out if you need to go and buy more books for your bookstore, it'll wait on your approval until you wake up the next morning and you're like, "Yeah, okay, that's the one I want." And it'll do
Speaker A
that for you. Of course, it has adapters like every single vendor that you'd ever want, OpenAI, Anthropic, and all the rest.
Speaker A
And it it gives you access to really anything you want. If you wanted to use chat, for sure, we we do that and we do that in spades. We do that with streaming and tool calls and all the
Speaker A
rest. But you can also get images and videos and all the rest. And we're augmenting that chat now so that it handles what's called MCP apps, which is the built allows MCP servers actually inject novel UI into a
Speaker A
chat flow. It also gives you the generative UI spec V1 V 0.2, which allows you to actually allows the LM actually write code into HTML and then embed it in the chat. So it can like you can say, "Hey, I need a
Speaker A
you know, a graph of how we're doing, you know, with sales." And it just right there, right in your chat, and it actually even continuously connects to the real-time data. It is bonkers cool. And there's one thing that
Speaker A
I kind of left out here, which is one of my favorite features, which is called code mode. What code mode allows you to do is actually have the LM write TypeScript in response to your query. And that TypeScript actually
Speaker A
runs in an isolate. And the really cool advantage of this is that it's hyper efficient and it's hyper accurate. So got to tell you, LMs are absolutely atrocious at math.
Speaker A
Code mode makes them fantastic at math and gives you real results with real code. And then you can kind of cash that off.
Speaker A
It's it's great stuff. And of course, that one just hit 1.0 last night. So we are super excited about that. That has been I would say a year in the making. And yeah, so Alan has done some fantastic
Speaker A
work on that and just really appreciate all the hard effort he's put in there.
Speaker A
So, yeah, if you want a chatbot, telling you, forget the AI SDK, TanStack AI is the way you want to go. I think that's up near 500,000 downloads a week now, so that's really taken off and we really
Speaker A
appreciate that. So, the last one I'll talk about is TanStack Intent. Now, this is really cool. So this is a uh skill-based a skill extractor, right?
Speaker A
You install all these libraries, how do I use them? Cool. TanStack Intent is how you do that. We have skills with every single one of our libraries and TanStack intent puts them into your environment so your so that your AI's know exactly how to
Speaker A
use our TanStack code. And of course, I think, you know, our biggest motto is that we built for humans first, but the agents love it, too. So, we we might not go in depth on every single line of
Speaker A
code, but we really make sure that those APIs are really, really friendly. So, that when you look at the code later, you can understand all of the code.
Speaker A
All right, that's it for me. Here's all my socials and a nice little QR code there you can get and yeah, that'll do it.
Speaker A
[applause]
Topics:TanStackTanStack StartReactframeworksAI codingClaudeSSRNext.jsTanStack QueryLLM











