Agents on the Canvas in tldraw — Steve Ruiz, tldraw

Hello. Hey. I'm going to kick off—sorry, we're starting a little bit late here. I am Steve Ruiz, from tldraw. Does anyone know tldraw?

Yes.

Hey. Allright. Fans in the room. If you don't know tldraw, tldraw is kind of a couple of things. tldraw is an online whiteboard. You can go to it. It's free. It's really nice.

I'm going to be using it for my slides. tldraw is also a startup. We're based here in London. And tldraw is also an SDK, something that you can use to build other products. So if you've used Replit's new

Agent Canvas, then that's built with our canvas. If you've used

Luma AI's new canvas, that's built with our canvas. If you've used Stitch's new canvas, that's not built with our canvas. But if you go into this kind of annotate mode,

this actually is our canvas. So we're in there somewhere. It's an Angular app. So anyway, tldraw is, again, a company that makes a whiteboard, makes a whiteboard SDK. Part of the idea with the SDK is that you can build cool things with the SDK, which means it's part of my job to build cool things with the SDK to improve it.

And a lot of those things recently have involved AI,right? Hackable canvas, runtime, built with React. In fact, the canvas is also just React components, so component, component, component, which means you can do some pretty cool stuff.

First one, I'm going to be at the mercy of the demo gods here and the internet and other things. So bear with me here. Does anyone remember this app, MakeReal? Maybe this tweet about MakeReal. Did anyone remember seeing this tweet back in 2023 when the vision models came out?

Allright, cool. This was one of the first projects to kind of break containment

in AI.

We're going to go to my phone. Fuck this.

Which may itself be a disaster. So the basic idea with MakeReal was that you could use the canvas, draw, then send that to a model and have it make it into a working prototype, which sounds very, very quaint in 2026.

But in 2023, it was all the rage because there was no lovable, there was no kind of vibe coding hadn't been termed or coined as a term. And so this was one of the first projects where non-technical people could make technical stuff without having to code or to look at code.

And so again, we are at the mercy of the several fucking various

internet gods. So we'll see. I'm going to let this cook for a minute. But at the risk of leaking my API keys, I will try and switch to a faster model. Oh, no. It's coming. Allright. Maybe I just gave it something hard to work on.

We'll see. Hang on a second.

Basic idea, something like that. Yeah, there's my API keys. I always got to rotate them every time I do this demo, sadly. We'll see how much gets spent before I get done with the talk. But it's a very simple prompt.

It's just like, go make this interactive. This is not what I asked for, but we'll see. Allright. Good job. Cool. And this is a working thing, which is great. It's like a real little bit of HTML. But you could also annotate on top of it.

You could say, like, hey, actually, make this green. And why don't we use these colors? I'm going to make this red

and black. Use these colors. Right? And so you're kind of constructing a prompt here, even the prompt that includes the old website. And

there's not many apps that have actually kind of used this. Only now, only like the Google Stitch that I mentioned before, this idea of, well, it didn't do anything that I asked it for. Terrible demo. We're going to move on.

Things have changed in the last couple of years. Anyway, there was another one in tldraw computer, which I'm actually not going to go into, but this was a couple of chains of prompts. But eventually, we were like, hey, AI on the canvas actually might be pretty cool.

Maybe we should have the AI kind of work as a collaborator, like work with you on the canvas. So the first one of those was a pretty straightforward idea where you could say, like, draw a cat. You could do anything.

You could say, draw me a diagram, finish my slides, complete this graph that I'm working on, things like that. And unlike maybe image models, like diffusion models and things like that, it's not building an image. It is using text structured outputs to make the same things that I could make,right?

So I have tools. I have circles and shapes and like this. And it's funny to see how these things have changed over to, oh, it's sad.

But

the fun of this is, as a way of exploring the model and what the model knows and how it can comport all that stuff, make the cat blow out the candle.

It's pretty cool. You could also do something like draw a mouse. So multiple prompts at the same time on different parts. And so even though I didn't tell it what a candle was, and I certainly, like the application doesn't know what a candle is, and I'm not even sure that cats can blow.

But it's correctly interpreted that and kind of incorporated it into the design. So

with incredible detail as well. It's still sad. So this was really fun because this is like solving a lot of problems that might not be obvious. Like vision models, when it comes to structured data, number one, there's much less vision training data than there is for text.

Number two, a lot of that training data conflicts in ways that text does not and other types of things don't. So for example, the y-axis on a Cartesian graph, as you go up, that number goes up,right? So 0, 1, 2, 3, 4, 5.

On the web, the y-axis goes up in this direction,right? The top left corner is 0, your top left corner here. But as you go down, the y goes up. There's left,right? Like there's your left. There's stage left. There's all sorts of things that conflict within language and within images.

So

training the model to kind of behave predictably and produce things like this was really, well, I used training, prompt engineering the model to do it was really tricky. But this was fun. But we felt like it didn't go really far enough because it was just one-shotting,right?

We wanted to do an agent. So this is what Cursor looked like back in 2025 or something like that when I did this. Draw a diagram of the life cycle of a butterfly.

So this put it into a kind of like agentic loop, like you might have seen elsewhere. And I'm sure you interact with dozens of times a day by now for this crowd, where you have it produce an output and then review the output and kind of iterate until it thinks it's done.

And we really tried to heed the conventions at the time of coding agents that were where this agent loop was seen most often, where, yeah, there's kind of like a lot of sub-features, like rejection, seeing its thinking, seeing how it works.

Great. And now we have the butterfly life cycle on the canvas. Pretty cool. However, this was still not really enough because as cool as this was, it still felt like, I don't know, it felt like I was handing my keyboard to some other AI rather than someone collaborating with me.

Although this model has been used really well in a lot of design apps that use tldraw, things like LovArt or MagicPath, and in education especially, where you have this kind of tutor of, like, help me with my homework and help me fill out my

oh, gosh. Let's see if I can do this on the fly. Steve Ruiz, class, age, whatever. And you can kind of ask it to, like,

complete my D&D character sheet. Allright. And it'll kind of pick up what you're doing and fill out forms and do fun stuff like this. Maybe I'll come back to that as it kind of chooches along. What I really wanted is to bring the agent out of the sidebar and into the canvas itself.

Oh, I'm a fighter.

Nice. Nice. Allright. I'll take that. And so we did. And we did it with fairies, which maybe you saw, maybe not. These are like little guys on the canvas. You can kind of throw them around. They don't like to be held for very long.

They'll start freaking out. Yeah. Oh, yeah.

But you can do the same thing, like draw a cat or something like that. Now, putting the agents on the canvas have a whole bunch of interesting things. You can see the state of the agent,right? These are multiple agents that I'm kind of running.

In coding terms, these would be multiple terminal windows or something like that. Or this would be in Composer. But you can kind of see what they're doing in a way that hang on. I'm zoomed way out. I did all the sprites myself.

And not only can you kind of see its thinking, but you can see its action. You can see where in the project it's sort of like acting relative to the other agents. And these other agents can see each other of what they're doing.

So if I ask this one to draw a hat on the cat, and I draw this one, draw the cat's neck.

We missed the neck.

They'll get to work,right? And they're able to kind of work with each other's stuff at the same time. But we could also ask them to work together. So if I grab all three of these fairies, Fairys, Helen, and Joan, draw some more animals.

One of them will be elected leader. So this one is the leader. And it's going to go scout kind of what's going on on the canvas. And then it's going to create a to-do list. And it's going to delegate that to-do list to the other agents,right?

We were doing this in like December, October of last year. And we're figuring this stuff out at the same time that a lot of people were figuring out agent orchestration. This idea of, like, OK, how do we give them shared state?

How do we have a leader follower? How do we manage the fact that these things are essentially blind while they're working and prevent them from kind of overlapping in terms of what they're doing? And so you can kind of see the leader here isn't doing any of the work.

But it is going to kind of like observe. Oh, no. That's the leader. That's the leader. It's observing

and judging and establishing whether this is done or not and whether it's done correctly.

Still not enough,right? Fairies are fun. If you want to play with this, by the way, this is at fairies.tldraw.com. In the same way that MakeReal was a really good introduction

to just AI at all,right? Draw something, click a button. Fairies is a great way to talk about multiple agents kind of working together. And they can do real work. Let me try and grab a this is a big description of an e-book or something like that.

And if I summon my fairies,

make the wireframes for this app. Cool. And I'll just kind of let them get to work while we keep talking.

I started 10 minutes late. I'm going to take another five minutes before I jump. The next step for this one is to kind of give more access to the canvas to the agents.

And we started to kind of run into the barriers of safety. Like, what is actually safe to do with our hackable thing for users? Because we have a runtime API. You could just code against it,right? And AI is really good at coding.

So maybe we could do some sandboxed stuff. But no, because we need the DOM. We need kind of the browser as a way to see what's going on. We need to be able to generate screenshots, all this stuff.

So we decided to use our desktop app instead. So over the holidays, I threw together an app, Electron Wrapper, that wrapped tldraw. And I opened a port, essentially. I said, OK, Claude, make a little HTTP server and put up an endpoint.

And anything that gets posted to that endpoint, treat it as JavaScript and run it, which

is a terrible idea. Not a good idea. Don't do that on your app. However, for an offline desktop app that is file-based, what's the worst that could happen? You could hurt yourself, I guess, but you're not going to hurt the rest of me.

And look at them going. Building my little e-book reader. That's fantastic. Thank you, fairies.

So what does that give us, though? I'm going to skip the demo where, as you can imagine, I could say, hey, visualize this code. Make a diagram. Cool. Allright. I'm going to change the diagram. Update the code to match the diagram.

Easy,right? You can have these kind of, like, let's pull up the level of abstraction that we want. But the more surprising stuff was actually where I was like, OK, check this out. I'm going to draw a little user interface or whatever,right?

And I want this to be a leg length. And I want this to be a t-shirt color. And even though tldraw doesn't really have the ability to we don't have primitives for on hover or on click. It's not a fully thing.

This thing can write code against the editor. So make it interactive. And we'll see where it gets to. So far, the results on this have been really, really cool in ways that are super strange and disturbing. Because asking the AIs are like, sure, let's do some script injection,right?

That's the way that it documents itself is like, this is how you should do this. It has no qualms at all, by the way, changing stuff that's on your desktop, on your computer. If you've ever wanted to, for example,

one of our team, Max, was like, you know what? I don't like podcasts in my Spotify. I want to get rid of podcasts in my Spotify app. Claude, can you just do that? And it's like, sure. Let me go through the minified code of the bundle of the thing.

Let me just rip and tear. And it's happy to do it. It makes them happy. They like it. Oh, what the fuck was that?

I don't even know what you did. You made an HTML? What the

it created a new HTML site out of this? And this is the pointer? It's not even a slider. No. I want it in the tldraw. Allright. Yeah. We love it. It's blinking as well. I don't know if you caught that.

Come on. Do it. Yeah. There we go.

Let's see if it can come on. Let's go.

So yeah, there's really no limit to what it can do with a desktop app. And it's happy to do it in a way that I can almost tell that it would love to do this to websites. Like, let me just get my claws in there.

Allright. Come on. Come on. Still not working. We're

going to go. Set up the interactivity. Come on. This is going to be really fun. I think we're going to just release this.

I mean, it is released. But the notion that you can take

I love local-first apps. I love file over app. There's all these ideas that up to now have kind of been curiosities and almost like

hang on. Oh, come on. Oh, that's such a disappointment.

We're going to have to catch me later. I'll make it work. But now, actually, the idea of a local file-based thing that is able to expose itself

to Claude and agents locally in order to essentially script inject kind of motivates a lot of that stuff, which previously was idealistic, into like, well, that's the only way that you could do this. If you really want to maximize the agency in order to maximize what it can do and take the risk and take on that risk, then you kind of just need to hand that to the user and say, good luck.

I think OpenClaw does this pretty well. These are sharp tools. Have fun. Anyway, that is my agents on the canvas talk. Work continues. If you want to play with the fairies, I highly recommend it because it's super fun.

And you will find things that surprise you that have surprised me. They have IRC as well. Let me see. Yeah. Anyway, and if you want to follow along with tldraw, we are on Twitter, X,@tldraw. And then I'm@SteveRuiz. OK?

And I post a lot about this stuff. So thank you for going. Cheers.