Canvas vs. Chat — Why Spatial Interfaces Win for AI Collaboration

Chat dominated the AI era because ChatGPT made it intuitive. You ask a question, the model responds, you ask a follow-up. One-dimensional, turn-based, easy to understand. Every AI product since has inherited that paradigm: Claude, Gemini, Copilot, all chat.

But Steve Ruiz ran an experiment late in 2025 that challenged the assumption. He built AI agents that lived on tldraw’s canvas — spatially located, moving around, working on tasks in parallel. The result revealed something the chat-first world has been missing: canvas isn’t just better for human collaboration. It’s fundamentally better for AI collaboration.

The problem with chat is that it’s linear. Messages arrive in a queue. Everyone waits for the previous response before the next action begins. That works fine for Q&A. It breaks down as soon as you have multiple agents trying to coordinate, multiple tasks running in parallel, or a human wanting to jump in and contribute without disrupting the workflow.

A canvas removes those constraints. Multiple agents can work in parallel, in different areas of the same space. They can see each other’s work, build on it, and coordinate without waiting. It’s the difference between a conference call (chat) and a whiteboard room (canvas).

The Multi-Agent Problem Chat Can’t Solve

When Ruiz experimented with agents on the canvas, he quickly realized something about the limitations of chat: “Having group chats with an AI actually isn’t a very good experience. Group chats in general aren’t a very good experience because they’re one dimensional. Every message kind of pushes to the end of a big list.”

Compare that to a designer working with a colleague on a Figma file, or three engineers collaborating on a tldraw whiteboard: they can work on different parts simultaneously. If one person is working on the header, another can focus on the sidebar. Nobody is blocked. Nobody is waiting for turn-based responses.

The same principle applies to multi-agent systems. An agent team tackling a complex task — say, generating a 10-scene visualization based on a film, or building a multi-stage marketing campaign — needs to parallelize work. One agent can be fetching data while another is generating images while a third is evaluating quality. Chat forces serialization: one response at a time, one agent per turn.

“On the canvas, why the canvas is one of the reasons why the canvas is so popular in situations like Figma or Miro for collaboration is that everyone can kind of be working at the same time in parallel with each other,” Steve explains. “And there’s really no limit on how many people that paradigm can support. Because if it’s busy, you just go somewhere else, right?”

Spatial Presence Changes Agent Coordination

Ruiz’s implementation added something chat systems don’t have: spatial location. Agents existed at coordinates on the canvas. You could move them, select them, have a conversation with one or all of them.

That seemingly simple affordance solved coordination problems that chat architectures struggle with:

Problem: Where is the agent working? On a canvas, if agents are near each other, there’s a natural assumption they’re working on the same thing. They’re spatially grouped. You can see task dependencies by drawing connections between elements. Chat forces you to manually specify scope and context in every message.

Problem: What is the agent doing right now? In the fairies experiment, agents had visible states: working (animated motion), reviewing (hands on hips), thinking (distinctive pose). You could glance at the canvas and know. Chat gives you a spinning loader.

Problem: Can multiple humans and agents work together without collision? On a canvas, a human can draw something in one corner while agents work in another area. They’re not fighting for the same message queue. They’re collaborating in parallel spaces.

Why This Matters for Real AI Products

This isn’t a theoretical problem. Companies are already trying to solve multi-agent workflows for marketing, content creation, and product design. Most of them are doing it in chat, which means they’re building complexity on top of a linear, turn-based architecture.

Take a marketing workflow: you have one agent generating copy, another generating images, a third evaluating whether the copy matches brand voice, a fourth optimizing for SEO. That’s four sequential steps in chat. On a canvas, you draw four connected boxes (or in tldraw’s case, you use tldraw Computer, which is exactly this — computational nodes on a canvas), and all four can run in parallel with data flowing between them.

Ruiz demonstrated this explicitly: “So this is kind of like a notation language for graphs and stuff. In order to do this, they need to know a lot about mermaid, which I didn’t tell them about. They need to understand what it is and the relationships between all these different things. But this is much more ambitious task. So they’re starting by getting rid of… clearing the canvas a little bit. And then now they are getting to work on the various parts of the diagram.”

The agents self-organized because the canvas gave them spatial context. No agent had to say, “Okay, I’m going to step aside now.” They just saw space and started working in it.

The Hidden Insight: Collaboration Principles Transcend Medium

The deeper lesson is that principles of good human collaboration — parallel work, spatial grouping, visible context, asynchronous contribution — don’t disappear when you replace humans with AI. If anything, they become more important.

“All of the things that make a Canvas good for collaborating with other people make it good for collaborating with AI,” Steve says. “So as an example, when I have a conversation with a model or something, if I’m doing some work with whether that’s ChatGPT or whether that’s like a coding agent or something like that, there’s not a lot of room in that experience for anyone else.”

This is why Ruiz designed tldraw Computer the way he did: as spatial nodes connected by arrows, not as a chat interface. It’s why the best multi-agent systems in the wild (N8N, Make, Zapier) use visual workflows, not chat. And it’s why Figma, Miro, and other collaborative tools remain generative AI’s natural home — they’re already built for parallel, spatial work.

What Chat Got Right (And Why It Won’t Be Replaced)

This isn’t a prediction that chat will disappear. Chat is perfect for Q&A, for exploratory thinking, for the “talk to an AI” use case that made ChatGPT popular. It’s intuitive. It requires no spatial reasoning.

But chat’s dominance in the AI world is largely an accident of timing: ChatGPT happened to be chat-shaped, so everyone built chat. As AI systems become more sophisticated, as multi-agent workflows become common, as coordination matters more than turn-based conversation, the architectural limitations of chat will become obvious.

“I think that Canvas has a place, just like I think terminal apps have a place,” Steve says. “I guess it’s question about for whom and in what setting, and what are the different constraints of those different platforms.”

For agents collaborating with agents, especially at scale, canvas wins.

FAQ

Isn’t chat better because it’s simpler for users to understand?

Chat is simpler for single-turn Q&A. But once you need coordination, state visibility, or parallel execution, chat becomes harder, not simpler. You’re manually managing context in messages instead of letting spatial organization do it for you.

Does every AI application need a canvas interface?

No. Q&A systems, writing assistants, and exploratory tools work fine in chat. But anything involving task coordination, multi-step workflows, or agent teams becomes easier on a canvas.

Can you build multi-agent systems in chat?

Technically yes, but you’re working against the architecture. You have to manually track context, serialize steps, and manage coordination through prompt engineering. A canvas lets the medium itself enforce good practices.

What’s the difference between tldraw Computer and visual workflow tools like N8N?

tldraw Computer is more expressive and visual-feedback-heavy (you can draw on the canvas, agents can see and modify drawings). N8N is more structured (predefined node types, more rigid workflows). Both use spatial canvas architecture instead of chat.

Could chat interfaces evolve to handle multi-agent coordination?

Theoretically yes, but you’d essentially be rebuilding canvas features (parallel threads, spatial grouping, visible context) within chat. At that point, you’re just reinventing the problem.

Why did Steve’s “fairies” experiment use agents on a canvas instead of a chat group?

Because having three agents trying to collaborate in a chat thread is disorienting: Who’s responding to whom? Which agent is working on which part? The canvas makes it obvious: each agent has a location, visible actions, and spatial relationship to the task.

Does this mean chat AI is a dead-end?

Chat is perfect for what it does. But if you’re building systems where agents coordinate with each other, chat is a limitation. The best hybrid might be: chat for exploring ideas, canvas for executing them.

Can a human use a canvas without understanding spatial reasoning?

Some people find canvas interfaces harder than chat. But tldraw’s success shows that with good design, spatial interfaces are learnable. And the payoff — being able to see and manipulate multiple elements at once — justifies the learning curve.

Is this why AI coding tools like Cursor don’t use chat?

Partly. Cursor uses a chat-like interface but organizes content spatially (code on one side, chat on the other). It’s a hybrid that acknowledges chat’s limits while keeping its simplicity for certain tasks.

What’s the future of AI UX if canvas wins for agents?

Likely a mix: chat for exploration and refinement, canvas for coordination and execution. Similar to how creative professionals use both chat-based prompting and spatial design tools depending on the task.