The canvas is the prompt
A year of asking what AI in Miro would look like if it were built from the canvas rather than borrowed from a chat app, and the four experiments that answered it.
Most AI products today look the same: a prompt box, often in a panel on the side, and a conversation that scrolls upwards. You trigger your agent only by sending a prompt. When we brought AI into Miro, we brought that pattern with us, because it's how the industry defined interaction with AI - through chat. This article is about the year we spent questioning it.
The question I kept coming back to was simple. Miro is an infinite, multiplayer canvas. What would AI look like if it were built from Miro's DNA, rather than borrowed from a chat app? My answer, after a lot of experiments across product, design and engineering, is that in a product like Miro the canvas itself should be the prompt and the response. The canvas should be the chat. And the triggers shouldn't be limited to the user having to prompt an agent (that already feels like a boring job anyway, doesn't it?).
There's one more tension running through this story. Miro is multiplayer first, but most AI patterns we saw in the wild are single-player. And you need to enable both.
The problem
When we talk to our users, there are clearly two reasons people choose Miro: to visualise something or to collaborate. They come for the infinite canvas. Nobody comes to Miro for a chat.
That has had a big consequence for Miro's AI. A chat panel is a single-player, text-first surface that sits next to the canvas rather than on it. Whatever the AI does, you experience it as a conversation in a sidebar your team can't see. And users expect AI to be just as good as a human. In the context of a multiplayer collaborative canvas, the AI in Miro had not met that bar until this year's October 2026 releases, which were presented at major keynote events in San Francisco, London, Tokyo and Sydney. That was our first step towards the vision, but the AI was still not Miro-native. Our AI design team - Kasper Marx Andersen, Mark Boyes-Smith, Josh Kriese, Robert - has explored many concepts to achieve the vision, and I have especially invested my time in testing a radical approach, where there are no chat panels, just the canvas, to see where it gets me. This is the story of us exploring all these concepts and what I learned along the way.
Experiment 1: a multiplayer chat panel
The first instinct was to keep the chat panel but make it multiplayer. In July we built it in our Miro prototype: a multiplayer chat panel with thread-based conversations that several people could have with other human or agent collaborators on the same board. It had controls to switch between private and shared conversations, because users would naturally expect them. It basically felt like bringing Slack to Miro. Conceptually, it felt quite close to Figma's recent AI chat release too. And that made it smell fishy. It felt like entering the vicious cycle of copying OpenAI and Anthropic again (oh, I have spent enough time in AI to really know when it happens, and I can assure you, it never ends well). AI was still stuck in a narrow panel, while 80% of the product surface was still the "old Miro experience".
So I ran a couple of user tests. No big reactions. Mostly scepticism, or surprise that Miro even has a chat: "I don't come to Miro to chat." It wasn't surprising, but it was necessary for us to really validate it and move on. A multiplayer chat panel asks people to adopt a new chat tool inside a canvas tool, and the experiment showed it wasn't an ask they wanted (and fun fact
The sharing problem
The second problem is the challenging one for all single-player AI products (so basically the current industry standard; at this point it doesn't seem that anybody has reached adoption for multiplayer AI). It's the context problem, the "who knows what, who sees what" problem. As soon as anything becomes multiplayer, it explodes in complexity. The moment other humans enter the interaction with AI, users start asking: "Will my context leak from my private chat into the shared one?", "Does the AI remember what I asked in private when my colleague asks in public?", "I don't want others to see how I prompt."
The amazing thing about the foundation Miro has is that it's visual and it's multiplayer. And this foundation has already solved these problems, if the canvas is the "chat". "I see what you see, and I know, what you know", and I stay in control, and that control is the sharing settings. I want a private AI interaction? I just don't share the board (same as with humans). I wanna get AI to help in a jam session? Get the humans in! As a user I stay in control, and lack of control is exactly the biggest adoption barrier for AI
Experiment 2: agents in comments
If people don't come to Miro to chat, where do they already talk? In comments. Comments are like a chat, and they're already multiplayer. They're attached to a specific thing on the board, so they carry its context, and (fun fact) comment usage is one of the features that help drive adoption.
So the thinking evolved from "bring humans to AI" to "bring AI to humans" - conceptually much closer to our product DNA, and to making agents feel like equal collaborators to humans. You leave a comment on a frame, tag the agent, and it picks up the task right there, in a thread everyone on the board can already see. And we actually shipped this!
Comments aren't the whole answer, though. On big boards, finding the right comment thread becomes a scavenger hunt. Comments are right for work that should stay attached to the board, but we might also need something lighter.
Experiment 3: cursor chat for agents
In October, @Tilo Krueger and @Bhavika's team introduced cursor chat for people: press a key, type a short line, and it floats next to your cursor for your collaborators, then disappears. It was built for humans talking to humans: quick, ephemeral, right here, right now.
At the same time, decision models started popping up, so I started experimenting with them for this. The ephemeral cursor chat felt like the perfect use case: when does AI react to these messages (and when doesn't it - surprisingly, that's much easier to define)? In a "quiet cursor chat" concept I compared three triggers: say its name, let a small decision model decide whether you're talking to it, or both. Naming alone is predictable but misses implicit asks. A model alone catches those, but then naming the agent isn't guaranteed to work. In the prototype the agent stays quiet on a low score, offers to help above about a third, and only acts when the model is very confident, around 0.86.
Experiment 4: Cursor Town
All of these experiments still assumed the human starts every exchange. And I wanted to radically rethink AI triggers. So I tested the radical opposite: no chat panels at all, all interaction triggered from user behavior in the canvas.
I called it Cursor Town: an infinite canvas where a team of agents works with you towards your goal. You write the goal on the board, and the agents break it down, work on it, and show their thinking as content on the canvas. You steer them like you do your colleagues: drop a sticky note, leave a comment, draw a shape. "I tell you, you do" becomes "we work towards a shared goal."
Each agent has a role, like the Devil's Advocate who raises risks, or the Chief Drawing Officer who defines the visual layout for other agents to respond. But the personas aren't the important part. What matters is that the agents react to everything that lands on the canvas, and fast. A small decision model picks what agents should do (approach user? respond to user? continue working? delegate subtask?), in a fraction of the time a full model would need to read the whole board. That speed is what makes it feel like collaborating with humans and agents.
The best part is that none of this needs new building blocks in Miro. It's a remix of functionalities that Miro already has. I recently heard of a user who already works with their Claude this way: prompts on sticky notes, which Claude picks up via MCP. For them, the canvas is already the prompt. If it seems like some power users might be already hacking their way into this experience, when would the industry catch up? Probably sooner than later (hey, it's AI after all!).
Reflections and what's next
The learnings from all this are the classics.
What to build is still the hard decision. AI makes it easy to build things fast, but building the wrong solution efficiently is still building the wrong solution. The multiplayer chat panel worked, and we could have spent months polishing it. A couple of user tests told us nobody needed it.
Start from the radical solution. As a designer who thinks a lot about product vision, I've learned that the biggest constraint is often the product itself: its limitations, existing solutions and dependencies. Design with all of those in mind and you soon end up just improving the system instead of thinking big. So start from scratch, question everything, and be radical. Don't try to build the bridge from today to tomorrow before you defined what you want tomorrow to look like. Don't dilute you ideas, don't shoot them down in your head, just build. Build and see where they take you. And they will take you places!
It's also the way out of the vicious cycle of chasing the big AI players (though let's be honest, you'll always be chasing a bit). Copying them was never the reason your product succeeded, so it won't suddenly bring more growth. Make AI native to your product's DNA instead.
Experiments 3 and 4 haven't been tested with users yet, so even though I'm excited about them, they might turn out to be wrong too. That's fine: the point is to find out quickly. I remain a strong believer in multiplayer AI and in the power of infinite canvases to accommodate this interaction.