Why anyone thinks the canvas is finished
The strongest version of this argument is not made by people who dislike Figma. It is made by Figma.
At Config 2026 they shipped Code Layers, which turn a design layer into an interactive code layer with one click or one prompt. Figma Motion exports real CSS, JSON and React-ready animation rather than a video of an animation. Code Connect links published components in a library to the actual components in your repository. There is an MCP server so agents can read design context directly, and agents now create and edit on the canvas using your own components and variables. Dylan Field's framing is that code is a design material, equal to images and vectors.
Read that list as a strategy rather than a changelog. Every one of those features exists to shorten the distance between a drawing and the running thing. A company does not spend a year of engineering closing a gap it believes is valuable.
Underneath that, the plumbing finally set. The W3C Design Tokens Community Group format stabilised through 2025, with $value and $type settling into the de facto syntax. A token file now travels from a design tool through Style Dictionary into a Tailwind config without anyone writing custom glue. The design system stopped being a document about the product and became a dependency of it.
The job market moved first, as it usually does. Linear, Vercel, Stripe, Anthropic and Cursor all ship through design engineers. Cursor's own listing asks for somebody equally comfortable in Figma and their code editor. The role is a designer who ships, and who owns the live source of truth rather than a description of it.
Put it together and the case is simple. If an agent can render a real component from your real system in the time it takes to draw a rectangle, the rectangle has lost its job.
The numbers point the other way
Then you look at what the market is actually doing, and it is not what that argument predicts.
Figma reported Q1 2026 revenue of $333.4 million, up 46 per cent year on year. That is the second consecutive quarter of accelerating growth, following 40 and 38 per cent before it. Paid customers reached roughly 690,000, up 54 per cent. Customers spending over $100,000 a year grew 48 per cent. Net dollar retention reached 139 per cent, the highest in more than two years, which means existing customers are spending materially more each year rather than quietly leaving. Full-year guidance went up, not down.
Whatever is dying does not bill like this. Businesses in structural decline do not accelerate, and they do not accelerate on expansion revenue from the customers who know the product best.
The research is less comfortable still. In a controlled study of sixty participants published at CHI 2024, Wadinambiarachchi and colleagues gave designers an AI image generator during ideation. The AI group fixated harder on the first example they were shown, and produced fewer ideas, with less variety and lower originality, than the group working without it. Not slower. Less original, measurably, in the one phase where originality is the entire point.
A 2026 paper by Imteyaz and colleagues names a second problem: design theater. Generative UI tools articulate sophisticated design rationale that does not correspond to what they actually implement. The system explains itself fluently and the explanation is decoration. If you have been reading that rationale as evidence the design was considered, you are being told a story.
Then there is the thing everybody has noticed without needing a paper. These tools reach for the same component library every time, and you can now recognise that design language across half the web. Left alone, they converge.
Field's own summary is the sharpest line anyone has offered against me. When code is a commodity, design is the competitive edge.
Businesses in structural decline do not accelerate.
Two different things are being called the canvas
Both of those sections are true, which usually means the question is badly formed. It is.
The canvas is doing two jobs, and they have almost nothing to do with each other.
The first is exploration. Somewhere cheap to think, where twenty bad versions can sit next to each other, where nothing is committed and being wrong costs nothing. A sketchbook that happens to be vector based.
The second is specification. The canvas as source of truth: the file that gets approved, the frame that gets redlined, the artefact an engineer opens in order to find out what to build. Design finishes here, then gets handed over there.
The first job is in excellent health, and the fixation research above is a defence of it rather than an attack. The second job is finished, and Figma's roadmap is a year of engineering aimed squarely at its funeral.
So ask which one your team actually runs on. If the honest answer is that the design file is the spec and the build is a translation of it, the question in the title of this piece is not rhetorical for you.
The translation step was the whole job
Before agents, the handoff had a real defence. A designer could not reasonably be expected to implement a responsive component, in the production system, with every state, keyboard support and a dark theme, in the time available. Somebody had to translate. Translation needs a specification. The canvas was the specification.
That defence has gone. A designer working inside the repository with an agent can produce the real component, in the real system, wired to the real tokens, in roughly the time it used to take to draw a convincing picture of one. Not because the agent has taste. Because the agent removes the part of the work that required a second person.
It is the same mechanism that is currently repricing agencies, and it turns up wherever an industry built a role around moving intent between two people. The role is not being automated. The distance it existed to cross is being closed.
A colour changed on a canvas is a proposal. A colour changed in the tokens file is a deploy.
Explore on a canvas, decide in the repo
So, no. Canvas design is not dead, and anyone selling you that is not reading the evidence. What died is the handoff.
Take the fixation study seriously and you should not begin in a prompt. Starting from generated output anchors you to the first thing you see, and the measured cost is variety and originality. The cheapest place to be wrong twenty times is still a canvas, and quite possibly still paper before that.
Take design theater seriously and you should not accept a generated interface because it arrived with an explanation. Judgement is the scarce input, and it has to come from somebody who can say why this and not that.
But take Code Connect and the token format seriously, and the decision has to land in the repository, because that is now the only place where a decision is real.
A colour changed in a design library is a proposal. The same colour changed in the tokens file is live: in every component, on every surface, through the build, in the next deploy. Those two actions look almost identical and they are not the same kind of event. One of them changes the product. The other asks somebody else to change the product, then waits, then gets implemented in three places out of four, then drifts.
Every design system I have watched drift, drifted for that exact reason. The canvas said one thing, the code said another, and the code won, because the code is what users actually get.
How to tell a design engineer from a designer with a GitHub account
If any of this holds, it changes what you hire for, and the job titles have not caught up. Four things worth checking.
Where does your design system actually live? If the authoritative version is a design library and the code is kept in sync with it, you have two design systems and one of them is lying to you. Pick the repository and make the canvas downstream of it.
Can this person change a token and ship it? Not whether they have used GitHub. Whether they can open the tokens file, make the change, understand what it cascades into, run the build, read the failure, and raise a pull request somebody will approve. That is a different skill from being comfortable near a terminal.
Ask what they would not let the agent decide. This is the best single question I know for the role. A good answer is specific and slightly opinionated: contrast ratios, motion durations, the empty state, whatever makes the product feel like itself instead of like the default. An answer along the lines of "I check its work" means they will ship whatever came out of it.
Ask to see something they made that does not look generated. The homogenisation is real and it is the failure mode of this entire way of working. Anyone who has genuinely done it will have opinions about where the defaults were wrong, and will be able to point at them in their own work.
Then the structural fix, which costs nothing and solves more than a hire will. Put the tokens in the repository, make them the contract between design and engineering, and give one person the authority to change them. Most of the drift I get called in to look at is not a skills problem. It is two sources of truth and nobody empowered to collapse them into one.
That part I can do with you, and it is usually a shorter engagement than people expect.