How To Design In The Agent Era

Why does it matter? Because the design tool built for human handoff turned out to be the only one AI agents can actually use.

Paper's most powerful AI advantage wasn't designed for AI at all. The rendering choice that made agents love it — HTML and CSS as the underlying engine — was originally built to fix designer-developer handoffs. The fit with agents came later, as a surprise. What this conversation reveals is that the design tool category is making a structural mistake, and Paper stumbled into the right architecture before anyone knew to ask the question.

• Agents working in HTML/CSS spend fewer tokens, run faster, and hallucinate less than agents fighting a proprietary rendering engine • AI-generated design has a specific fingerprint — bold fonts, purple gradients, too many font sizes — and stripping it out takes minutes • Design becomes more valuable as software gets cheaper to build, not less • The new workflow has inverted: leave comments before bed, wake up to hundreds of agent-generated variations to curate

Paper's HTML/CSS rendering wasn't built for agents — and that's exactly why agents love it

Most design tools fight their own rendering engines. Custom formats agents have to decode before acting add cost at every step: more tokens, slower execution, more hallucinations. Paper sidesteps this entirely — not through clever AI integration, but through a decision made years earlier for an entirely different reason.

"We actually did it to help with designer-developer handoff," Haney explains. "If the designer can have this constraint-free system where it feels like Figma or Sketch, but by the way, it's HTML..." The AI advantage was discovered after the fact. "It turns out if you build a great handoff for humans, you've also built a great handoff for agents because agents understand HTML and CSS and it's in their training data."

An agent working in Paper reads CSS directly, writes HTML directly, needs no translation layer. The canvas and the codebase already speak the same language — token spend drops, execution speeds up, accuracy holds. Paper's competitors see slower, costlier, less reliable agent work because their custom rendering engines sit between the agent and the format it actually trained on.

The question this opens for any "AI-native" tool: not how polished the AI integration is, but whether the underlying format already lives in model training data. The best AI handoff turns out to be one designed first for human clarity.

AI-generated design has a fingerprint — and knowing it lets you strip the look out in minutes

Purple gradients, icon badges, bold everything: these aren't aesthetic choices, they're tells. Haney has catalogued the AI design fingerprint with uncomfortable precision, and once you know the list, you can't stop seeing it.

Font weight is the biggest offender. "Models love bold styles. Try to pull your font weights back as light as you can get and it'll just magically look better designed." Right behind it is font size proliferation — agents reach for six or seven distinct sizes; collapsing to three removes most of the noise immediately.

Then comes the decorative scaffolding. Numbered badges on card corners. Icons in all-caps headers with extra letter spacing. The two-pixel color swoop on every card. "A lot of avoiding the agent look is pulling back... a lot of design is deleting." Purple gradients get their own category: Linear made the aesthetic great, the models encoded it, and now it appears everywhere, signaling nothing about taste and everything about prompting.

The corrective is mostly subtractive — three font sizes, weights pulled back from bold, decorative widgets deleted. In Paper, that deletion is direct: select something, hit delete, the layout snaps closed. You're not prompting a correction and waiting; you're editing.

The stakes are higher than aesthetics. In products handling money or health, visible vibe-coding erodes trust before the user reads a word.

More software than ever is being shipped — and almost none of it is getting better

Quantity is exploding; quality is flat. "There's more software than ever being created and engineering teams are moving so fast now," Haney says. "And I haven't felt like the software is getting better necessarily." He doesn't hedge this.

When the baseline is AI-generated mediocrity, exceptional design becomes the clearest differentiator in a market where features are commoditizing faster than anyone can track. "If you look at every great company of the last 10, 20 years, basically all of them have exceptional design." He can't name an exception.

The consequences reach beyond aesthetics. Design influences who applies to work at your company, who trusts you with their data, who roots for you when you're still missing features. Founders most tempted to cut corners on design — those in the most crowded markets — are exactly the ones for whom it matters most.

The blunt version: "I don't think it helps you stand out from the crowd. I think it makes you look like one of a million other projects." In a world launching millions of projects every month, that's a real cost, not a philosophical one.

The designer-developer handoff was never solvable — Paper just stopped maintaining two copies

Two copies of the same design, perpetually drifting apart. "We see so many teams maintain two copies of their design. And it's not a tractable problem to keep them in sync. It's impossible."

Paper's answer isn't better syncing. It's eliminating the second copy. Because Paper files are HTML and CSS, the design and the codebase are the same artifact. Agents can read the canvas and write directly to the repo. "The code is the source of truth. Both systems know how to work on that same code base."

What this looks like in practice surprised Haney. He expected handoffs to get smoother; instead, they've disappeared for many teams. "We see a lot of folks are just kind of like this is the whole loop now — like there isn't even a handoff anymore. It's just like one person making these changes and iterating and shipping."

For new companies, the upshot is clear: start with code as the canvas from day one. The organizational overhead of the two-copy pattern — two systems, two sources of truth, two people whose job is keeping them reconciled — is a tax that compounds invisibly until it becomes impossible to pay.

Leave comments at midnight, curate in the morning — the new design loop has inverted

The practice is showing up at scale: designers leaving comments on Paper canvases before they log off, agents running overnight loops generating hundreds of variations, designers returning to curate rather than create.

"We see a lot of folks leaving these running overnight," Haney says. "They're setting up loops of an agent just generating a lot of feedback and then just creating hundreds of variations." The volume caught Paper off guard — their servers are straining under the load.

The creative inversion this enables is real. When variation generation is cheap, the bottleneck shifts from production to judgment. You're no longer trying to arrive at one good design — you're filtering from many. "I find that it'll do things that I know but I wouldn't have thought of because I was too busy. It lets me scale myself a little bit."

Practical consequence: the quality of your comments matters more than the quality of your first draft. Specificity about what to elevate, what to avoid, and what the design needs to communicate turns overnight agent loops into a genuine creative multiplier — not a pile of slop to sort through.

Agents will master letter spacing. They won't master what your audience actually needs.

Execution-layer design — contrast ratios, font weight, spacing decisions — will become reliably agent-territory. Haney sees this coming and welcomes it. Fable already shows more apparent thoughtfulness than earlier models, and the trajectory will continue.

Strategic design is a different problem entirely. "Do I ever think it's going to replace a human making these design decisions? No, no way. Because a lot of design is actually a very human task of like decision-making in an organization." The pixels at the end are the output of a process — requirements, stakeholder alignment, competitive positioning — that happens between people arguing about what the product should be.

"Was this showing value prop or not? That stuff's more important... Can a model make that decision? I think we're pretty far away from that." An agent can't tell you whether your market needs restraint or excitement, trust or energy. It can't read the room.

The division of labor this implies: use agents hard for execution work — resizing, translation, variation generation, catching contrast problems. Guard strategic decisions as human territory. Knowing which layer you're operating on is the judgment agents can't supply.

25,000 followers before a product existed — because values-aligned audiences don't churn on features

A year before Paper shipped anything, Haney had 25,000 Twitter followers. The approach was almost aggressively simple: go on podcasts, talk about caring about design, actually mean it.

"Being values aligned with your market, they will forgive so much about your product missing things. They'll wait for them to come out. They'll root for you if they can tell you're authentic." In a category Figma has dominated for years, that forgiveness bought time that feature releases couldn't.

The results are now measurable. Paper appears in real subscription data alongside Figma and Sketch — "the first company since Figma has come out to make a dent in that market at all." They got there by being visibly genuine before being feature-complete.

For creative tools, this mechanism is unusually strong. Designers talk to each other, share what they use, wear the merch. Someone at a dinner in New York City was wearing a Paper Mono shirt when fewer than a hundred of them exist in the world. That's not a marketing outcome you buy. It's one you earn by being publicly credible in the domain you're building for.

Paper reads every line of code before shipping — and still beats well-funded competitors with more features

Twelve people at Paper read every single line of code before it ships. Usually more than one person reads each line. In 2026, this is practically contrarian.

The logic is architectural: design tools require 120 FPS precision and edge-case reliability that agent-generated code cannot yet consistently deliver. "You can't vibe code a design tool at the level of a Figma today. It's a lot of really precise, really careful work. It's very multidimensional in terms of the systems."

Competitors have shipped faster and accumulated more features. Users still prefer Paper. "We have competitors that have every single feature. But they don't have the attention or people don't care." Haney's explanation is direct: designers feel care in software. Precision is legible. Sloppy tooling signals that the team doesn't respect the craft it's supposed to support.

There's a FOMO dimension he acknowledges honestly — every other company is moving faster by not reading their code. His bet: quality compounds in ways a feature list doesn't. For any precision-critical software — rendering engines, financial interfaces, medical tools — treat agent-generated code as a strong draft requiring expert review, not a finished output.

The scarce thing in an AI world isn't capability — it's clarity

The thread connecting every insight here points somewhere the conversation never quite names. As AI collapses the cost of making software, the scarce resource isn't capability — it's clarity. Clear architecture. Clear typography. Clear values. Clear code review. Paper's entire approach, from its rendering engine to its pre-launch Twitter presence to its handoff philosophy, is a single bet: that being legible — to humans, to agents, to users — compounds over time in ways that simply shipping faster never will.

Legibility is the moat.

---

Topics: design tools, AI-native software, designer-developer handoff, agentic workflows, startup growth, product design, founder market fit, creative AI, design systems, typography