Design tools are being rebuilt around agents
Paper is an AI-native design tool built on the bet that agents will do much of the drawing. Rendering in HTML and CSS, which models already know from training data, makes the canvas a visual interface for agents and makes the code base the source of truth. Humans keep the design decisions, because design inside a company is decisions and stakeholders.
Takeaways
- Agents design best on HTML and CSS. Paper renders in HTML and CSS because models know both from training data. Custom rendering engines cost more tokens, run slower and land less accurately. Paper added an MCP server so agents write HTML directly, and the Claude Code boom pushed it onto the desktop.
- The code base replaces the design file. Handoff broke because designs were not built as intended and code drifted from design. Paper treats the code base as the source of truth for both, and Stephen says keeping two copies of a design system is impossible. Components are still to come.
- Models have design tells you delete. Models reach for bold weights, five to eight font sizes, pills, badges, two-pixel colour swoops and the purple gradient Linear popularised. Pull weights back, work with three sizes, check contrast, and never ship raw model output.
- Agents iterate, humans still decide. Stephen says agents will keep improving at tactics like letter spacing, and Fable seems more thoughtful than earlier models. He does not expect them to take over design decisions, because design in a company is decisions and stakeholders, perhaps in ten years for higher-order calls.
- Ask for many variations, then curate. Users leave agents running overnight to generate hundreds of variations and return to curate them. Selecting an element and deleting it beats re-prompting, and naming a frame or leaving a comment gives the agent its context.
- Paper sold itself before it shipped. Paper reached about 25,000 Twitter followers before it had a product by being values-aligned with designers, and its open-source Paper Shaders library serves as the marketing budget. Ramp spend data now lists it beside Figma and Sketch, and the team stays at 12.
Teardown notes on three user sites
- Legion Health has a generic pill icon, alignment and contrast problems that make the brand feel untrustworthy, and could use a YC badge as social proof. Stephen ran Paper Snapshot and Conductor on GPT-5.5 for three variations, then left a comment for the agent.
- SciTechs pairs a bold headline and purple glow with five font sizes. Asked through Cursor to reduce the frame to three sizes and no bold, it looked intentional.
- Moreta, a QR payments company, has energy but a hero that should work harder and a QR code that belongs above the fold. The site gets more trustworthy the further you scroll.