Overview
AI can generate interfaces faster than ever, but faster production does not automatically produce better software. In this interview, Paper founder Steven Haney argues that design is becoming more valuable precisely because generic model output makes products easier to build and harder to distinguish. Paper addresses this tension with an agent-native canvas powered by HTML and CSS, allowing agents to read, generate, and modify the same structures that designers manipulate visually and developers ship as code. The demonstration covers editable imports from live websites, multi-model image generation, customizable WebGL shaders, agent-generated layout variations, comments as agent instructions, and export to React or Tailwind. Haney also identifies recurring signs of unedited AI design—excessive bold type, too many font sizes, purple gradients, gratuitous cards, pills, icons, and dark mode—and recommends reducing visual noise through direct manipulation and deliberate curation. Paper itself uses agents selectively: heavily for marketing, coding assistance, and peripheral production, but cautiously in its precision-sensitive core product. The larger conclusion is that models will improve at design tactics, yet human designers remain essential for interpreting requirements, aligning stakeholders, judging context, and creating work that stands apart from prevailing patterns.
Sections
Higher-Order Insights
Implications that emerge across the product demonstration, design critiques, and company-building discussion.
- As generation becomes cheaper, differentiation shifts from producing an interface to exercising judgment over what should exist, what should be removed, and what best expresses the product's context.
- Agent-native software is not defined merely by adding a chat box; it depends on representations and interfaces that agents can manipulate accurately while remaining directly editable by humans.
- The emerging designer role combines generation, curation, product judgment, and implementation, potentially allowing one person to carry work from concept through production without a formal handoff.
- Open-source creative assets such as shaders and fonts can function as product-led marketing by giving the community something useful and culturally distinctive instead of purchasing attention through advertising.
- Visual simplicity is presented as a trust mechanism: deleting unsupported decoration signals prioritization and care, particularly in healthcare, finance, and other credibility-sensitive markets.
Technical and Workflow Details
Specific implementation choices and demonstrated capabilities of Paper's agent-native workflow.
- Paper uses HTML and CSS as its rendering engine, allowing agents to read CSS and send HTML directly to the canvas without an intermediate design-file translation layer.
- Canvas content can be copied as React or Tailwind, while an agent connected to the repository can convert a selected design into a new component that follows existing coding conventions.
- Paper Snapshot imports sections from a live website as editable layers rather than flattened screenshots, enabling direct visual iteration from the interface users currently see.
- The desktop application includes an MTP server that exposes operations such as discovering available fonts, reading page context, and writing HTML to the Paper canvas.
- Paper Shaders is an open-source WebGL animation library whose effects can be customized, blended, exported as React components, and used as inputs to image-generation workflows.
- The image-generation library can run four models through a rotating variety pack, accept image inputs, extract colors, vectorize results, and support rapid visual exploration.
- Comments, named frames, and current selections provide different targeting mechanisms for agents; comments are preferred when selection state may change during navigation.
- Agents can capture rendered screenshots during execution, detect issues such as collisions or text overflow, and iteratively revise their output.
- The team targets 120 FPS for the design experience and uses multiple human reviews for every line of core-product code because current agents do not yet satisfy its precision requirements.
- A planned prototype-management feature will embed externally hosted or local HTML prototypes in iframes on the Paper canvas so teams can organize them and attach feedback in one place.
Key Comparisons
Explicit contrasts between conventional workflows, agent-native workflows, model capabilities, and interaction modes.
- Custom rendering engines require agents to understand or translate proprietary formats, whereas HTML and CSS give Paper a representation models already understand, improving speed, token efficiency, and accuracy.
- Prompting is effective for broad generation and batch operations, while direct canvas manipulation is faster for precise visual edits such as deleting, dragging, or adjusting a specific element.
- Fast models are suitable for demonstrations and rapid iteration but make more mistakes; taste-oriented frontier models can produce more thoughtful visuals but may be considerably slower.
- Agents excel at generating alternatives, resizing, translating, and applying tactical rules, while humans remain stronger at stakeholder alignment, requirement synthesis, contextual tradeoffs, and deciding what should stand out.
- The original Maridata design carried energy and international character but felt visually uncontrolled; the agent-generated revision improved professionalism and trust while sacrificing some excitement.
Memorable Quotes
Statements that capture the interview's central positions on design, agents, quality, and company building.
- Paper is a design tool first. It's a human design tool first that happens to be really really good with agents.
- I think a lot of design is deleting.
- There isn't even a handoff anymore. It's just like one person making these changes and iterating and and shipping.
- The agents can speed you up. They can save you a lot of work, and they can do tasks like translation or resizing for you.
- What do you need next and just do the most straightforward thing you can possibly do to get there.