Coding agents can build fast, but they miss the design rationale
According to a new post on the Vercel blog, the company has identified a critical blind spot in how coding agents approach product design. While agents can replicate existing UI patterns, match component styling, and even follow established conventions, they fundamentally cannot grasp the reasoning behind those decisions. That reasoning — why one design choice won out over another — lives in design reviews, PR comments, Slack threads, and in the institutional memory of the people who were in the room.
Vercel's engineering and design teams found that agents are excellent at producing working UI quickly. They can copy your product's style, match its patterns, and try to follow its conventions. But code only shows agents what shipped, not why one component, phrase, or interaction became your standard. For an agent, context that isn't in the codebase simply doesn't exist.
This is a problem that goes beyond Vercel. As AI coding tools from GitHub Copilot, Anthropic's Claude, and Cursor become more capable, developers are increasingly relying on them to generate entire frontend experiences. But design is not just about assembling visual elements; it's about making intentional, often collaborative decisions about user behavior, accessibility, and business goals.
Why this matters for AI-powered development
The implications are significant for any team using AI to build products. When an agent generates a button that matches your brand colors but places it in a context where users typically need a confirmation dialog, it's not making a mistake — it's following patterns without understanding the problem it's solving.
Vercel's approach to teaching agents product design involves capturing the decision-making process that surrounds code. Instead of just feeding the agent a style guide, they incorporate design review notes, PR comments, and even Slack summaries into the agent's context. This means the agent learns not just what was built, but why it was built that way.
For example, if a design review revealed that users were confused by a checkout flow and the team moved the order summary to the left, that rationale becomes part of the agent's knowledge base. The agent can then apply that same logic when generating new checkout screens, rather than blindly matching the first pattern it finds.
Practical lessons for developers and teams
For developers and businesses relying on AI coding tools, Vercel's approach offers several actionable takeaways:
- Document design decisions outside code. The most valuable context for agents lives in design reviews, PR comments, and chat threads. Start capturing that rationale explicitly.
- Treat agent knowledge as a living artifact. Just like a codebase, the context you feed agents needs to evolve with your product. Old reasoning may become irrelevant as user needs change.
- Don't assume agents understand intent. Even if an agent produces visually accurate output, validate that it aligns with the design's purpose, not just its appearance.
- Invest in structured design systems. A well-documented design system with clear usage guidelines helps agents make better decisions, but it's not a substitute for sharing the reasoning behind those guidelines.
Broader implications for AI in design
Vercel's work touches on a larger debate in the AI community: can agents ever truly understand design intent, or will they always be pattern matchers? Some experts argue that as models improve, they'll naturally infer intent from code context. Others believe that design is inherently human — it's about empathy, cultural nuance, and business strategy that can't be encoded in training data.
What's clear is that for now, the gap between generating UI and understanding design is real. Vercel's approach of explicitly feeding agents design rationale is a pragmatic stopgap, but it also points to a future where AI tools are not just code generators but active participants in the design process.
For teams using AI coding assistants, the lesson is simple: the better you are at documenting why you made design decisions, the better your agents will perform. And if you're building AI tools for development, incorporating design rationale into agent training data could be the differentiator that separates a tool that generates usable UI from one that generates truly thoughtful products.
Source: Vercel Blog. This article was produced with AI assistance and reviewed for accuracy. Editorial standards.
