What AI Agents Need from Design

Coding agents – Cursor, Claude Code, GitHub Copilot, and others – have access to extraordinary resources for building software. They have been trained on vast bodies of production code. They can generate, debug, refactor, and compose with speed and accuracy that would have seemed implausible five years ago. For the logical structure of software, they are well equipped. They know what a button is. They know how to build a form. They know how to assemble components into a layout.

What they do not have – unless it is explicitly provided – is design judgment. Not judgment in the sense of preference (agents do not have preferences), but judgment in the sense of encoded brand intent: the knowledge of what this particular brand means to feel like, what this composition should express, what level of hierarchy this content demands.

When a brand publishes a design system – tokens distributed via Style Dictionary, components documented in Storybook, a Figma library shared with the team – a coding agent can, with the right tooling, consume parts of it. It can reach for the correct token values. It can use canonical component patterns. This is genuinely useful. Token-equipped agents produce fewer color violations and spacing inconsistencies than agents working from scratch.

But a design system gives agents vocabulary, not judgment. It says what the options are. It does not say which option fits. An agent with a token set still makes its own decisions about hierarchy, density, emphasis, and composition – and those decisions are not grounded in the brand's expression identity, because that identity is not encoded in the token set.

First, they need an expression identity – a structured representation of what the brand means to feel like. Not a mood board and not a guidelines document, but a machine-readable encoding of brand intent that the agent can query at the moment of decision.

Second, they need compositional guidance – principles for how to arrange elements, establish hierarchy, calibrate density, and make layouts feel considered for this brand in this context. Token systems encode individual values. Compositional guidance encodes the relationships between them.

Third, they need runtime access – the ability to invoke this guidance at the moment of building, not as a document to read in advance, but as a service available when the decision is being made. Humans carry all three of these in their heads after years of working with a brand. Agents need them encoded as infrastructure.

See also: what is expression infrastructure.

The instinct when agents produce aesthetically flat output is to improve the prompt – to tell the agent more precisely what the brand looks like, what style to apply, what tone to strike. This works at the margin. It does not solve the problem.

A prompt is a one-time instruction. It does not persist across agent sessions. It does not update when the brand evolves. It does not carry nuance from one composition to the next. And it places the burden of encoding brand knowledge on every person who writes a prompt – requiring them to carry in their head what should be encoded in infrastructure.

The answer to "what do agents need from design" is not better prompts. It is infrastructure that makes the prompts shorter, because the brand's expression identity is already there.

This set of needs – expression identity, compositional guidance, runtime access – has always been present. What has changed is the agent. Coding agents that build production interfaces at scale are new. The volume of design decisions being made without a human present is new. The gap between what agents can build and what agents can build well is now visible at a scale that makes it impossible to ignore.

Brands that rely on AI to build any part of their product surface are already encountering this gap. The ones that name it as an infrastructure problem – not a prompting problem, not a design-system-update problem – are the ones that will solve it.

See also: the post-static design era.