The limits of design as context

R/GA’s BRDNA makes a compelling case for machine-readable 'design'. The harder question is whether the judgment behind design decision-making can ever be encoded. The most interesting argument here isn’t really about AI generating design. It’s about the infrastructure required for an agent to make useful design decisions in the first place.

Paul Aaron of R/GA's Addition studio argues that the quality of agentic design is largely a context problem. Give an agent a well-structured design system, machine-readable tokens, examples and carefully written skills, and its output improves dramatically. Their own systems illustrate the shift: what began as thousands of lines of conventional code has evolved into something closer to an English codebase, with one deployment consisting of 76% markdown.

That inversion feels significant. If software increasingly interprets intent rather than simply executing predefined logic, natural language starts functioning as part of the system itself. Design guidelines, principles and examples are no longer documentation sitting beside the product. They become executable context.

Where I’m less convinced is the leap from encoding a brand’s existing design language to encoding the judgment required to evolve it.

The BRDNA idea is essentially an attempt to make an organisation’s accumulated creative knowledge machine-readable: its components, rules, campaigns, examples and patterns of decision-making. That seems genuinely useful, particularly for production work. But describing this as a brand’s “DNA” risks implying that the important parts of creative judgment can be extracted from previous work and represented as instructions.

A design system can tell an agent what belongs. A set of skills can tell it how the organisation tends to work. Examples can demonstrate what has previously been considered good. None necessarily tells it when those rules should be ignored, when something should be removed, or when the right decision is to do something the existing system has never seen before.

Interestingly, the article almost acknowledges this in its discussion of the final 10%. Once the output gets close enough, humans return to Figma to select, remix, edit and refine it. That remaining gap is described as subjective, but I suspect it contains much of what we actually mean by design judgment: taste, restraint, prioritisation and knowing when the technically correct solution still feels wrong.

That doesn’t diminish the argument for agent context. If anything, it clarifies where these systems are most powerful. Encoding a design organisation’s accumulated knowledge can dramatically expand what an agent is capable of producing. But context and judgment are not quite the same thing.

The more interesting question is therefore not whether we can make a brand machine-writable. It’s how much of the judgment that created the brand can ever be written down in the first place.