Code got infrastructure because code is text
Every software project runs on the same machinery. Version control remembers every change and who made it, tests fail a build before a person sees it, and review puts a second reader on every line. None of these tools writes code; together they let many hands work on the same code for years.
That machinery exists because code is text, and text can be diffed, merged, linted and executed. Once a craft lives in files a program can read, a program can hold its standards.
Design never got there. Its standards lived in a brand book nobody opened and in a designer's eye. The tools were built for a person at a canvas and made that person faster, but none of them turned the brand itself into something a program could hold, check and version. That was enough while every pixel that shipped passed someone who carried the brand in their head.
The design workflow is inverted
For decades software followed design. A designer drew the interface, an engineer built what was drawn, and every change passed between them. The handoff was where the brand was kept: whoever passed the file across carried the standard with it.
Agents broke that sequence. An agent takes an idea straight to running software: the landing page, the link card and the launch post come out of the same session that writes the backend. The canvas drops out of the sequence, and the handoff where the brand was guarded goes with it. The faster agents build, the harder a brand is to keep intact, because nothing in the new sequence holds the brand between one piece and the next.
The wrong problem
Most design tooling still treats the canvas as the centre of production and the agent as a faster pair of hands inside it. The rest of the industry is working to make agents better designers. That is the wrong problem. Agents already draw good pages. What none of them has is one brand they all work for.
The canvas-first workflow defends a position production has already left. Its authority came from sitting upstream of every build, and builds now start without it. Making each agent a better designer improves every piece in isolation and leaves the brand where it was, in no file any agent reads.
Designers move upstream of every agent. The decisions, references and musts and must-nots they set are what every agent executes, on every piece it makes.
What breaks without it
Start with the intent gap. A person's idea passes through four steps before it ships: what they see in their head, the brief that puts it into words, the agent's reading of that brief, and the rendered design. Something is lost at each step. Designers closed this gap with pictures, with references to react to, options drawn on real formats and a finger on the one they meant. Agents start from words alone, and the intent gap widens at every step.
Then there is what we call slot-machine design. Ask an agent for a landing page and you get a good one. Ask again tomorrow and you get a different good one, and neither is your brand, because each session reads the brief afresh, interprets it afresh and renders afresh. One great pull is luck; a brand means the same result on every pull.
Brands also drift across agents and months. Claude Code builds the site in March. By the time Codex makes the launch cards in May and Cursor touches the docs in July, each has read a different slice of context, and none of them knows what the others decided. The colour shifts a shade and the type steps a weight, and nobody chose either change; it accumulated.
None of these failures is about taste. A stronger model draws a better page and still starts every session from zero. The fix is structure.
The design system has a new reader
Until now a design system was reference material. Designers and front-end engineers consulted it, and a person decided in each case how a rule applied. It could stay loose because its readers could ask and remember.
From now on its first reader is an agent, which executes it on every piece it makes, and the design system becomes a production protocol. Every gap in it is filled afresh by each agent in each session, which is how slot-machine design and drift begin. A protocol that many agents execute has to be versioned, readable by every agent and checked by programs.
What the design layer must be
A brand is a sequence of decisions, so the layer has to be versioned. Each decision gets a version, a date and a note, and every piece knows which version it was made under. When the brand changes, the system names every piece that is now stale, and a brand change becomes as reviewable as a code change.
It has to be readable by every agent. The brand lives in one plain file the agent reads before it draws, holding the decisions in words, the exact values measured from a real render, and the client's own musts and must-nots. That file sits in the brand's own repository, next to the code it styles, where review, branches, permissions and history apply to it as they apply to code, and no vendor's context window owns it.
It has to be checked by programs on every rendered piece, for every model. A headline clipped at the edge, a fallback font where the brand face belongs and grey text nobody can read are never matters of taste. Taste stays with the model, which improves every year; a written rubric would cap it and become a target the work bends to. The layer holds what can be proven and leaves the rest to judgement.
The control plane
Briefatto is the design control plane for AI agents. Whichever model generates and wherever it runs, every piece goes through one brand system that constrains it before it is made, checks it after it is rendered and versions every decision behind it.
Generation improves on its own schedule, inside every model. The assurance that the latest piece, made by a different agent in a different month, belongs to the same brand as the earliest lives in no model at all. A brand outlives every model and environment that works on it, so that assurance sits in the one layer they all pass through.
Why it is open
A protocol wins by being adopted by every agent and every project. A control plane that only some agents read controls nothing, so this one installs with one command under any model, and its rules sit where the people who make the brand can read them.
Open work compounds: our checks run Impeccable's open-source detector. The brand file is plain Markdown any tool can parse, and every check is code anyone can run on their own machine.
What Briefatto Lab is building
Briefatto Lab builds the design control plane for AI agents. Briefatto Board is its first release. It is a skill and a command-line tool that take a brand from references, directions and a client's pick to a frozen, versioned DESIGN.md, or adopt the brand of a site that already exists. It tells every agent in AGENTS.md and CLAUDE.md to read the guide first, checks every rendered piece against a quality floor, and names the stale pieces after every brand change. One day, every pixel an agent ships will pass through Briefatto.
In a blind test on the open-source project Issue Composer, all four judges ranked the Briefatto Board kit first overall, ahead of Impeccable, Taste Skill and Hallmark (see every kit side by side). It is Apache-2.0 and works with any coding agent that reads AGENTS.md.
npx briefatto-board install