Study · Updated July 2026
Design tools aren’t the future. The terminal is.
Two years ago almost nobody directed an AI agent from the command line. Today hundreds of thousands do, and about one in four are building a user interface, with no design tool anywhere in the loop. Here is what the data says, and where it points.
What changed
For most of software’s history, building an interface meant two separate places: a design tool where you drew it, and an editor where someone rebuilt it in code. AI help, when it existed, was invisible autocomplete. Around early 2025 a different shape appeared. You hand an agent a set of written instructions, and it builds and edits the project for you, from the terminal. That instruction file is the fingerprint. It only exists when someone is directing an agent, so counting the files counts the people working this new way.
The old way
Two places. A handoff loses detail in between.
The new way
One loop. No handoff.
The takeoff
When each repo first committed its agent-config file, by quarter. Near zero before 2025, then the jump from 2025 Q4 to 2026 Q1 (29 to 103). The dominant conventions (Claude Code, AGENTS.md) are a 2025 and 2026 phenomenon; Cursor has a smaller tail reaching back to 2024.
* 2026 Q3 is a partial quarter (July).
How many are building UI
We classified 278 sampled repos by their dependencies. About 23% build a user interface (React, Next, Vue, Svelte, Tailwind, or a static site), give or take 5 points. That slice, roughly one in four, is dwic’s audience: people shipping interfaces from a terminal, with no traditional design tooling.
- 76%CLAUDE.md: Claude Code, 639K files.
- 20%AGENTS.md: agent-agnostic, 164K files.
- 4%Cursor, Windsurf, Cline, Aider: other tools, 31K files.
What this means
A large, brand-new group is shipping interfaces in a workflow with no design tool in it at all. An agent is very good at the design you can see in a screenshot. It is blind to the design you cannot.
What shows in a screenshot
- Layout, spacing, and alignment
- Color and type
- The happy-path look of a screen
An agent optimizes for this.
What doesn’t
- Contrast ratios
- Keyboard and focus order
- Labels, roles, and landmarks
- Reduced-motion and edge states
This is where it fails.
Our first study measured the result. Of 123 frontends built by AI coding tools, 74% would fail a basic quality check, and accessibility was the number one defect. That is what this workflow ships when nothing minds the invisible half.
Where this points
This is a snapshot, not a forecast, and every number here is a floor. But three signals are hard to walk back.
- A vertical, two-year-old curve. Adoption went from almost nothing to hundreds of thousands in under two years. That is the shape of a beginning, not a fad leveling off.
- A quarter is already interfaces. The design surface is moving into the terminal whether the incumbent tools follow or not.
- A measured cost when ungoverned. Study one showed what this workflow ships without a guardrail: most of it would fail a basic quality check.
If the curve holds, the design tool of the next decade looks less like a canvas you open and more like an agent you instruct, with the guardrails built in. That is the future we are building dwic for.
Method, and what this can and cannot see
We counted public files on GitHub matching each agent-config convention (CLAUDE.md, AGENTS.md, .cursorrules, .cursor/rules, .windsurfrules, .clinerules, .aider.conf.yml) via code search, then sampled 278 repos and, for each, read the first commit that introduced the file (its real adoption date, not the repository’s creation date) and its dependencies (to classify frontend).
Deterministic queriesFirst-commit datesFiles, not people
Every number here is a floor. Repo fingerprints only appear once someone commits a config file. AI-assisted work that predates the convention (Copilot from 2021, ChatGPT from late 2022, Cursor autocomplete from 2023) leaves no trace, so the real movement is older and larger than we can show.
The file counts are a July 2026 snapshot that climbs week to week, so we report them as a floor. They index files on GitHub, not unique repos, and the sample is relevance-ranked, not random, so treat the magnitudes as directional. Check our work: the raw sample of 278 repos and the full method are on GitHub.