---
title: "Vision"
url: /vision
description: "Where AI design systems are heading, and how one should work. Five shifts already under way, the target loop from decision to check, and my bets, marked as judgment."
---

## §0 · vision

# The design system becomes the contract an agent works under

This is where I think design systems go, and why. I mark each judgment as a judgment.

For fifteen years a design system was written for people: a library, a guide and a few experts who knew the rest. Agents now write a large part of the interface, and an agent has no tacit knowledge. It reads what is written and obeys what is enforced.

So the system has to change shape. It stops being documentation and becomes a contract: decisions with reasons, tokens as data, components as code, rules an agent can read and checks that prove the result. The parts already exist. Few teams have joined them.

## §1 · shifts

### Five shifts, already under way


**1. Rules move into the repository.** A short rules file next to the code tells an agent what goes wrong and what to do instead. The format is shared now: AGENTS.md and MCP moved to the Agentic AI Foundation under the Linux Foundation on 9 December 2025 [claim ind-aaif: Announcement by OpenAI, co-founded with Anthropic and Block; source: https://openai.com/index/agentic-ai-foundation/; confirmed]. It works: in one test of the Supernova tool, a 12-line rules file removed every rule break [claim ind-supernova-rules: One test of the Supernova tool, published by its maker, round two: 9 runs (3 tasks × 3 runs) per set-up; rules file: 8 rules in a 12-line CLAUDE.md; source: https://www.supernova.io/blog/shadcn-mcp-design-system; confirmed].

**2. Tokens become one source with contexts.** Brand, mode and density stop being copies of a theme. The DTCG Resolver Module 2025.10 combines token sets through modifiers such as light and dark [claim ind-dtcg-resolver: Final Community Group Report, 28 October 2025. Not a W3C Standard; source: https://www.w3.org/community/reports/design-tokens/CG-FINAL-resolver-20251028/; confirmed]. This site runs on one resolver with two modifiers, brand and mode.

**3. Components ship as registries an agent can search.** An agent installs owned code from a registry instead of guessing an import. The shadcn MCP server works with any registry that follows the shadcn registry specification [claim ind-shadcn-v4: shadcn/ui changelog and MCP documentation; source: https://ui.shadcn.com/docs/changelog ; https://ui.shadcn.com/docs/mcp; confirmed].

**4. Design files open in both directions.** Agents can write to the Figma canvas through Figma's MCP server, using the team's components and variables [claim ind-figma-write: Figma developer documentation. Remote server only; in beta and free during the beta; source: https://developers.figma.com/docs/figma-mcp-server/write-to-canvas; confirmed]. The design tool becomes one more client of the system, not its only home.

**5. Design tools move into the conversation.** Claude Design is moving into Claude: a research preview in April 2026, inside every Claude conversation since September, the standalone site closing on 14 December [claim ind-claude-design-dates: Anthropic release notes, product page, help centre; source: https://www.anthropic.com/news/claude-design-anthropic-labs ; https://support.claude.com/en/articles/12138966-release-notes ; https://support.claude.com/en/articles/17440474-move-from-standalone-claude-design-to-claude; confirmed]. Its own working rule is the one this site follows: code is the truth, screenshots are lossy.

## §2 · system

### How it should work

One loop, from a decision to a check and back. Every step is a file in the repository, so a person and an agent read the same facts.


1. **Decisions.** Numbered, each with its reason, the option rejected and the check that enforces it. Where nothing checks a rule, the register says so.
2. **Tokens.** DTCG files and one resolver. A build turns them into code; nobody edits the output by hand.
3. **Components.** Owned code that reads tokens only, published as a registry.
4. **Rules for agents.** A short `AGENTS.md` that names what goes wrong, with the correct replacement. Skills apply the rules; they never make them.
5. **Agents.** They read all of the above through MCP. They select and fill. They do not invent a colour, a spacing value or a component.
6. **Checks.** Deterministic scripts in CI judge the result: the same input gives the same verdict, byte for byte. A failed check names the rule it protects.
7. **Proposals.** When an agent meets a gap, it writes a proposal instead of a new rule. A person approves it, and it becomes a numbered decision. The loop starts again.

This is the ladder from level 3 to level 5. [Find your level](/ladder)

## §3 · bets

### What I expect

These are bets, not facts. I will check them against this page in a year.

- **Judgment:** the main output of a design-system team becomes the decision register and its checks. The component library stays, but it is no longer the product.
- **Judgment:** open files win the format question. Tokens in DTCG, rules in `AGENTS.md`, components in a registry. The lock-in moves to the checks a vendor runs for you.
- **Judgment:** design tools and coding agents meet on the same protocol. A designer's change and an agent's change pass the same checks.
- **Judgment:** self-maintaining systems stay mostly demo through 2027. Agents propose; people approve.

## §4 · gaps

### What is still missing

The industry has the parts, not the trust.


 68% of developers use AI to write code, but only 32% of designers and developers trust the output [claim ind-figma-trust: Figma's own AI report; sample size not stated on the page. Vendor: Figma; source: https://www.figma.com/blog/design-systems-ai-mcp/; confirmed]. In another survey, 59% of practitioners report UI that bypasses the design system [claim ind-zeroheight-bypass: Vendor survey of 123 practitioners, 2026. Vendor: zeroheight; source: https://zeroheight.com/resources/state-of-ai-in-design-systems/; confirmed].

No tool I reviewed stores a team's usage rules with an id, a severity and a link to the check that enforces it. That layer is open. This site is one working version of it, with its own gaps in public.

[All decisions](/decisions) · [Today against demo](/today-against-demo)
