This kit is the workflow from the talk "Built for Two: What Design Systems Are Missing When AI Agents Become the Third User", AGNTCon + MCPCon Japan 2026, September 11, Tokyo. Everything here is deliberately small: one test, one map in two halves, one set of tiers, one doc template.
What's in the kit
| File | What it is | Time to adopt |
|---|---|---|
| workflow.md | The workflow the other files support. Five steps here; grouped into the three moves used on stage, mapped at the top of the file | A read |
| two-minute-test.md | The test: what does an agent get wrong about your best-documented component? | 2 minutes |
| alert.md | The document from the talk, the one that looks finished. Run the test on it and see what it can't answer | 2 minutes |
| component-index.template.md | Half one of the map: one file, every component, one line each | An afternoon |
| structure-map.template.md | Half two of the map: where things live, and how a workflow shows up on screen | An afternoon |
| tiers.md | Core / Extended / Experimental: status criteria and promotion rules | A team discussion |
| component-doc.template.md | Component doc template with multi-author review built in | Per component |
Where to start
Run the two-minute test. Don't build anything first. What the agent gets wrong is your starting point, not for the agent, for your team.
The idea in one paragraph
The agent didn't bring new problems into your design system. Every failure it surfaces, naming collisions, gaps the docs never covered, no map of what's load-bearing, was already there. Your team had learned to navigate around them with tribal knowledge; the agent can't. Fixing the docs for the reader with no tribal knowledge fixes them for every human who joins after the current team walks out.
The same shape carries beyond design systems: an MCP server whose tool descriptions sit in one author's head, a skill whose instructions assume context that never made it into the file, a README an agent has to trust. Same failures. Same fix.