deep-research-plan

Recommended next research phase

This plan is based only on discovery. It does not define a future design system or maturity model.

Systems to crawl deeply

1. IBM Carbon

Why: Broadest public Level 1 evidence across guidance, components, tokens, Figma references, code, accessibility and an MCP server.

Crawl: Carbon for AI; AI label; every core component’s AI-presence section; code/accessibility tabs; AI Chat docs/Storybook; Carbon MCP docs; relevant news/roadmap/issues.

Repositories: carbon, carbon-ai-chat, carbon-mcp, carbon-website, and relevant Carbon Labs history.

Capture: Package APIs, component source, tests, accessibility status, explainability/revert behavior, AI token definitions, MCP tool/resource schemas, examples and agent instructions. Pull token packages. Record Figma asset names/access conditions; do not assume public file access.

Questions: What moved from experimental to stable? Which AI behavior is encoded in components versus docs? How are AI states represented in APIs and tokens? What does Carbon MCP expose, and how is correctness tested?

2. Microsoft Fluent and Microsoft agent guidance

Why: Strong cross-platform AI kit signal and the most developed public agent lifecycle/accessibility guidance.

Crawl: Fluent design resources and Copilot kit documentation; content engineering; Microsoft Design’s AI-first articles; full Agents design-guideline tree; Adaptive Cards; HAX library and underlying research.

Repositories: Fluent UI React/web-components and Adaptive Cards; locate any public Copilot component packages or examples before claiming code alignment.

Capture: Public tokens/components, agent response patterns, lifecycle states, accessibility requirements and content/system-prompt artifacts. Record kit access and licensing.

Questions: Which AI patterns are formal Fluent assets? How do agent lifecycle patterns map to reusable UI? Where does content engineering sit relative to the DS? Which implementation details remain internal?

3. Atlassian Design System

Why: Clearest formal claim that a mature DS is becoming a semantic “context engine,” while also integrating AI product patterns.

Crawl: Entire Rovo UI/Dialect and Rovo AI pattern trees; ADS context-engine article and linked talks; Panel/composition work; accessibility and token foundations.

Repositories: ADS packages and public mono-repositories; search specifically for MCP, skills, context files, Markdown instructions, generation templates and AI/Rovo assets.

Capture: Component/pattern inventory, design foundations, accessible public code, token packages and exact public/private boundary for ADS MCP and skills. Inspect Figma/library access.

Questions: What is implemented versus directional? What semantics are encoded? How is context versioned and evaluated? How do Rovo product patterns relate to the agent-facing context layer?

4. SAP Fiori and the Compositional Design System

Why: Richest formal AI pattern library and the most explicit model of semantics/relationships/guardrails for generative UI, with a stated production use.

Crawl: All current AI and Joule pages across web/iOS/mobile; archived intelligent-system guidance; Joule Work and Autonomous Enterprise design stories; explainability, feedback, ranking, matching and recommendation patterns.

Repositories: Public Fiori/UI5 controls, SDK examples and token/theme packages. Search for public compositional metadata or schemas without assuming their existence.

Capture: Pattern text/examples, controls referenced, platform differences, tokens/packages, public Figma resources, and every verifiable artifact behind the Compositional Design System claim.

Questions: Which guidance predates generative AI and which is new? What is the actual representation of relationship logic? How is generated UI constrained, tested and made accessible? Which details are proprietary?

5. Figma + Storybook infrastructure

Why: These are becoming common delivery surfaces through which many design systems become legible to agents.

Crawl: Figma MCP developer docs, Code Connect integration/mappings/instructions, MCP-agent responsibility boundary, code-to-canvas features; Storybook MCP architecture, manifests, documentation tools, story generation and tests.

Repositories/packages: figma/code-connect, Storybook MCP packages and schemas.

Capture: Schemas, supported frameworks, generated payloads, versioning, permissions, test hooks, limitations and security considerations.

Questions: What context is actually transmitted? What remains implicit? Can behavior and semantic intent be represented? How portable are mappings? What evidence shows improved DS adherence?

Bounded supporting crawls

• SGDS: map the AI section, skills, component/token/code links and Figma/Code Connect workflow; valuable as a transparent, smaller implementation with documented limitations.
• WordPress/Gutenberg: inspect the DS MCP package, generated data, Storybook ingestion, package releases and tests.
• zeroheight: preserve survey methodology/raw numbers if available; map MCP delivery and distinguish vendor product claims from field evidence.

Non-design-system sources to preserve

Research and guidelines

• Microsoft HAX 18 Guidelines and 2019 CHI paper.
• Google People + AI Guidebook and research on how practitioners use it.
• 2024 Design Principles for Generative AI Applications.
• 2026 “How Agents Ask for Permission” survey.
• 2026 workplace human–AI agent UX principles paper.
• Research on design-system-aware AI development, with careful review of method and sample size.

Production and safety evidence

• OpenAI ChatGPT Agent System Card sections on confirmations, watch mode and memory.
• OpenAI prompt-injection mitigation article, especially consent and constrained effects.
• Official MCP Apps specification/repository for interactive UIs delivered by tools.

Practitioner material

• Into Design Systems 2026 conference pages, talks, slides and any benchmark artifacts, especially Indeed’s reported prompt benchmark.
• Romina Kavcic’s 156-system inventory; seek dataset/method to verify the “26 systems” result.
• Smashing/ServiceNow agentic UX series.
• The selected Reddit threads in discussions.md, archived with dates and explicitly labeled as opinion.

Tools and projects

• Figma Code Connect, Storybook MCP, WordPress DS MCP and Carbon MCP.
• Independent DS MCP/registry projects only as experimental comparison cases, never as proof of widespread practice.

Proposed crawl method

1. Freeze URLs and capture dates/commit hashes.
2. Inventory navigation trees before extracting individual pages.
3. Capture guidance, source/API, tokens, tests, Figma access and issue/discussion history separately.
4. Trace every high-level claim to a repository, shipped product, formal guideline or clearly labeled unverified statement.
5. Build comparable evidence records without forcing a capability framework: artifact, owner, problem addressed, implementation status, user, and public/private boundary.
6. Preserve contradictions and historical renames (for example, Carbon “AI slug” to “AI label”).

Gaps discovery cannot answer

• Whether internal teams achieve materially better quality, speed or DS adoption with these approaches.
• A validated definition or threshold for “AI-ready.”
• A shared schema for component intent, relationships, anti-patterns or agent permissions.
• How design systems should govern generated compositions over time.
• Whether users understand current AI labels and explanations across contexts.
• Which agent lifecycle patterns belong in a cross-product DS versus a platform-specific orchestration layer.
• How to measure semantic correctness of generated UI, not merely visual similarity or valid code.
• How accessibility testing works for dynamic/generative UI before it reaches users.
• How public Figma assets, private documentation and proprietary product constraints are securely exposed to agents.
• Maintenance cost, context-window/token cost, latency, version skew and security risk of MCP-based delivery.
• Whether agent instructions and MCP tools remain reliable across different models and coding environments.
• How often AI-specific components are truly needed versus AI states on existing components.

Stop condition

The next phase should be a controlled crawl of the Tier 1 surfaces above. It should not yet convert findings into a future-state design, component list, token proposal or maturity model.