How components actually move from idea to production in 2026 — the four-stage spine (RFC, design review, tech review, ship), where AI agents enter each stage as assistance, and the surfaces reviews now happen across (Figma frames, MCP context, PR diffs, component stories).
Workflows used to be a calendar of meetings — sync RFC review, sync design crit, sync tech review, sync ship. The 2026 default is artifact-driven: RFC is a doc, design review is annotations on a Figma frame, tech review is comments on a PR, ship is a merge. Synchronous time goes to disagreement, not status.
Design crit Mondays, eng review Thursdays, ship meeting Friday. The meeting was the artifact.
Async RFCs in Notion or Confluence. Comments instead of meetings. Status pages and project boards.
RFC, Figma frame, PR, codemod, story — every stage has a permalinked artifact. Agents grade the artifacts; humans review the agent.
The spine doesn't change. What changes is what each stage produces — and the artifact that survives after the meeting (if one happens) is over.
A short written proposal — problem, motivation, the shape of a solution, alternatives considered, and what success looks like. Lives in a doc, not a meeting.
The proposed visual artifact gets checked against the system: token contract, density, motion vocabulary, accessibility floor. Happens against the Figma frame and the MCP context, not on a Zoom call.
Engineering feasibility, API surface, performance impact, and the codemod plan. The codemod is now the artifact, not an afterthought — if a change can't be codemoded, the change has to be reshaped.
Final sign-off, PR opened, codemod published, changelog entry written, consumer teams notified. Adoption tracking begins on day one.
Agents are assistance, not authorship. They draft, surface, run, and post — but the human still owns the artifact, the decision, and the consumer-team conversation.
There is no single review meeting. There is a stack of surfaces, each holding part of the artifact. Reviewers move across the stack; the decision falls out when the surfaces agree.
The visual proposal
Source of intent. Variables and component graph readable via the Dev Mode MCP server.
Machine-readable handoff
Figma Dev Mode MCP + the system's own MCP. The agent grounds its work in both.
Canonical review locus
Where consequence-conversations happen now. Comments, codemod results, CI checks all converge here.
The living artifact
Storybook or Ladle. The same surface designers spot-check and engineers wire up.
A new workflow generation comes with a new set of failure modes. These are the ones that bite design system teams in the first year.
A proposal that doesn't name the codemod is a wish list. The codemod plan is part of the RFC, not a follow-up.
Synchronous reviews don't scale and don't leave a trail. Reviews happen against the PR and the Figma frame — both have permalinks.
Agents write codemods, draft RFCs, propose API shapes. Humans still own the decision and the artifact. The agent's output is the first draft of one section, not the spec.
Major-version cuts, breaking renames, sunset announcements need a written cooling period — 48 hours minimum between the proposal landing and the merge.
Single-reviewer approval lets institutional knowledge calcify in one person's head. Pair reviews are cheap insurance.
Five concrete signals that the workflow is alive, not just documented. If a quarter goes by and none of these are true, the workflow exists on paper only.
Figma Dev Mode MCP, agent integrations, modern token pipelines — the stack the workflow assumes.
Who owns what — centralized vs federated vs open models, RFC discipline, funding.
What ships after the workflow runs — npm, registry, codemods, the adoption funnel.
