Your Design System Just Became Executable

This week crystallized something that’s been building for six months: AI agents now operate inside the artifacts your team already produces — not alongside them. In late March, Figma opened its canvas to agents through a new tool that lets Claude Code, Cursor, Codex, and other MCP clients write directly to Figma files, building components wired to your actual design system rather than generating decorative mockups. Days later, Microsoft shipped Agent Framework 1.0 — a consolidation of Semantic Kernel and AutoGen with MCP support built in — joining OpenAI’s Agents SDK, Google’s ADK, and Anthropic’s Agent SDK as the fourth production-grade agent stack from a major lab. Four SDKs, one protocol, and suddenly your tools are the surface the agents act on.

The more interesting signal, though, is what travels between those surfaces: skills. Anthropic released Agent Skills as an open spec last December; OpenAI, Microsoft, and Spring AI have all adopted it. A skill is a markdown folder — a SKILL.md and a few supporting files — that teaches an agent a domain, a convention, or a taste. Figma’s launch ships with skills as a first-class way to encode design conventions on the canvas. Claude Code uses them. Cursor supports them. The copy-pasted prompt era is ending.

What actually changed

If you’re a product or design leader, the practical shift is this: your conventions just became executable artifacts. Before this quarter, your component naming rules, your PRD template, your accessibility checklist, your tone guide — they lived in Notion pages that agents couldn’t reliably consume. Formatted as a skill, the same documents become deterministic inputs to every AI tool your team uses. The skill you write once runs in every session, for every designer, on every file.

That’s a governance moment disguised as a tooling release. The teams that win the next twelve months won’t be the ones with the best model access — that’s commoditized. They’ll be the ones whose taste is written down in a form an agent can follow.

The practical implications

Three things worth doing this month:

Audit your tribal knowledge. Every “the way we do it here” rule that lives only in someone’s head is a blocker to agent leverage. Convert the top ten into skills.

Treat your design system as runtime, not documentation. If your tokens, components, and variables aren’t wired into your Figma library cleanly, agents will paper over the gaps by inventing their own. The cost of sloppy design system hygiene just went up an order of magnitude — because now the sloppiness ships at machine speed.

Pick a stack intentionally, but don’t overthink it. With four major agent SDKs and MCP as connective tissue, vendor lock-in is lower than it looks. The real switching cost lives in your skills library. Write skills that are portable — plain markdown, minimal tool-specific syntax.

The takeaway

“AI for product teams” spent 2024 meaning autocomplete. In 2025 it meant copilots. This week it started to mean something much more load-bearing: agents that operate on your files, guided by your written conventions, inside the tools you already pay for.

The work isn’t picking a tool anymore. It’s deciding what your team actually believes — and writing it down well enough that a machine can follow it.