Agent Plugins: Google Just Made Your AI Skills Portable

Google has joined Amazon, Microsoft, OpenAI, Cursor, and Vercel as a core maintainer of Agent Plugins 1.0.0 — an open, vendor-neutral standard for bundling AI agent skills and tools into portable packages. Two Google products, Agents CLI and Data Agent Kit, already support it. If you're building with AI coding tools, this is worth understanding now, before it quietly becomes infrastructure you depend on.

The Problem It Solves (And Why Designers Should Care)

Here's the situation before this standard existed: say you built a small AI skill — a reusable set of instructions that pulls data from a spreadsheet and writes a weekly summary. You also connected it to a tool (technically, an MCP server — think of that as a bridge between your AI agent and an external service like a database or API). Together they do one useful thing.

Then you want to use that same combo in a second AI environment — say, switching from Cursor to Gemini CLI. Suddenly you're copying files, reformatting config, and maintaining two nearly-identical versions that slowly drift apart. The skill and the tool weren't the problem. The wrapper — the box you shipped them in — was different every time.

Agent Plugins standardizes that box. One directory structure, one manifest file, readable by any client that adopts the spec. You build it once; it travels.

For designers building their first AI-powered tools, this matters because it's one less invisible wall between "I made this work in one place" and "I can ship this somewhere real."

What an Agent Plugin Actually Looks Like

No code experience required to grasp this. An Agent Plugin is literally just a folder with a predictable layout:

• A plugin.json file at the top — just a name and a schema reference, two lines of substance
• A skills/ folder containing your reusable AI instructions
• An mcp.json file listing any external tools or services your plugin connects to
• An optional namespace folder (like com.example.cursor/) where a specific client can stash its own extras — and every other client simply ignores it

That last part is elegant design thinking: the portable core stays small because the non-portable parts have a legitimate place to live without polluting everything else. Clients that don't recognize an extension just skip it. And if one tool fails to load, it doesn't take the whole plugin down — components fail independently.

The restraint is the point. Agent Plugins deliberately does not define how plugins get installed, distributed, discovered, or permissioned. Those things are genuinely different depending on whether you're working in an IDE, a command-line tool, or an enterprise platform. The spec names those gaps openly in its future considerations rather than pretending they don't exist.

How to Actually Use This as a Designer-Builder

Right now, the two places to experiment are both from Google.

Agents CLI packages a set of expert skills for agent building, evaluation, and deployment — and it works across AI coding environments including Gemini CLI, Claude Code, and Cursor. If you're already using one of those tools to vibe-code your way through a project, Agents CLI can slot in as a skill set without you needing to rewire anything.

Data Agent Kit is more specialized — it's aimed at connecting agents to Google's data services like BigQuery and Cloud SQL. If your project involves pulling in real data (user research exports, analytics, a content database), it's worth knowing this exists and that it uses the same portable format.

Beyond Google's tools, the practical move right now is to think in terms of composability. When you're building a skill or connecting a tool in your AI coding environment, ask: is this something I'd want to reuse? If you're assembling something that genuinely belongs together — a set of instructions plus a service connection — that's when the plugin format earns its keep. A single MCP server connecting to one service? You probably don't need to wrap it in a plugin yet.

The Ecosystem Around It

Agent Plugins is one layer in a stack that's taking shape. There's also Agentic Resource Discovery (ARD), an open protocol for helping a client find available skills and tools for a given task — think of it as search, but for AI capabilities. And there's AI Catalog, a format for indexing those resources so they're findable. Each layer is independently adoptable; you don't have to buy into all of them to use any one of them.

This layered thinking — packaging separate from discovery, separate from execution — is actually good product architecture, and it's worth internalizing as a mental model even if you're not implementing any of it directly.

A Grounded Takeaway

Agent Plugins 1.0.0 is a packaging format, not a finished ecosystem. Installation, trust, permissions, and the user experience of managing plugins are all explicitly unresolved. The spec is new, Google's support is brand new, and real-world adoption across tools will take time to develop.

But the direction is clear, and the coalition behind it — spanning most of the major AI coding environments — is unusually broad for something this early. If you're building skills and tools as part of your design-to-build practice, keeping them portable from the start costs you almost nothing. Ignoring portability until you need it costs you a painful rewrite.

Build the box right the first time. The spec now tells you how.