Anthropic’s published docs on Agent Skills + Plugins are the most careful thinking I’ve seen about the shape of agent-portable tooling. The key insights (paraphrased from memory, not a direct quote):
- Skills are named, versioned, and discoverable. An agent asks “what can I do here?” and gets a list of skills - not a raw list of functions. That distinction unlocks everything downstream.
- Skills carry intent, not just interface. A skill’s description tells the agent when to use it as much as how to call it. That’s what makes an agent able to choose well.
- Plugins are composable. A plugin bundles skills; a workspace composes plugins. An agent inherits capabilities without the skills themselves needing to know about each other.
This is the right shape for the next 18 months of agent tooling. It’s
also very close to the shape I’ve been building with
rice-shared-skills from the bottom up - one skill source
feeding multiple agent runtimes. Anthropic’s take is more formal, and
I’ll likely converge my approach toward the spec over the next couple
of iterations.
Worth reading twice if you’re building CLI agents or MCP servers.