Agent Plugins standard sidesteps governance—ceding control of core behavior to Anthropic specifications

Vercel, AWS, Cursor, Microsoft, OpenAI, and Google this month released Agent Plugins 1.0.0, a shared standard for packaging AI skills and tools that developers can reuse across platforms without forking. Yet the specification performs a deliberate act of restraint: it governs only where components live in a plugin directory, not what they do—handing that authority to separate specifications created by Anthropic, which sits outside the project's steering committee.

The governance split that reshapes portability#

When Anthropic published the Model Context Protocol (MCP) and Agent Skills specifications, it donated them to open governance separately. MCP went to the Agentic AI Foundation under the Linux Foundation; Agent Skills became an openly developed standard. The Agent Plugins Technical Steering Committee—made up of individuals from the five founding companies plus Google—has no control over either.

This split is not accidental. The project's own documentation describes v1 as "a package format and nothing more," not a full interoperability contract. The specification explicitly delegates Section 7.1 (skill behavior) to Agent Skills and Section 7.2 (MCP wire protocol) to separate governance. That narrowness is the core of every fragmentation problem that follows.

Google's representation on the steering committee is still in progress—Kevin Hou is listed as joining, but the official maintainers file has not yet been updated. Anthropic has no seat on the committee, and no primary source explains why.

What the spec actually standardizes#

Agent Plugins does one job cleanly: it locks down file locations. A plugin is a directory with plugin.json at the root, skills in skills/, and MCP servers declared in mcp.json. The manifest itself requires only two fields—$schema and name—with everything else optional.

By fixing these addresses, the standard eliminates the redundant forking problem it was built to solve. Developers write a skill once, reference it in the manifest, and multiple clients can find it in the same place. No per-client file rewrites needed.

But behavior? That belongs elsewhere. The specification points to Agent Skills and MCP rather than constraining them. It avoids defining profiles, mandating subsets of optional features, or prohibiting extensions. This is a deliberate choice: treat Agent Skills and MCP as separately governed contracts, standardize addresses only.