AI titans to tidy agent frontier with plugin prescription

A new packaging standard, Agent Plugins 1.0, lets engineers write agent skills and tools once and move them between platforms. The format combines the existing MCP protocol for external connectivity with the Agent Skills standard for task instructions. By bundling both in a predictable directory structure, the spec creates a portable unit that can be copied across any supporting agent client. This reduces the effort of reimplementing the same capability for each platform, but it does not eliminate platform-specific work, security, UI, and marketplace integration remain the responsibility of the client.

The spec is deliberately narrow, focusing only on the essential manifest, skills, and MCP configuration. Additional extensions can be added in a namespace directory, but portability is not guaranteed for these extras. This trade-off keeps the format simple and predictable while allowing experimentation. Engineers who adopt the standard gain immediate portability for core functionality but must accept that some custom features may not transfer. The open governance model means future versions could incorporate widely used extensions, but there is no commitment to do so.

Platform support is already in place for major tools like VS Code, Cursor, and GitHub Copilot, but the standard’s success depends on continued adoption. The spec is managed independently by the Agentic AI Foundation, with input from the founding companies. Competing approaches, such as Google’s Agentic Resource Discovery and AAIF’s Skills Over MCP, exist but are positioned as complementary. Engineers must decide whether to adopt this standard now or wait to see if a broader consensus emerges around a different approach.