The Pulse

Google engineer joins shared stewardship of Agent Skills and MCP

Google says its engineers are co-maintaining Agent Skills after Anthropic contributed the specification to the Agentic AI Foundation. Google Cloud’s Kurtis Van Gent outlined work on packaging, discovery and versioning for the standard.

AI.info Team · · fact-checked · against 2 sources ↓ · Published 12 Nov 2026

Google engineers are co-maintaining the Agent Skills specification with contributors from Anthropic, OpenAI and the wider community after Anthropic donated it to the Agentic AI Foundation, Google said in a November 11 blog post. The format joins the Model Context Protocol, or MCP, under the Linux Foundation’s agent-focused foundation.

Kurtis Van Gent, a Google Cloud engineer and core maintainer of both standards, described how the company adopted MCP for database tools and outlined priorities for Agent Skills. Those priorities include work on how collections of skills are packaged and discovered.

A shared home for agent standards

The foundation announced Anthropic’s contribution of Agent Skills on October 22. Its other additions were OpenAI’s Open Responses, Hugging Face’s OpenEnv and Alibaba’s OpenSandbox. The additions brought the foundation to 10 hosted projects in its first year, up from three launch projects in December 2025.

MCP connects agents to live tools, while Agent Skills supplies procedural instructions and scripts that can be loaded when needed. Van Gent said the foundation gives organizations and open-source communities using Agent Skills a way to participate in shaping the specification.

Tools and instructions

Van Gent described a problem that can arise as agents connect to many tool servers: deciding which tool descriptions and operating instructions to place in the agent’s context can become a bottleneck. Agent Skills uses a method called progressive disclosure, loading a name and description at startup and pulling fuller instructions, reference files or scripts into context when a task requires them.

The maintainer said skills package domain knowledge about how to approach a task, while MCP supplies the interface to execute it. He pointed to proposals for serving skills over MCP so platform maintainers can distribute instructions and live tools together.

From databases to the protocol

Van Gent said his team faced an integration problem as it worked to connect agents and databases: Google Cloud supports more than 15 databases, while new open-source agent frameworks were also emerging. The team created GenAI Toolbox for Databases, an open-source framework for building and serving tools, with its own server, clients and protocol.

After Anthropic announced MCP, the team contacted its co-creators and Van Gent advocated adopting it. The project evolved into MCP Toolbox for Databases, and he contributed capabilities from the original protocol to the upstream specification.

His work also included leading the MCP Transports Working Group, which sought a stateless, horizontally scalable architecture. He said early assumptions about persistent sessions and initialization were limiting MCP at production scale. A two-day summit brought core maintainers to Google’s San Francisco offices to work on scaling; the resulting July 28, 2026 specification moved MCP to a stateless model.

Database plugins and next steps

For database workflows, Google Cloud’s database team combines MCP and Agent Skills in Agent Plugins. Van Gent described developers using Antigravity with a PostgreSQL database on Cloud SQL or AlloyDB installing the corresponding plugin. The bundled skills provide task instructions for work such as provisioning an instance or configuring database authentication.

Looking ahead, Van Gent said packaging and discovery are priorities. The specification defines the structure of an individual skill but does not yet standardize how collections are bundled, how agent harnesses discover them or how naming conflicts are resolved. Versioning is another priority: he wants a portable way for skill authors and harnesses to declare which version of the specification they support.