Google tightens oversight of its AI Agent Skills repository

Google has outlined how it builds, tests and governs its open-source Agent Skills repository, which has drawn more than 15,000 stars on GitHub.

The repository packages Google Cloud domain knowledge into structured instructions for AI coding agents. Its goal is to make those agents more accurate and reduce errors in development tasks.

The project began as a rapid internal effort ahead of Google Cloud Next 2026. A cross-functional group of Developer Advocates and Technical Writers created the initial framework to turn product knowledge into agent-readable guidance.

Interest soon spread beyond the original launch group. Engineering and product teams across Google wanted to add skills for their own services, including work outside the core cloud business.

Quality controls

That growth created a governance problem. A skill with weak instructions, dead links or incomplete handling of edge cases can undermine the wider repository and lead to poorer results from AI agents.

To manage that risk, Google standardised the repository structure for each skill and introduced automated checks before any contribution can be accepted. Those checks cover metadata validation, directory layout, naming rules and line counts.

It also tests every URL to catch broken or invented links before a change is merged. Automated checklist reviews assess whether instructions follow required structural patterns and internal guardrails.

Testing regime

Each new skill must go through internal evaluation before public release. Authors must provide prompt suites and scoring rubrics so Google can test whether the skill improves an agent's output.

Google also runs scheduled weekly checks across the full skills library to detect regressions caused by changes in documentation, APIs, large language models or agent frameworks.

For each evaluation suite, it compares agent performance with and without a given skill. It measures both accuracy, including task completion and response quality, and efficiency, including token use and completion time.

Those tests are repeated across different agent frameworks to determine whether a skill delivers a measurable improvement. Google described this as a way to stop the library degrading as more teams contribute material.

Internal and public use

Google develops and evaluates skills internally before exporting selected work to the public GitHub repository. Automated export rules strip out internal assets, ownership details and evaluation suites before publication.

It has also created a separate internal programme called DevRel Skills. The initiative focuses on agent skills for internal workflows such as content transformation, search engine optimisation and reporting.

Google said the internal project is intended to improve consistency in day-to-day work by encoding recurring processes into reusable instructions. This suggests the company sees the same methods used for external developer tools as applicable to internal knowledge work.

Design approach

At the architectural level, Google prefers to reference remote Model Context Protocol tools where possible rather than rely on command-line interfaces or direct API calls. It said remote MCP servers are better suited to agent-based workloads because they provide tools while also handling authentication and identity governance.

The structure also reflects a broader effort across the AI software market to impose discipline on prompt-based systems. As companies build more agent tools on top of large language models, the challenge has shifted from simple experimentation to maintenance, testing and version control.

Google's approach places responsibility on both central maintainers and individual authors. Repository maintainers are responsible for overall health, pipelines and standards, while skill owners are expected to update their work when APIs change or quality problems emerge.

The company has also built internal tools to support authors creating new skills and evaluation suites. These include skills that help contributors write other skills, along with agent-based tools built with the Agent Development Kit for drafting and self-critique.

The combination of structured templates, automated checks and recurring evaluations shows how large technology groups are trying to turn AI instructions into managed software assets rather than ad hoc prompt libraries. Google said a key lesson from the project is that a skill should be treated as a living product rather than a one-off document.