How Google Releases Agent Skills Like Products: A Behind-the-Scenes Look Two Months Later

At a glance

Google manages the "skills" used by AI agents just like regular products. A skill is a set of instructions that tells an AI what to do, so even a small change can alter how the AI behaves. For this reason, every change goes through automatic checks and evaluation before it is deployed.

Key points

1. 1

   Each time a skill is changed, its version number must be raised, or the change is automatically blocked.

2. 2

   Change proposals must first pass six automatic checks covering structure, syntax, version, evaluation materials, and links.

3. 3

   Evaluation materials include a report comparing how well the AI performs with and without the skill.

4. 4

   Final approval is given by a human subject-matter expert, and an AI review agent also examines the skill's quality.

Why it matters

Because skills are shared across many teams and AI platforms, a system is needed to keep small mistakes from spreading to many users.