Google Agent Registry skills reach GA with scanning controls
Google Cloud made standalone skills in Agent Registry generally available on November 6, giving teams a central place to manage reusable instructions and supporting files for their AI agents. The same release notes announce skill security scanning and governance controls. Standalone skills had been in Preview since July, so this is a maturity milestone rather than a new agent format.
The practical appeal is change control. A team distributing the same procedure across several agents needs to know which version each one consumes, who can update it and what happens when a package needs review.
Reusable packages with fixed revisions
Google’s skills architecture separates a stable Skill resource from immutable SkillRevision packages. A revision holds the instructions, optional scripts and other assets. Consumers can request a particular revision or use the skill’s default revision pointer.
That arrangement lets teams move the default forward or restore an earlier version without rewriting a published package. It supports repeatable configuration, although the same instructions still do not guarantee identical model responses. These standalone packages are distinct from A2A skill descriptions in an agent’s Agent Card, which describe capabilities other agents can invoke.
What scanning catches, and what it can miss
The default scanning policy is Enabled Blocking. Google’s scanning guide says uploaded revisions are inspected for risks including prompt injection, exposed credentials and unauthorized data exfiltration. In this mode, revisions with findings cannot be downloaded, while findings or pending scans prevent promotion to the default revision.
A nonblocking audit mode is available, and organizations can enforce scanning policies centrally. Changing a skill’s scan settings does not inspect older revisions retroactively. Teams must upload a new revision to scan previously uploaded material.
Google explicitly warns that the scanner uses probabilistic LLM analysis. It can miss issues, produce false positives and hallucinate details in findings. Developers must verify its results before making fixes. A clean package scan therefore needs to sit alongside runtime permissions, testing and human review.
Package limits and first setup
The registration workflow requires an enabled Agent Registry project and suitable IAM permissions. A ZIP package must contain a root SKILL.md file with valid YAML metadata. The documented console flow starts a custom skill in a disabled state, with default-revision selection and activation required before discovery.
Google’s published limits cap compressed skill archives at 1 MB. Scanned packages may contain up to 10 MB uncompressed, with a 1 MB limit per file. Default quotas allow 100 standalone skills and 500 revisions shared across skills in each project and location.
Free registry operations, paid scanning from January
The pricing schedule keeps registration, storage, search and registry operations free. Skill scanning is free through December 31, 2026, then costs $8 per million uploaded input tokens from January 1, 2027. Internal scanner prompts, reasoning and output tokens are excluded. Google-managed skills do not incur scanning charges in the customer’s project. Separate Cloud Storage and agent-runtime charges may apply.
For an initial deployment, the useful test is whether the registry improves a real release process: publish one small skill, review its findings, pin a revision and test how updates reach consumers. ByteForward has not independently tested the service.
The release extends Google’s broader agent-management tooling. Our App Topology coverage explains a complementary view of deployed agents and their dependencies.
Featured image: AI-generated editorial illustration of reusable agent skills. Google and the Google logo are trademarks of Google LLC. ByteForward is not affiliated with or endorsed by Google.
