Registering an agent turns it into a discoverable component within Agent Registry. By publishing your agent's A2A skills, you help other developers and orchestrator agents in your organization find and reuse your work, accelerating overall AI development.
Agent Registry seamlessly integrates with the
Agent2Agent (A2A) protocol and its
specifications. A2A is an open standard that lets agents declare their
capabilities and identity. Agent Registry supports versions 0.3 and 1.0 of
the A2A specification. For more information, see
Agent Card schema.
Depending on your agent runtimes, Agent Registry supports two primary registration mechanisms: automatic registration and manual registration.
This document helps you choose the registration mechanism that fits your agent runtimes.
Automatic registration
Supported Google Cloud runtimes register your agents with Agent Registry either by default or through an opt-in deployment step.
Automatic registration is supported for the following Google Cloud runtimes:
- Agent Runtime on Gemini Enterprise Agent Platform: Agents deployed using the SDK are registered without additional configuration.
- Built-in Google agents: Agents that Google provides directly, such as Google Workspace and Gemini Enterprise agents, are discoverable without any configuration.
- Google Kubernetes Engine: Containerized agents deployed to Google Kubernetes Engine (GKE) are registered when they include a specific Agent Registry functional type label.
The metadata and A2A skills of an agent captured during its registration depend on the agent's protocol:
- Agents implementing the Agent2Agent (A2A) protocol: A2A is an open standard that lets agents declare their capabilities and identity. If your agent implements this protocol, Agent Registry automatically extracts the agent's A2A skills from its Agent Card and populates the registry.
- Agents not implementing the A2A protocol: Agent Registry registers the agent, but cannot automatically capture metadata such as the inference URL and A2A skills. To provide this information so your agent is discoverable, use manual registration.
For information about Google Cloud runtimes that support automatic registration, see Use automatic registration.
Manual registration
For agents hosted on external platforms, on-premises, or on Google Cloud runtimes that don't support automatic discovery, you must use manual registration.
To manually register an agent, you use the Agent Registry API, Google Cloud CLI, or
Terraform to create a Service resource that defines the agent's endpoint and
configuration. This writable Service resource also contains the agent's
specific details. After the Service is created, the registry generates a
read-only Agent resource that you can use for discovery.
You can manually register two types of agents:
- A2A-compliant agents: Agents that implement the A2A specification
(
A2A_AGENT_CARD). The registry scans their Agent Card (agent-card.json) endpoint to index their A2A skills. - Standard REST agents: Remote agents that don't implement the A2A
specification (
NO_SPEC), such as a standard REST API endpoint.
To learn how to manually register agents, see Use manual registration.
Integrate standalone skills
Agent Registry lets you manage and govern standalone
skills for agents as top-level resources.
Standalone skills are registered as Skill resources in your registry,
letting administrators apply fine-grained access policies and manage version
history using skill revisions.
When building and managing agents in Gemini Enterprise, you can reference these registered skills to extend your agent's capabilities. For information on how to register and manage standalone skills, see Register skills and Manage skills.