`before_tool_callback` can skip one tool by returning a dict. The agent callback is the admission check. Skipping one tool does not decide the rest. An allow list is a later check on the subject.

## After the subject is known

| Check | Where |
| --- | --- |
| Bearer accepted | The hook on [Google ADK](/docs/agent-frameworks/adk) |
| Allowed tools | Trust-store row, when your process enforces it |
| The action itself | [Authorization](/docs/concepts/authorization) using the subject |

Revoke in [Agents](/docs/console/agents) stops new tokens when the subject is the OIDC client id. Tokens already issued last until `expires_in` unless a later check rejects them.

When the token is an [RFC 8693 delegation](/guides/agent-delegation), the person stays the subject and the agent stays the actor. Authorizing the agent as the user skips the delegation.

## Related

| Read | To learn how to |
| --- | --- |
| [Custom channel auth for Google ADK](/docs/ai/adk/channel-auth) | Admit the caller |
| [Enriching instructions for Google ADK](/docs/ai/adk/enriching-instructions) | What the model is allowed to see |
| [Authorization](/docs/concepts/authorization) | Pick the check after the subject |