Agentic Governance
Agentic governance controls what an AI agent can reach and whom it acts as. It uses the same manifests, pull requests and engine as every other resource.
The agentic layer is early and in active development. The agent runtime supports ADK only. Test an agent again in each environment before you use it in production. Governance covers the shipped Agent, AgentDeploymentConfig and McpConfig kinds. Broader policy controls are not complete. Treat the examples as reference patterns, not as guarantees.
Three controls
| Control | What it does | Source |
|---|---|---|
| Access boundary | Lists the databases, secrets and Pub/Sub topics that the agent can reach. | The accessControl block of the Agent manifest. A change is a pull request. |
| Managed identity | Gives the agent its own scoped identity and refreshes its tokens. | The engine, for every Agent. No manifest field exists. |
| User consent | Lets an agent act as one person, with the consent of that person. | The code of the agent, through the ADK credential flow. The manifests do not enforce it. |
Agent manifest kinds
Agent declares an AI agent that runs on Vertex AI Agent Engine.
apiVersion: lowops.manifests.v1
kind: Agent
metadata:
name: compliance-auditor
application-set: compliance-agents
release-track: main-track
organizational-unit: platform-engineering
organization: example-co
spec:
description: "SOC2 compliance auditor — reads infrastructure state, generates findings"
target: AGENT_ENGINE
project: platform-core
buildDefinition: compliance-agentic-build # BuildDefinition (type: AGENT) that packages this agent
accessControl:
database:
secretSource:
envVar: DATABASE_PLATFORM_CORE_CONNECTION_STRING
schema: platform-core
name: main
privileges: [USAGE] # Read-only — principle of least privilege
secrets:
audit-api-key:
envVar: AUDIT_API_KEY
AgentDeploymentConfigpins anAgentto a version in each environment.McpConfigdeclares an MCP server that runs as a Cloud Run service and gives tools to agents.- See the
Agentreference for the full schema.
The engine reads only the registry field of the agent block, to grant CI permission to push artifacts. The schema requires entryModule, packageName and agentDirectory, but no consumer reads them. This type has no CI workflow builder yet. Build and publish the agent wheel by hand.
Authentication, Managed by Default
Agent identity
Each agent gets its own scoped identity. It does not share a broad CI/CD service account. On deploy, the engine does these actions:
- It grants the IAM roles that the agent needs to reach its registered tools (
agentregistry.viewer) and to call out through the network boundary (iap.egressor). - It issues audience-scoped ID tokens and refreshes them before they expire.
You do not write a manifest field for this. The engine does it for every Agent.
End-user authorization (3-legged)
Some actions must be attributable to one person. Examples are a pull request that a named developer opens, or a read of a private repository. For these, the agent asks the person for consent through the ADK credential-request flow. The agent code must call this flow. The engine has no consent code and no manifest field for it.
Example: an agent that sets up a new service opens a pull request as the developer who asked for it. The pull request, CODEOWNERS review and audit history then show the real requester, not a shared bot account.
What agents cannot do
- Create IAM bindings outside their declared project.
- Change egress rules. The
Projectmanifest holds them, outside the reach of the agent. - Reach secrets or databases that
accessControldoes not list.
A change reaches production through the release track, where a stage can require approval.
Next
Agentreference andMcpConfigreference- Use Beck
- The Agent Path: what your own AI agent will be able to do through MCP, for each tier and environment
- Agent access model: an agent never has more access than its person, and production stays with people
- The Security Framework