---
title: How do you govern agents deployed on Vertex AI Agent Engine?
url: https://opencontroller.lyzr.com/blog/govern-google-vertex-ai-agent-engine-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern agents deployed on Vertex AI Agent Engine?

**Give each agent its own Agent Identity instead of the shared default service account, grant it narrow IAM roles, restrict who can query it, put it inside a VPC Service Controls perimeter, and turn on Data Access audit logs and tracing. Google has renamed Vertex AI Agent Engine to Agent Runtime on the Gemini Enterprise Agent Platform; the controls below use current documentation. They cover Google Cloud agents only.**

Agent Engine made it simple to deploy an agent with one call. The defaults are built for speed, not least privilege. The most important governance step is changing the identity an agent runs as.

## Why is the default identity a risk?

By default, a deployed agent runs as the AI Platform Reasoning Engine Service Agent, or as a custom service account you set. Google's [access docs](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agent-access) state the consequence plainly: every deployed agent can reach whatever its service account can reach. If ten agents share one service account, each has the permissions all ten need.

Agent Identity fixes this. It is SPIFFE-based and gives each agent its own IAM principal, tied to its reasoning engine resource. Certificates last 24 hours, access tokens are bound to them, the identity cannot be impersonated, and it allows no service-account keys. You opt in at deployment with `identity_type=AGENT_IDENTITY`. Google's docs warn against granting the agent `roles/aiplatform.user`; grant only what its tools need.

Two details catch teams out. A redeployed agent gets a new principal, so bindings must be re-applied. And IAM bindings persist after an agent is deleted, so clean them up. Google labels Agent Identity preview on one docs page and generally available in a July 2026 blog post, so confirm the status for your project.

## Which controls map to which risks?

| Risk | Control | Note |
|---|---|---|
| Over-broad agent permissions | Agent Identity plus narrow IAM roles | Opt-in; new principal on redeploy |
| Anyone can call the agent | Custom role with only `aiplatform.reasoningEngines.query` on that agent | Google calls this coarse access control to the agent interface |
| Data exfiltration | VPC Service Controls | Some deployment types need extra ingress rules |
| Private egress | Private Service Connect interface | Needed for internet access inside a perimeter; changing it means recreating the agent |
| Key control | CMEK via `encryption_spec` | Cannot change after deploy; not on `us`/`eu` multi-region endpoints; some metadata not covered |
| Missing audit trail | Cloud Audit Logs | Admin Activity always on; Data Access off by default |
| No record of reasoning | Cloud Trace with OpenTelemetry | Prompt and response capture is opt-in |
| Prompt injection | Model Armor | Only through Agent Gateway |

## What should the governance checklist include?

1. Every agent deploys with Agent Identity, not the shared default.
2. Each identity has only the roles its tools need, and bindings are removed when the agent is deleted.
3. Query access is granted per agent with a custom role.
4. Agents run inside a VPC Service Controls perimeter, with a Private Service Connect interface where they need egress.
5. CMEK is set at deployment if your policy requires customer-managed keys.
6. Data Access audit logs are turned on for the project.
7. Tracing goes to Cloud Trace. Decide deliberately whether to set `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT`, since it records prompt and response content.
8. Inbound traffic passes through Agent Gateway if you rely on Model Armor. Model Armor screens ADK `streamQuery` inbound, not LangChain payloads.
9. Agent Platform Threat Detection in Security Command Center is evaluated. It is preview and needs the Premium tier.

## What does Google's stack not cover?

Agent Registry catalogs agents, MCP servers and endpoints, and Agent Gateway blocks connections by default unless IAM allows them. Cross-cloud support is limited preview and covers only Kubernetes clusters outside Google Cloud. The docs do not mention SaaS agents such as those on Salesforce, Microsoft or SAP. For a multi-cloud estate, you still need one inventory and one policy set across all of them, which is what an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane) provides. Our guide to [discovery permissions on AWS, Azure and Google Cloud](/blog/ai-agent-discovery-permissions-aws-azure-google-cloud) lists the read-only roles a discovery tool needs.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Google Cloud, AWS and Azure and has a runtime adapter for Vertex AI. Discovered agents enter a governed registry with owners, versions, tools and MCP servers. Posture flags risky permissions, such as agents sharing a broad service account. Per-agent Block and Quarantine act through the connected platform, and policies map to OWASP, MITRE ATLAS, NIST and ISO/IEC 42001. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What identity does a Vertex AI Agent Engine agent run as by default?**
The AI Platform Reasoning Engine Service Agent, or a custom service account you specify. Every agent can reach whatever that account can reach, which is why per-agent Agent Identity is recommended.

**Are prompts and responses logged by default?**
No. Data Access audit logs are off by default, and capturing message content in traces is opt-in through an OpenTelemetry environment variable.

**Is Vertex AI Agent Engine still the product name?**
Google has renamed it Agent Runtime on the Gemini Enterprise Agent Platform, and current documentation uses the new name. Check the name-changes page for how product and API names map.

## Key facts

Vertex AI Agent Engine is now Agent Runtime on the Gemini Enterprise Agent Platform. By default agents run as the Reasoning Engine Service Agent or a custom service account; per-agent SPIFFE-based Agent Identity is opt-in (identity_type=AGENT_IDENTITY). Data Access audit logs are off by default. Model Armor applies only through Agent Gateway. Opencontroller (by Lyzr) discovers agents across Google Cloud, AWS and Azure and has a runtime adapter for Vertex AI. Pricing starts from $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [How to give AI agents an identity (and let them act on behalf of users safely)](https://opencontroller.lyzr.com/blog/ai-agent-identity-on-behalf-of-delegation.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
