# Omnia vs Gemini Enterprise Agent Platform

> Omnia rolls out in steps, keeps every version, enforces policy on tool-call arguments, and deploys to Google’s Agent Runtime too. On Agent Runtime the tool gateway and a canary can’t run together, and policy sees which tool is called but not its arguments.

Category: Cloud agent platforms. Checked against Google Cloud’s published material on 2026-10-22. Canonical page: https://www.altairalabs.ai/vs/google-agent-platform/

## Google’s agent platform can’t canary behind its own gateway. Omnia can, and deploys to Google too.

Omnia gives you both at once: a stepped rollout behind the gateway, and policy on what a tool is called with. Google’s platform, formerly Vertex AI, has a simulation suite and a default-deny gateway, but on Agent Runtime the gateway and a canary can’t coexist.

## When it matters: Halfway through a canary, security asks for the agent to go behind the gateway.

**On Omnia:** The policy broker has sat beside the agent since its first version, and the canary carries on at its share of traffic. Every earlier version is still there to roll back to. (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

**On Gemini Enterprise Agent Platform:** Changing an Agent Gateway binding archives every earlier revision, archiving is irreversible, and binding changes are only allowed while all traffic is on the latest revision. (source: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-gateway-runtime-deploy)

## Policy that reads the arguments

A rule can allow transfer_funds under £500 to accounts the caller owns. CEL sees the call’s arguments and the caller’s identity, on any tool. *(enterprise)* (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

Gemini Enterprise Agent Platform: Access-policy conditions cover the tool name and annotation hints such as read-only or destructive, for MCP tools. Argument values aren’t an attribute. (source: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/policies/iam-overview-uap)

## A rollout that decides for itself

The traffic split starts once the suite passes and widens as the metrics hold. If they slip, traffic goes back to the previous version. *(enterprise)* (Omnia docs: https://promptarena.altairalabs.ai/arena/how-to/interfaces/run-in-ci/)

Gemini Enterprise Agent Platform: Revisions and manual traffic splits are in preview. An eval gate, metric-gated promotion and automatic rollback aren’t documented. (source: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-revisions-and-traffic)

## Telemetry into your own stack

OpenTelemetry with a documented span inventory, live to the collector you already run, plus Prometheus metrics and Grafana dashboards. (Omnia docs: https://omnia.altairalabs.ai/reference/platform/telemetry/)

Gemini Enterprise Agent Platform: Instrumented with OpenTelemetry and exported to Cloud Trace. Export to your own collector isn’t documented. (source: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/tracing)

## At a glance

| Topic | Omnia | Omnia docs | Gemini Enterprise Agent Platform | Gemini Enterprise Agent Platform docs | Source |
|---|---|---|---|---|---|
| Simulated users | Multi-turn, against real tools, adversarial included. | https://promptarena.altairalabs.ai/arena/explanation/scenario-design/ | Personas, with virtualised tools and fault injection. | documented | https://docs.cloud.google.com/gemini-enterprise-agent-platform/optimize/evaluation/evaluate-simulated |
| Voice testing | Barge-in, overlapping speech and dead air. | https://promptarena.altairalabs.ai/arena/how-to/voice/setup-voice-testing/ | Not documented. | not documented | https://docs.cloud.google.com/gemini-enterprise-agent-platform/optimize/evaluation/evaluate-simulated |
| Argument-level tool policy | A CEL rule on every call. *(enterprise)* | https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/ | Tool name and hints only. | partly documented | https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/policies/iam-overview-uap |
| Canary and policy together | Yes. *(enterprise)* | https://omnia.altairalabs.ai/tutorials/progressive-rollouts/ | Binding needs all traffic on the latest revision. | not documented | https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-gateway-runtime-deploy |
| End-user identity to tools | Carried through the agent to every call. | https://omnia.altairalabs.ai/explanation/platform/delegation-tokens/ | Three-legged OAuth for external tools, preview. | partly documented | https://docs.cloud.google.com/iam/docs/auth-with-3lo |
| Runs outside Google Cloud | Your cluster, AWS or Azure. | https://promptarena.altairalabs.ai/arena/explanation/deploy/overview/ | Gateway on other clusters is coming soon. | not documented | https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/gateways/agent-gateway-overview |

## Questions

**Is this the same as Vertex AI Agent Builder?**

Google describes the Gemini Enterprise Agent Platform as the evolution of Vertex AI, and Agent Engine is now Agent Runtime.

**Can I use Omnia with Google Cloud?**

Yes. Omnia deploys a pack to Google’s agent runtime with one command, uses Gemini on Vertex natively, and runs its scenarios against the deployed agent.