# Omnia vs Agent Bricks

> Omnia tests every version on simulated users, gates the release, rolls it out in steps and runs it in your own Kubernetes cluster, for agents whose tools reach past one data platform. Agent Bricks keeps agents inside Databricks.

Category: Agent platforms. Checked against Databricks’s published material on 2026-10-22. Canonical page: https://www.altairalabs.ai/vs/databricks-agent-bricks/

## Agent Bricks keeps agents next to your data. Getting each new version into production safely is Omnia’s job.

Omnia runs in your own cluster for agents whose tools reach past one data platform, and gets each new version into production through a test gate and a stepped rollout. Agent Bricks puts agents next to Unity Catalog, with its own judges and tracing.

## When it matters: A finance agent gets a new prompt the week before quarter close.

**On Omnia:** Before it serves anyone, the new version meets simulated controllers through the real tools, one of them pushing an out-of-policy journal entry. It passes and goes to part of the traffic first. Its numbers hold, and it reaches everyone before close. (Omnia docs: https://promptarena.altairalabs.ai/arena/explanation/scenario-design/)

**On Agent Bricks:** Multi-turn conversation simulation is experimental, and its API and behaviour might change. An eval-gated release, staged rollout and automatic rollback aren’t documented. (source: https://docs.databricks.com/aws/en/mlflow3/genai/eval-monitor/conversation-simulation)

## Simulation you can build a release on

Omnia’s simulated users are generally available, with adversarial and voice testing alongside. A version that fails them doesn’t ship. (Omnia docs: https://promptarena.altairalabs.ai/arena/explanation/scenario-design/)

Agent Bricks: An LLM-driven simulator with goals and personas, marked experimental. Voice testing isn’t documented. (source: https://docs.databricks.com/aws/en/mlflow3/genai/eval-monitor/conversation-simulation)

## Stepped rollout

The old version keeps serving while the new one takes traffic in steps. A conversation stays on the version it started on, and if the new version’s metrics drop, the traffic goes back. *(enterprise)* (Omnia docs: https://omnia.altairalabs.ai/tutorials/progressive-rollouts/)

Agent Bricks: Agents deploy with the Agent Bricks CLI. Versioning, staged rollout and rollback aren’t documented. (source: https://docs.databricks.com/gcp/en/agents/custom-agents/agent-bricks-cli)

## Runs where the rest of your stack runs

Omnia installs into your Kubernetes cluster and deploys the same pack to AWS Bedrock AgentCore, Google’s Agent Runtime or Azure AI Foundry. Telemetry goes live to your own OpenTelemetry collector. (Omnia docs: https://promptarena.altairalabs.ai/arena/explanation/deploy/overview/)

Agent Bricks: Agents run on Databricks-hosted compute, and traces are stored in OpenTelemetry format in Unity Catalog tables. (source: https://docs.databricks.com/aws/en/agents/custom-agents/build-agents)

## At a glance

| Topic | Omnia | Omnia docs | Agent Bricks | Agent Bricks docs | Source |
|---|---|---|---|---|---|
| Simulated users | Generally available, adversarial included. | https://promptarena.altairalabs.ai/arena/explanation/scenario-design/ | Experimental. | partly documented | https://docs.databricks.com/aws/en/mlflow3/genai/eval-monitor/conversation-simulation |
| Eval-gated release | The suite decides whether a version ships. | https://promptarena.altairalabs.ai/arena/how-to/interfaces/run-in-ci/ | Not documented. | not documented | https://docs.databricks.com/aws/en/agents/concepts/ |
| Stepped rollout, auto rollback | Traffic shifts in steps and rolls back on its own. | https://omnia.altairalabs.ai/tutorials/progressive-rollouts/ | Not documented. | not documented | https://docs.databricks.com/gcp/en/agents/custom-agents/agent-bricks-cli |
| End-user identity to tools | Carried through the agent to every call. | https://omnia.altairalabs.ai/explanation/platform/delegation-tokens/ | The user’s token is forwarded, scoped to the workspace. | documented | https://docs.databricks.com/aws/en/dev-tools/databricks-apps/auth |
| Argument-level tool policy | A CEL rule on every call, before the tool runs. *(enterprise)* | https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/ | Access, guardrails and rate limits on models, MCP servers and skills. | partly documented | https://docs.databricks.com/aws/en/agents/custom-agents/build-agents |
| Runs in your own cluster | Yes, installed with Helm. | https://omnia.altairalabs.ai/how-to/workspaces/manage-workspaces/ | Databricks-hosted compute. | not documented | https://docs.databricks.com/aws/en/agents/custom-agents/build-agents |
| Voice | Duplex voice, and tests for it. | https://promptarena.altairalabs.ai/arena/explanation/duplex-architecture/ | Not documented. | not documented | https://docs.databricks.com/aws/en/agents/concepts/ |

## Questions

**Can an Omnia agent use data in Databricks?**

Yes. Point it at the APIs or MCP servers you already expose, with the end user’s identity carried to the call.

**Which models can each use?**

Both are broad. Omnia speaks seven providers’ native protocols and any OpenAI-compatible endpoint, including a model you host.