# Omnia vs Arize Phoenix

> Omnia runs the agent, so it can redact, refuse a tool call, hold a release and roll back on live scores. Arize traces and scores agents on OpenTelemetry, and acting on the scores is left to you.

Category: Tracing and evaluation. Checked against Arize AI’s published material on 2026-10-22. Canonical page: https://www.altairalabs.ai/vs/arize-phoenix/

## Arize shows you how the agent is doing. Omnia acts on it.

Omnia is the runtime the agent sits in, so redaction, tool policy and the release gate act on what the scores say. Its traces still go to Phoenix over OpenTelemetry.

## When it matters: A guardrail evaluator flags an agent leaking an account number.

**On Omnia:** The account number was a named field in the tool response, so it was replaced before the model saw it. A request for someone else’s account was refused by policy before the tool ran. Both are in the audit. (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

**On Arize Phoenix:** Arize’s guidance for real-time guardrails: AX doesn’t block requests, your app does. Evaluators run after the fact, asynchronously. (source: https://arize.com/docs/ax/cookbooks/observe/designing-realtime-guardrails)

## Enforcement where the agent runs

Policy compares each tool call’s arguments with the caller’s identity and fails closed. Named fields are redacted before they reach the model, the session recording or long-term memory. *(enterprise)* (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

Arize Phoenix: Guardrails are recorded as spans; your application does the blocking. (source: https://arize.com/docs/ax/cookbooks/observe/designing-realtime-guardrails)

## Tests that behave like users

Simulated callers hold whole conversations with the agent through its real tools. Some try to break it, some call by duplex voice, and the release waits on the results. (Omnia docs: https://promptarena.altairalabs.ai/arena/explanation/scenario-design/)

Arize Phoenix: Datasets, experiments and LLM, code and human evaluators. Simulated users aren’t documented. (source: https://arize.com/docs/phoenix)

## Live scores that move traffic

A new version starts on part of the traffic. Its live metrics decide whether it gets more or is rolled back. *(enterprise)* (Omnia docs: https://omnia.altairalabs.ai/tutorials/progressive-rollouts/)

Arize Phoenix: Online evals score sampled sessions; acting on them is up to you. (source: https://arize.com/docs/ax/evaluate/trace-and-session-evals)

## At a glance

| Topic | Omnia | Omnia docs | Arize Phoenix | Arize Phoenix docs | Source |
|---|---|---|---|---|---|
| Licence | PromptArena and PromptKit are Apache 2.0. | https://promptarena.altairalabs.ai/arena/explanation/scenario-design/ | Phoenix is Elastic License 2.0. | documented | https://github.com/Arize-ai/phoenix |
| OpenTelemetry | Emits and ingests. | https://omnia.altairalabs.ai/reference/platform/telemetry/ | Built on OpenTelemetry. | documented | https://arize.com/docs/phoenix |
| CI gate | Built into the release. | https://promptarena.altairalabs.ai/arena/how-to/interfaces/run-in-ci/ | pytest and Vitest plugins. | documented | https://arize.com/docs/phoenix/evaluation/integrations/pytest |
| Simulated users | Against real tools. | https://promptarena.altairalabs.ai/arena/explanation/scenario-design/ | Not documented. | not documented | https://arize.com/docs/phoenix |
| Runtime enforcement | Policy and redaction before the call. *(enterprise)* | https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/ | Your app. | not documented | https://arize.com/docs/ax/cookbooks/observe/designing-realtime-guardrails |
| Runs and rolls out agents | Yes. | https://omnia.altairalabs.ai/tutorials/progressive-rollouts/ | Observes only. | not documented | https://arize.com/docs/phoenix |

## Questions

**Can I keep Phoenix and use Omnia?**

Yes. Omnia emits standard OpenTelemetry, which Phoenix ingests.