Relevance and Omnia run the same build, test and release loop, and Omnia runs it in your own cluster.
Relevance builds agents, tests them on simulated users and can block a failing publish, on its own multi-tenant cloud. On Omnia the conversations stay in your cluster, with personal data removed before they’re stored.
The number is redacted before the session is recorded, redacted again before anything reaches long-term memory, and replaced in tool responses before the model sees it. What’s left is stored in your cluster under your retention rules. docs
on relevance ai Relevance runs a multi-tenant architecture with customer data logically separated, in a US, EU or Australian region. Single-tenant is described as in the works. their docs
The data stays inside your boundary
Omnia installs into your Kubernetes cluster. Sessions move from Redis to Postgres to object storage you own, and the archive is Parquet you query from Snowflake, BigQuery or DuckDB. Traces go live to your OpenTelemetry collector. docs
Relevance AI: Data is stored on Relevance’s platform; telemetry can be exported in OpenTelemetry format to your S3 bucket. source
A bad version comes back out on its own
A version that passes the gate starts on part of the traffic and gets more while the metrics hold; if they drop, the previous version takes the traffic back. Someone mid-conversation stays on the version they started with. docs enterprise
Relevance AI: Restoring a version creates a new draft, which you then publish. Stepped rollout isn’t documented. source
Policy that knows who’s asking
The caller’s identity travels through the agent to every tool. A CEL rule compares each call’s arguments with it, and a refund to an account the caller doesn’t own is refused. docs enterprise
Relevance AI: Human approval per tool. Argument-level policy and end-user identity in tool calls aren’t documented. source
At a glance
| Topic | Omnia | Relevance AI, per its docs |
|---|---|---|
| Build from a conversation | yes The Arena builder writes the agent and its tests from a brief, and runs them on real providers. docs enterprise | documented Invent builds and manages agents from a conversation. source |
| Runs in your infrastructure | yes Your cluster, or AgentCore, Vertex or Foundry in your account. docs | not documented Multi-tenant SaaS. source |
| Simulated users | yes Multi-turn, against real tools, adversarial included. docs | documented Multi-turn, with live or simulated tools. source |
| Eval-gated release | yes The suite decides whether a version ships. docs | documented Optional publish block, Enterprise first. source |
| Voice testing | yes Barge-in, overlapping speech and dead air. docs | not documented Phone agents ship; testing them isn’t documented. source |
| Stepped rollout, auto rollback | yes Traffic shifts in steps and rolls back on its own. docs | not documented Not documented. source |
| Live telemetry to your collector | yes OpenTelemetry and Prometheus. docs | partly OpenTelemetry events to S3. source |
| Argument-level tool policy | yes A CEL rule on every call, before the tool runs. docs enterprise | not documented Not documented. source |
| Any model, including your own | yes Seven native protocols and any OpenAI-compatible URL. docs | partly OpenAI, Gemini, Azure, OpenRouter; Anthropic via your cloud account. source |
Common questions
- Can I self-host Relevance AI?
- Its security documentation describes a multi-tenant platform with regional storage, and single-tenant as in the works. Omnia runs in your own cluster from day one.
- Do both test agents with simulated users?
- Yes. Omnia adds adversarial scenarios and duplex voice testing, and runs them inside your boundary against your real tools.
- What does Omnia cost?
- It’s free while you build and the core is free in production. The enterprise licencelicense is one price per environment per year from £30,000, and nothing is metered.
Put your own agent through the gate.
Install Omnia in your cluster and run it free while you build.
Relevance AI claims are from Relevance AI’s published documentation, checked 22 October 2026. Spot something out of date? Tell us.