We already run agents on Kubernetes with kagent.
Then you already have the part the two agree on: agents as custom resources, each session pinned to an immutable revision. Kagent’s docs stop at the revision. Omnia tests it against simulated callers before it takes traffic and promotes it on the results.
Before it serves anyone, the new revision meets simulated customers on the real tools, a caller pushing for a policy exception and one who talks over it on the phone. It grants the exception more readily than the old model did, so it stays at the gate. docs
Kagent compiles the change into a new digest-pinned revision. Its docs describe no testing before the revision serves; Solo.io’s separate agentevals tool grades traces from runs that already happened. their docs
A test suite on every revision
Traffic in steps
Both pin a session to the revision it started on. Omnia also moves new conversations across in steps, checks the metrics before each one, and rolls the revision back if they drop. docs enterprise
Kagent: Sessions pin to their revision. Canary, eval-gated promotion and rollback aren’t documented. source
The user’s identity reaches the tool
OIDC or your gateway authenticates the caller, and the agent passes that identity to every tool and remote agent it calls. It’s in the free core. A CEL rule can also check a tool call’s arguments before it runs. docs enterprise
Kagent: Open source has Kubernetes RBAC, an oauth2-proxy and a per-binding tool allowlist. OIDC and on-behalf-of token exchange are in Solo Enterprise for kagent. source
At a glance
| Topic | Omnia | Kagent, per its docs |
|---|---|---|
| Agents as Kubernetes resources | yes Custom resources and controllers, installed with Helm. docs | documented Harness, AgentTemplate and Agent resources. source |
| Immutable versions | yes A versioned pack: prompts, tools, policy and tests together. docs | documented Digest-identified revisions. source |
| Simulated users before release | yes Multi-turn, against real tools, adversarial included. docs | not documented Not in the kagent docs. source |
| Voice | yes Duplex voice, and tests for it. docs | not documented Not documented. source |
| Stepped rollout, auto rollback | yes Traffic shifts in steps and rolls back on its own. docs | not documented Not documented. source |
| Argument-level tool policy | yes A CEL rule on every call, before the tool runs. docs enterprise | partly Allowlist by tool name, optional approval. source |
| End-user identity to tools | yes In the free core. docs | partly In the enterprise edition. source |
| Deploy to AgentCore, Vertex or Foundry | yes The same pack, one command. docs | not documented Not documented. source |
Common questions
- Which kagent version is this?
- The 1.x documentation, which kagent describes as an alpha and a clean break from the 0.x line.
- Does either need a service mesh?
- Kagent’s open-source install doesn’t mention one; Istio ambient appears in Solo’s enterprise edition. Omnia doesn’t need one.
- Can Omnia run agents built with other frameworks?
- Yes. They run under the same policy, audit and testing as Omnia’s own, and any agent is reachable over A2A or as an MCP tool.
Put your own agent through the gate.
Install Omnia in your cluster and run it free while you build.
Kagent claims are from The kagent project’s published documentation, checked 22 October 2026. Spot something out of date? Tell us.