Issue

When routing traffic through AgentGateway to an InferencePool backend, the translator fails to resolve the backend reference and logs the following error:

{"time":"2026-03-20T12:24:39.701030772Z","level":"error","msg":"erroring building agent gateway destination","component":"agentgateway/translator","error":{"Type":"ResolvedRefs","Status":"False","Reason":"BackendNotFound","Message":"backendRef default/inferencepool not found"}}

Despite the pool being apparently located ("found pull pool for service"), the destination cannot be built and the route condition is set to ResolvedRefs: False.

Explanation

AgentGateway’s translator successfully looks up the InferencePool object by namespace and name, but then fails to build a valid destination from it. This happens because support for InferencePool backends relies on the Inference Extension feature, which is disabled by default in the Helm chart.

Without this feature flag enabled, AgentGateway does not register the necessary reconcilers and endpoint resolution logic for InferencePool resources. As a result, even though the object exists in the cluster, the translator cannot construct a valid backend destination and raises BackendNotFound.

Resolution

Enable the inferenceExtension feature in your AgentGateway Helm values:

inferenceExtension:
  enabled: true

Refer to the official values.yaml for the full list of available options under inferenceExtension.

After updating the chart, redeploy AgentGateway. The translator will then correctly resolve InferencePool backends and the BackendNotFound error should be gone.