diff --git a/agent/relay_runtime.py b/agent/relay_runtime.py
index 080f5161ddbe..0b1c341d34c9 100644
--- a/agent/relay_runtime.py
+++ b/agent/relay_runtime.py
@@ -66,7 +66,7 @@ class RelayRuntime:
             self._execution_consumers.discard(consumer)
 
     def managed_execution_enabled(self) -> bool:
-        """Return whether an active interceptor or subscriber needs the pipeline."""
+        """Return whether a Hermes-managed consumer needs the Relay pipeline."""
         with self._execution_consumers_lock:
             return bool(self._execution_consumers)
 
diff --git a/plugins/observability/nemo_relay/README.md b/plugins/observability/nemo_relay/README.md
index 8dd838908a69..85f7d039031f 100644
--- a/plugins/observability/nemo_relay/README.md
+++ b/plugins/observability/nemo_relay/README.md
@@ -443,9 +443,11 @@ separate trajectories.
 
 ## Adaptive Execution Example
 
-Hermes core always hands LLM and tool calls to NeMo Relay managed execution.
-The `observability/nemo_relay` plugin can install adaptive components on those
-boundaries.
+Hermes core owns the LLM and tool boundaries and enters NeMo Relay managed
+execution while a Hermes-managed Relay consumer is active. With no shared
+metrics subscriber or explicitly configured Relay plugin, Hermes calls the
+provider or tool directly. The `observability/nemo_relay` plugin retains the
+managed path while its adaptive components are installed on those boundaries.
 
 Minimal `plugins.toml`:
 
diff --git a/tests/agent/test_auxiliary_relay.py b/tests/agent/test_auxiliary_relay.py
index a69387a35573..0c9e8cc51388 100644
--- a/tests/agent/test_auxiliary_relay.py
+++ b/tests/agent/test_auxiliary_relay.py
@@ -185,6 +185,8 @@ def test_auxiliary_stream_uses_streaming_relay_primitive(monkeypatch):
 
 def test_auxiliary_attempt_uses_real_relay_request_intercepts(relay_turn):
     relay, turn = relay_turn
+    consumer = "test.auxiliary-request-intercept"
+    turn.lease.host.retain_managed_execution(consumer)
     captured_requests = []
     client = SimpleNamespace(
         chat=SimpleNamespace(
@@ -228,6 +230,7 @@ def test_auxiliary_attempt_uses_real_relay_request_intercepts(relay_turn):
         result = run("compression")
     finally:
         relay.intercepts.deregister_llm_request("hermes-auxiliary-request")
+        turn.lease.host.release_managed_execution(consumer)
 
     assert result.choices[0].message.content == "ok"
     assert captured_requests[0]["temperature"] == 0.25