diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py
index 39ae9a759c3..c24cc972a2e 100644
--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -5711,6 +5711,8 @@ def call_llm(
     timeout: float = None,
     extra_body: dict = None,
     api_mode: str = None,
+    stream: bool = False,
+    stream_options: dict = None,
 ) -> Any:
     """Centralized synchronous LLM call.
 
@@ -5731,9 +5733,16 @@ def call_llm(
         tools: Tool definitions (for function calling).
         timeout: Request timeout in seconds (None = read from auxiliary.{task}.timeout config).
         extra_body: Additional request body fields.
+        stream: When True, return the raw SDK streaming iterator instead of a
+            validated complete response. The caller is responsible for consuming
+            chunks (and for any fallback). Used by the MoA aggregator so its
+            output can stream to the user.
+        stream_options: Passed through to the request when stream is True
+            (e.g. {"include_usage": True}).
 
     Returns:
-        Response object with .choices[0].message.content
+        Response object with .choices[0].message.content, OR — when stream=True —
+        the raw streaming iterator from client.chat.completions.create().
 
     Raises:
         RuntimeError: If no provider is configured.
@@ -5835,6 +5844,20 @@ def call_llm(
     if _is_anthropic_compat_endpoint(resolved_provider, _client_base):
         kwargs["messages"] = _convert_openai_images_to_anthropic(kwargs["messages"])
 
+    # Streaming path: return the raw SDK Stream iterator directly. This is used by
+    # the MoA aggregator so its tokens stream to the user. It deliberately skips
+    # _validate_llm_response and the temperature/max_tokens/payment fallback chain
+    # below — those all assume a complete response object, whereas a stream is
+    # consumed chunk-by-chunk by the caller. The caller (the agent's streaming
+    # consumer) owns chunk reassembly, stale-stream detection, and falling back to
+    # a non-streaming call on error. stream_options is best-effort: providers that
+    # reject it surface an error the caller's fallback already handles.
+    if stream:
+        kwargs["stream"] = True
+        if stream_options:
+            kwargs["stream_options"] = stream_options
+        return client.chat.completions.create(**kwargs)
+
     # Handle unsupported temperature, max_tokens vs max_completion_tokens retry,
     # then payment fallback.
     try:
diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index 4ca09c6895d..7a5919807af 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -1168,11 +1168,22 @@ def run_conversation(
                 # stream.  Mirror the ACP exclusion used for Responses
                 # API upgrade (lines ~1083-1085).
                 elif (
-                    agent.provider in {"copilot-acp", "moa"}
+                    agent.provider in {"copilot-acp"}
                     or str(agent.base_url or "").lower().startswith("acp://copilot")
                     or str(agent.base_url or "").lower().startswith("acp+tcp://")
                 ):
                     _use_streaming = False
+                # MoA streams only when a display/TTS consumer is present to
+                # receive the deltas. MoAChatCompletions.create() honors
+                # stream=True (runs the references, then returns the aggregator's
+                # raw token stream) and is reached here because, for provider
+                # "moa", _create_request_openai_client returns the MoA facade
+                # itself. Without consumers (quiet mode, subagents, health-check
+                # probes) we keep the complete-response path: the facade returns a
+                # whole response when stream is not requested, preserving the
+                # prior behavior for those callers.
+                elif agent.provider == "moa" and not agent._has_stream_consumers():
+                    _use_streaming = False
                 elif not agent._has_stream_consumers():
                     # No display/TTS consumer. Still prefer streaming for
                     # health checking, but skip for Mock clients in tests
diff --git a/agent/moa_loop.py b/agent/moa_loop.py
index 583a0d56ccb..fcc76c2cf0f 100644
--- a/agent/moa_loop.py
+++ b/agent/moa_loop.py
@@ -577,6 +577,24 @@ class MoAChatCompletions:
         # max_tokens is passed through from the caller (normally None → omitted
         # → the model's real maximum). The preset's old hardcoded 4096 default
         # is gone — it truncated long syntheses.
+        # When the agent's streaming consumer calls us with stream=True, run the
+        # references first (above) and then return the aggregator's RAW token
+        # stream so the acting model's output reaches the user live. The consumer
+        # reassembles chunks + tool_calls, runs stale-stream detection, and falls
+        # back to a non-streaming retry on error. The non-streaming path
+        # (stream=False) is unchanged — no stream/stream_options/timeout are
+        # forwarded, so its behavior is byte-for-byte identical to before.
+        stream = bool(api_kwargs.get("stream"))
+        stream_kwargs: dict[str, Any] = {}
+        if stream:
+            stream_kwargs["stream"] = True
+            stream_kwargs["stream_options"] = (
+                api_kwargs.get("stream_options") or {"include_usage": True}
+            )
+            # Forward the consumer's per-request (stream read) timeout so it
+            # actually governs the aggregator stream, not just call_llm's default.
+            if api_kwargs.get("timeout") is not None:
+                stream_kwargs["timeout"] = api_kwargs["timeout"]
         return call_llm(
             task="moa_aggregator",
             messages=agg_messages,
@@ -584,6 +602,7 @@ class MoAChatCompletions:
             max_tokens=agg_kwargs.get("max_tokens"),
             tools=agg_kwargs.get("tools"),
             extra_body=agg_kwargs.get("extra_body"),
+            **stream_kwargs,
             **_slot_runtime(aggregator),
         )
 
diff --git a/tests/run_agent/test_moa_streaming.py b/tests/run_agent/test_moa_streaming.py
new file mode 100644
index 00000000000..a61780aca08
--- /dev/null
+++ b/tests/run_agent/test_moa_streaming.py
@@ -0,0 +1,221 @@
+"""Tests for MoA aggregator streaming.
+
+MoAChatCompletions.create() honors stream=True by running the references first
+and then returning the aggregator's raw streaming iterator (from call_llm), so
+the acting model's output can stream to the user. stream=False is the original
+complete-response path and must stay byte-identical.
+"""
+from types import SimpleNamespace
+
+import pytest
+
+
+def _response(content="done", *, tool_calls=None):
+    message = SimpleNamespace(content=content, tool_calls=tool_calls or [])
+    choice = SimpleNamespace(message=message, finish_reason="stop")
+    return SimpleNamespace(choices=[choice], usage=None, model="fake-model")
+
+
+def _write_cfg(home):
+    home.mkdir()
+    (home / "config.yaml").write_text(
+        """
+moa:
+  default_preset: review
+  presets:
+    review:
+      reference_models:
+        - provider: openai-codex
+          model: gpt-5.5
+      aggregator:
+        provider: openrouter
+        model: anthropic/claude-opus-4.8
+""".strip(),
+        encoding="utf-8",
+    )
+
+
+def _facade(monkeypatch, tmp_path, on_call=None):
+    home = tmp_path / ".hermes"
+    _write_cfg(home)
+    monkeypatch.setenv("HERMES_HOME", str(home))
+    calls = []
+
+    def fake_call_llm(**kwargs):
+        calls.append(kwargs)
+        if on_call is not None:
+            r = on_call(kwargs)
+            if r is not None:
+                return r
+        if kwargs["task"] == "moa_reference":
+            return _response("reference advice")
+        return _response("aggregator acted")
+
+    monkeypatch.setattr("agent.moa_loop.call_llm", fake_call_llm)
+    from agent.moa_loop import MoAChatCompletions
+
+    return MoAChatCompletions("review"), calls
+
+
+# --------------------------------------------------------------------------
+# Facade-level: create() stream branch
+# --------------------------------------------------------------------------
+
+def test_create_streams_aggregator_when_requested(monkeypatch, tmp_path):
+    """stream=True: references still run, aggregator is called with stream=True
+    and stream_options, and create() returns the aggregator call's result
+    (the raw stream) verbatim."""
+    sentinel = object()
+
+    def on_call(kwargs):
+        if kwargs["task"] == "moa_aggregator":
+            return sentinel
+        return None
+
+    facade, calls = _facade(monkeypatch, tmp_path, on_call=on_call)
+    out = facade.create(
+        messages=[{"role": "user", "content": "q"}],
+        tools=[{"type": "function"}],
+        stream=True,
+    )
+
+    # create() returns the aggregator's streaming result untouched.
+    assert out is sentinel
+    # References still ran (MoA not bypassed).
+    assert any(c["task"] == "moa_reference" for c in calls)
+    agg = next(c for c in calls if c["task"] == "moa_aggregator")
+    assert agg["stream"] is True
+    assert agg["stream_options"] == {"include_usage": True}
+    # Tools still flow to the (streaming) aggregator.
+    assert agg["tools"] is not None
+
+
+def test_create_non_stream_path_unchanged(monkeypatch, tmp_path):
+    """Default (no stream): the aggregator call carries NO stream/stream_options
+    keys, so the non-streaming path is byte-identical to before."""
+    facade, calls = _facade(monkeypatch, tmp_path)
+    facade.create(messages=[{"role": "user", "content": "q"}], tools=[])
+
+    agg = next(c for c in calls if c["task"] == "moa_aggregator")
+    assert "stream" not in agg
+    assert "stream_options" not in agg
+    assert "timeout" not in agg
+
+
+def test_create_forwards_stream_read_timeout(monkeypatch, tmp_path):
+    """The consumer's per-request (stream read) timeout is forwarded to the
+    aggregator so it actually governs the stream."""
+    timeout_sentinel = object()
+    facade, calls = _facade(monkeypatch, tmp_path)
+    facade.create(
+        messages=[{"role": "user", "content": "q"}],
+        tools=[],
+        stream=True,
+        timeout=timeout_sentinel,
+    )
+    agg = next(c for c in calls if c["task"] == "moa_aggregator")
+    assert agg["timeout"] is timeout_sentinel
+
+
+def test_create_respects_caller_stream_options(monkeypatch, tmp_path):
+    """A caller-provided stream_options is forwarded as-is (not overwritten)."""
+    facade, calls = _facade(monkeypatch, tmp_path)
+    facade.create(
+        messages=[{"role": "user", "content": "q"}],
+        tools=[],
+        stream=True,
+        stream_options={"include_usage": False, "extra": 1},
+    )
+    agg = next(c for c in calls if c["task"] == "moa_aggregator")
+    assert agg["stream_options"] == {"include_usage": False, "extra": 1}
+
+
+def test_create_does_not_forward_timeout_when_not_streaming(monkeypatch, tmp_path):
+    """A stray timeout on a non-streaming call is NOT forwarded — the non-stream
+    path must remain unchanged regardless of incidental kwargs."""
+    facade, calls = _facade(monkeypatch, tmp_path)
+    facade.create(messages=[{"role": "user", "content": "q"}], tools=[], timeout=object())
+    agg = next(c for c in calls if c["task"] == "moa_aggregator")
+    assert "timeout" not in agg
+    assert "stream" not in agg
+
+
+# --------------------------------------------------------------------------
+# call_llm-level: stream branch returns the raw SDK stream
+# --------------------------------------------------------------------------
+
+def test_call_llm_stream_returns_raw_stream_and_skips_validation(monkeypatch):
+    """call_llm(stream=True) returns the client's raw stream object directly,
+    attaches stream/stream_options to the request, and does NOT run response
+    validation (which assumes a complete response)."""
+    from agent import auxiliary_client as ac
+
+    captured = {}
+
+    class _Completions:
+        def create(self, **kwargs):
+            captured.update(kwargs)
+            return "RAW_STREAM"
+
+    fake_client = SimpleNamespace(
+        chat=SimpleNamespace(completions=_Completions()),
+        base_url="http://localhost:8001/v1",
+    )
+
+    monkeypatch.setattr(
+        ac, "_resolve_task_provider_model",
+        lambda *a, **k: ("custom", "m", "http://localhost:8001/v1", "key", "chat_completions"),
+    )
+    monkeypatch.setattr(ac, "_get_cached_client", lambda *a, **k: (fake_client, "m"))
+
+    def _no_validate(*a, **k):
+        raise AssertionError("streaming must not go through _validate_llm_response")
+
+    monkeypatch.setattr(ac, "_validate_llm_response", _no_validate)
+
+    out = ac.call_llm(
+        provider="custom",
+        model="m",
+        messages=[{"role": "user", "content": "hi"}],
+        stream=True,
+        stream_options={"include_usage": True},
+    )
+
+    assert out == "RAW_STREAM"
+    assert captured.get("stream") is True
+    assert captured.get("stream_options") == {"include_usage": True}
+
+
+def test_call_llm_non_stream_still_validates(monkeypatch):
+    """Sanity: stream=False keeps the validated path (regression guard for the
+    early-return not leaking into normal calls)."""
+    from agent import auxiliary_client as ac
+
+    class _Completions:
+        def create(self, **kwargs):
+            return _response("ok")
+
+    fake_client = SimpleNamespace(
+        chat=SimpleNamespace(completions=_Completions()),
+        base_url="http://localhost:8001/v1",
+    )
+    monkeypatch.setattr(
+        ac, "_resolve_task_provider_model",
+        lambda *a, **k: ("custom", "m", "http://localhost:8001/v1", "key", "chat_completions"),
+    )
+    monkeypatch.setattr(ac, "_get_cached_client", lambda *a, **k: (fake_client, "m"))
+
+    validated = {"called": False}
+
+    def _validate(resp, task):
+        validated["called"] = True
+        return resp
+
+    monkeypatch.setattr(ac, "_validate_llm_response", _validate)
+
+    ac.call_llm(
+        provider="custom",
+        model="m",
+        messages=[{"role": "user", "content": "hi"}],
+    )
+    assert validated["called"] is True