diff --git a/python/src/agent_squad/agents/openai_agent.py b/python/src/agent_squad/agents/openai_agent.py
index 745a1d75..cf68768b 100644
--- a/python/src/agent_squad/agents/openai_agent.py
+++ b/python/src/agent_squad/agents/openai_agent.py
@@ -164,11 +164,14 @@ async def handle_single_response(self, request_options: dict[str, Any]) -> Conve
             raise error
 
     async def handle_streaming_response(self, request_options: dict[str, Any]) -> AsyncGenerator[AgentStreamResponse, None]:
+        """Yield text and the final message, skipping chunks without completion choices."""
         try:
             stream = self.client.chat.completions.create(**request_options)
             accumulated_message = []
 
             for chunk in stream:
+                if not chunk.choices:
+                    continue
                 if chunk.choices[0].delta.content:
                     chunk_content = chunk.choices[0].delta.content
                     accumulated_message.append(chunk_content)
@@ -208,4 +211,4 @@ def replace(match):
                 return '\n'.join(value) if isinstance(value, list) else str(value)
             return match.group(0)
 
-        return re.sub(r'{{(\w+)}}', replace, template)
\ No newline at end of file
+        return re.sub(r'{{(\w+)}}', replace, template)
diff --git a/python/src/tests/agents/test_openai_usage_stream.py b/python/src/tests/agents/test_openai_usage_stream.py
new file mode 100644
index 00000000..52ee1ebc
--- /dev/null
+++ b/python/src/tests/agents/test_openai_usage_stream.py
@@ -0,0 +1,69 @@
+from unittest.mock import AsyncMock, Mock, call
+
+import pytest
+from openai.types.chat import ChatCompletionChunk
+
+from agent_squad.agents import OpenAIAgent, OpenAIAgentOptions
+
+
+def _chunk(content=None, *, usage_only=False):
+    return ChatCompletionChunk.model_validate(
+        {
+            "id": "chatcmpl-test",
+            "object": "chat.completion.chunk",
+            "created": 0,
+            "model": "test-model",
+            "choices": [] if usage_only else [{"index": 0, "delta": {"content": content}, "finish_reason": None}],
+            "usage": {"prompt_tokens": 1, "completion_tokens": 2, "total_tokens": 3} if usage_only else None,
+        }
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("usage_position", ["start", "middle", "end", "only"])
+async def test_usage_only_chunks_preserve_stream_and_final_message(usage_position):
+    client = Mock()
+    agent = OpenAIAgent(
+        OpenAIAgentOptions(
+            name="TestAgent", description="Test streaming", api_key="test-key", client=client, streaming=True
+        )
+    )
+    agent.callbacks.on_llm_new_token = AsyncMock()
+    chunks = [_chunk(), _chunk("hello "), _chunk("world")]
+    if usage_position == "only":
+        chunks = [_chunk(usage_only=True)]
+        expected_text = ""
+        expected_calls = []
+    else:
+        chunks.insert({"start": 0, "middle": 2, "end": 3}[usage_position], _chunk(usage_only=True))
+        expected_text = "hello world"
+        expected_calls = [call("hello "), call("world")]
+    client.chat.completions.create.return_value = iter(chunks)
+
+    stream = await agent.process_request("hello", "user", "session", [])
+    responses = [response async for response in stream]
+
+    assert "".join(response.text or "" for response in responses) == expected_text
+    assert responses[-1].final_message.content == [{"text": expected_text}]
+    assert responses[-1].final_message.role == "assistant"
+    assert agent.callbacks.on_llm_new_token.await_args_list == expected_calls
+    client.chat.completions.create.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_stream_errors_after_usage_chunks_still_propagate():
+    client = Mock()
+    agent = OpenAIAgent(
+        OpenAIAgentOptions(
+            name="TestAgent", description="Test streaming", api_key="test-key", client=client, streaming=True
+        )
+    )
+
+    def failing_stream():
+        yield _chunk(usage_only=True)
+        raise RuntimeError("stream disconnected")
+
+    client.chat.completions.create.return_value = failing_stream()
+    with pytest.raises(RuntimeError, match="stream disconnected"):
+        async for _ in agent.handle_streaming_response({"stream": True}):
+            pass