diff --git a/trpc_agent_sdk/dsl/graph/_node_action/_agent.py b/trpc_agent_sdk/dsl/graph/_node_action/_agent.py
index adf1e7a7..85d064dc 100644
--- a/trpc_agent_sdk/dsl/graph/_node_action/_agent.py
+++ b/trpc_agent_sdk/dsl/graph/_node_action/_agent.py
@@ -181,11 +181,17 @@ async def execute(self, state: State) -> dict[str, Any]:
                         if isinstance(candidate, str) and candidate:
                             last_response = candidate
 
-                    if (not self._is_graph_event(event)) and (
-                            not event.partial) and event.content and event.content.parts:
-                        text_parts = [part.text for part in event.content.parts if part.text]
-                        if text_parts:
-                            last_response = text_parts[-1]
+                    if (not self._is_graph_event(event)) and event.is_final_response():
+                        # Only clean final answers may become last_response. Skip
+                        # thought parts (part.thought) so a thinking model's
+                        # reasoning monologue is never surfaced as the answer.
+                        # is_final_response() already excludes partial, function
+                        # call/response, and code-execution events, so plan and
+                        # transfer_to_agent rounds are naturally ignored.
+                        if event.content and event.content.parts:
+                            visible_text = [part.text for part in event.content.parts if part.text and not part.thought]
+                            if visible_text:
+                                last_response = "".join(visible_text)
 
                     if not event.visible:
                         if event.actions and event.actions.transfer_to_agent:
@@ -246,7 +252,7 @@ async def execute(self, state: State) -> dict[str, Any]:
 
         node_response: Any = last_response
         structured_output: Any = None
-        if isinstance(self.agent, LlmAgent) and self.agent.output_schema is not None:
+        if isinstance(self.agent, LlmAgent) and self.agent.output_schema is not None and last_response:
             node_response = json.loads(last_response)
             structured_output = node_response
 
diff --git a/trpc_agent_sdk/dsl/graph/_node_action/_llm.py b/trpc_agent_sdk/dsl/graph/_node_action/_llm.py
index 631e4d2a..4c0bcb16 100644
--- a/trpc_agent_sdk/dsl/graph/_node_action/_llm.py
+++ b/trpc_agent_sdk/dsl/graph/_node_action/_llm.py
@@ -251,7 +251,9 @@ async def _run_model_round(
                     continue
 
                 response_parts = list(llm_response.content.parts)
-                text_parts = [part.text for part in response_parts if part.text]
+                # Exclude thought parts so a thinking model's reasoning does not
+                # leak into response_text / STATE_KEY_LAST_RESPONSE.
+                text_parts = [part.text for part in response_parts if part.text and not part.thought]
                 response_text = "".join(text_parts)
 
             logger.debug(f"[{self.name}] LLM response received ({len(response_text)} chars)")