diff --git a/backend/app/main_graph/nodes/agent.py b/backend/app/main_graph/nodes/agent.py
index 225468f..7248269 100644
--- a/backend/app/main_graph/nodes/agent.py
+++ b/backend/app/main_graph/nodes/agent.py
@@ -68,22 +68,36 @@ def create_agent_node(llm_with_tools, llm):
         """
         info(f"[Agent] 第 {state.current_step} 步推理")
 
-        # 组装完整消息：系统提示 + 历史消息
-        full_messages = [SystemMessage(content=SYSTEM_PROMPT)] + state.messages
+        try:
+            # 组装完整消息：系统提示 + 历史消息
+            full_messages = [SystemMessage(content=SYSTEM_PROMPT)] + state.messages
+            
+            info(f"[Agent] 消息数量: {len(full_messages)}, 最后一条: {type(full_messages[-1]).__name__}")
 
-        # 判断是否达到步数上限
-        if state.current_step >= state.max_steps:
-            info(f"[Agent] 达到步数上限 {state.max_steps}，强制结束，不绑定工具")
-            llm_no_tools = llm.bind_tools([])
-            response = await llm_no_tools.ainvoke(full_messages)
-        else:
-            response = await llm_with_tools.ainvoke(full_messages)
+            # 判断是否达到步数上限
+            if state.current_step >= state.max_steps:
+                info(f"[Agent] 达到步数上限 {state.max_steps}，强制结束，不绑定工具")
+                llm_no_tools = llm.bind_tools([])
+                response = await llm_no_tools.ainvoke(full_messages)
+            else:
+                info(f"[Agent] 调用带工具的 LLM...")
+                response = await llm_with_tools.ainvoke(full_messages)
+            
+            info(f"[Agent] LLM 调用成功！响应类型: {type(response).__name__}")
+            if hasattr(response, 'tool_calls') and response.tool_calls:
+                info(f"[Agent] 检测到工具调用: {[tc['name'] for tc in response.tool_calls]}")
 
-        # 返回状态更新（注意：不原地修改 state，返回字典让 LangGraph 处理
-        return {
-            "messages": [response],
-            "current_step": state.current_step + 1,
-            "llm_calls": state.llm_calls + 1
-        }
+            # 返回状态更新（注意：不原地修改 state，返回字典让 LangGraph 处理
+            return {
+                "messages": [response],
+                "current_step": state.current_step + 1,
+                "llm_calls": state.llm_calls + 1
+            }
+        
+        except Exception as e:
+            error(f"[Agent] ❌ 第 {state.current_step} 步推理出错: {e}")
+            import traceback
+            error(f"[Agent] 堆栈: {traceback.format_exc()}")
+            raise
 
     return agent_node