Update multi_agent.py

multi_agent.py

@@ -32,12 +32,13 @@ class MultyAgentSystem:
         )
         self.qwen_model = LiteLLMModel("groq/qwen-qwq-32b", **common)
         self.fallback_model = LiteLLMModel("groq/llama3-70b-8k", **common)
+        self.llama_model = LiteLLMModel("groq/llama-4-scout-17b-16e-instruct", **common)
 
         self.verification_limit = int(os.getenv("VERIFY_WORD_LIMIT", "75"))
 
         # --- Web agent definition ---
         self.web_agent = CodeAgent(
-            model=self.qwen_model,
+            model=self.llama_model,
             tools=[WebSearchTool(), VisitWebpageTool(), WikipediaSearchTool()],
             name="web_agent",
             description=(
@@ -59,7 +60,7 @@ class MultyAgentSystem:
 
         # --- Info agent definition ---
         self.info_agent = CodeAgent(
-            model=self.qwen_model,
+            model=self.llama_model,
             tools=[PythonInterpreterTool(), image_reasoning_tool],
             name="info_agent",
             description=(
@@ -83,7 +84,7 @@ class MultyAgentSystem:
         # The manager starts with the smaller Qwen model to minimize token usage
         # and only relies on DeepSeek when verifying critical answers.
         self.manager_agent = CodeAgent(
-            model=self.qwen_model,
+            model=self.llama_model,
             tools=[FinalAnswerTool()],
             managed_agents=[self.web_agent, self.info_agent],
             name="manager_agent",
@@ -116,7 +117,7 @@ class MultyAgentSystem:
 
     def run(self, question, high_stakes: bool = False, **kwargs):
         start_time = time.time()
-        print("Generating initial answer with Qwen-32B")
+        print("Generating initial answer with llama-4-scout")
         max_completion_tokens = kwargs.get("max_completion_tokens", 512)
         prompt_tokens = len(question.split())
         consume(prompt_tokens + max_completion_tokens)