fix: retrieve function_calling_llm from registered LLMs in CrewBase

src/crewai/project/crew_base.py

@@ -137,13 +137,11 @@ def CrewBase(cls: T) -> T:
                 all_functions, "is_cache_handler"
             )
             callbacks = self._filter_functions(all_functions, "is_callback")
-            agents = self._filter_functions(all_functions, "is_agent")
 
             for agent_name, agent_info in self.agents_config.items():
                 self._map_agent_variables(
                     agent_name,
                     agent_info,
-                    agents,
                     llms,
                     tool_functions,
                     cache_handler_functions,
@@ -154,7 +152,6 @@ def CrewBase(cls: T) -> T:
             self,
             agent_name: str,
             agent_info: Dict[str, Any],
-            agents: Dict[str, Callable],
             llms: Dict[str, Callable],
             tool_functions: Dict[str, Callable],
             cache_handler_functions: Dict[str, Callable],
@@ -172,9 +169,10 @@ def CrewBase(cls: T) -> T:
                 ]
 
             if function_calling_llm := agent_info.get("function_calling_llm"):
-                self.agents_config[agent_name]["function_calling_llm"] = agents[
-                    function_calling_llm
-                ]()
+                try:
+                    self.agents_config[agent_name]["function_calling_llm"] = llms[function_calling_llm]()
+                except KeyError:
+                    self.agents_config[agent_name]["function_calling_llm"] = function_calling_llm
 
             if step_callback := agent_info.get("step_callback"):
                 self.agents_config[agent_name]["step_callback"] = callbacks[

tests/config/agents.yaml

@@ -8,6 +8,7 @@ researcher:
     developments in {topic}. Known for your ability to find the most relevant
     information and present it in a clear and concise manner.
   verbose: true
+  function_calling_llm: "local_llm"
 
 reporting_analyst:
   role: >
@@ -19,3 +20,4 @@ reporting_analyst:
     your ability to turn complex data into clear and concise reports, making
     it easy for others to understand and act on the information you provide.
   verbose: true
+  function_calling_llm: "online_llm"

tests/project_test.py

@@ -31,6 +31,13 @@ class InternalCrew:
     agents_config = "config/agents.yaml"
     tasks_config = "config/tasks.yaml"
 
+    @llm
+    def local_llm(self):
+        return LLM(
+            model='openai/model_name',
+            api_key="None",
+            base_url="http://xxx.xxx.xxx.xxx:8000/v1")
+
     @agent
     def researcher(self):
         return Agent(config=self.agents_config["researcher"])
@@ -105,6 +112,20 @@ def test_task_name():
     ), "Custom task name is not being set as expected"
 
 
+def test_agent_function_calling_llm():
+    crew = InternalCrew()
+    llm = crew.local_llm()
+    obj_llm_agent = crew.researcher()
+    assert (
+        obj_llm_agent.function_calling_llm is llm
+    ), "agent's function_calling_llm is incorrect"
+
+    str_llm_agent = crew.reporting_analyst()
+    assert (
+        str_llm_agent.function_calling_llm.model == "online_llm"
+    ), "agent's function_calling_llm is incorrect"
+
+
 @pytest.mark.vcr(filter_headers=["authorization"])
 def test_before_kickoff_modification():
     crew = InternalCrew()