Added an agent description field distinct from the system_message. (m…

…icrosoft#736)

* Added an agent description field that can be distinct from the system_message, and be used to for orchestration (e.g., GroupChatManager, etc.)

* Added debugging.

* Moved default descriptions to constants.

* Fixed conditions under which the assistant uses the default description.

* Removed debugging.

* Updated GroupChat prompt.

* Re-added debugging.

* Removed double [[ ]].

* Another update to GroupSelection prompt.

* Changed 'people' to 'participants' since agents are not people.

* Changed 'role' to 'name'

* Removed debugging statements.

* Restored the default prompt. Created a contrib class with new prompt.

* Fixed documentation.

* Removed broken link.

* Fixed a warning message.

* Removed GroupChatModerator contrib. Will re-add in another PR

* Resolving comment.

---------

Co-authored-by: Chi Wang <wang.chi@microsoft.com>

autogen/agentchat/assistant_agent.py

@@ -27,6 +27,8 @@ class AssistantAgent(ConversableAgent):
 Reply "TERMINATE" in the end when everything is done.
     """
 
+    DEFAULT_DESCRIPTION = "A helpful and general-purpose AI assistant that has strong language skills, Python skills, and Linux command line skills."
+
     def __init__(
         self,
         name: str,
@@ -36,6 +38,7 @@ def __init__(
         max_consecutive_auto_reply: Optional[int] = None,
         human_input_mode: Optional[str] = "NEVER",
         code_execution_config: Optional[Union[Dict, Literal[False]]] = False,
+        description: Optional[str] = None,
         **kwargs,
     ):
         """
@@ -63,5 +66,12 @@ def __init__(
             human_input_mode,
             code_execution_config=code_execution_config,
             llm_config=llm_config,
+            description=description,
             **kwargs,
         )
+
+        # Update the provided desciption if None, and we are using the default system_message,
+        # then use the default description.
+        if description is None:
+            if system_message == self.DEFAULT_SYSTEM_MESSAGE:
+                self.description = self.DEFAULT_DESCRIPTION

autogen/agentchat/conversable_agent.py

@@ -54,6 +54,7 @@ def __init__(
         code_execution_config: Optional[Union[Dict, Literal[False]]] = None,
         llm_config: Optional[Union[Dict, Literal[False]]] = None,
         default_auto_reply: Optional[Union[str, Dict, None]] = "",
+        description: Optional[str] = None,
     ):
         """
         Args:
@@ -95,11 +96,14 @@ def __init__(
                 for available options.
                 To disable llm-based auto reply, set to False.
             default_auto_reply (str or dict or None): default auto reply when no code execution or llm-based reply is generated.
+            description (str): a short description of the agent. This description is used by other agents
+                (e.g. the GroupChatManager) to decide when to call upon this agent. (Default: system_message)
         """
         super().__init__(name)
         # a dictionary of conversations, default value is list
         self._oai_messages = defaultdict(list)
         self._oai_system_message = [{"content": system_message, "role": "system"}]
+        self.description = description if description is not None else system_message
         self._is_termination_msg = (
             is_termination_msg
             if is_termination_msg is not None

autogen/agentchat/groupchat.py

@@ -80,13 +80,17 @@ def next_agent(self, agent: Agent, agents: List[Agent]) -> Agent:
                     return self.agents[(offset + i) % len(self.agents)]
 
     def select_speaker_msg(self, agents: List[Agent]) -> str:
-        """Return the message for selecting the next speaker."""
+        """Return the system message for selecting the next speaker. This is always the *first* message in the context."""
         return f"""You are in a role play game. The following roles are available:
 {self._participant_roles(agents)}.
 
 Read the following conversation.
 Then select the next role from {[agent.name for agent in agents]} to play. Only return the role."""
 
+    def select_speaker_prompt(self, agents: List[Agent]) -> str:
+        """Return the floating system prompt selecting the next speaker. This is always the *last* message in the context."""
+        return f"Read the above conversation. Then select the next role from {[agent.name for agent in agents]} to play. Only return the role."
+
     def manual_select_speaker(self, agents: List[Agent]) -> Union[Agent, None]:
         """Manually select the next speaker."""
 
@@ -170,15 +174,9 @@ def select_speaker(self, last_speaker: Agent, selector: ConversableAgent):
 
         # auto speaker selection
         selector.update_system_message(self.select_speaker_msg(agents))
-        final, name = selector.generate_oai_reply(
-            self.messages
-            + [
-                {
-                    "role": "system",
-                    "content": f"Read the above conversation. Then select the next role from {[agent.name for agent in agents]} to play. Only return the role.",
-                }
-            ]
-        )
+        context = self.messages + [{"role": "system", "content": self.select_speaker_prompt(agents)}]
+        final, name = selector.generate_oai_reply(context)
+
         if not final:
             # the LLM client is None, thus no reply is generated. Use round robin instead.
             return self.next_agent(last_speaker, agents)
@@ -205,11 +203,11 @@ def _participant_roles(self, agents: List[Agent] = None) -> str:
 
         roles = []
         for agent in agents:
-            if content_str(agent.system_message).strip() == "":
+            if agent.description.strip() == "":
                 logger.warning(
-                    f"The agent '{agent.name}' has an empty system_message, and may not work well with GroupChat."
+                    f"The agent '{agent.name}' has an empty description, and may not work well with GroupChat."
                 )
-            roles.append(f"{agent.name}: {agent.system_message}")
+            roles.append(f"{agent.name}: {agent.description}".strip())
         return "\n".join(roles)
 
     def _mentioned_agents(self, message_content: Union[str, List], agents: List[Agent]) -> Dict:

autogen/agentchat/user_proxy_agent.py

@@ -16,6 +16,13 @@ class UserProxyAgent(ConversableAgent):
     To customize the initial message when a conversation starts, override `generate_init_message` method.
     """
 
+    # Default UserProxyAgent.description values, based on human_input_mode
+    DEFAULT_USER_PROXY_AGENT_DESCRIPTIONS = {
+        "ALWAYS": "An attentive HUMAN user who can answer questions about the task, and can perform tasks such as running Python code or inputting command line commands at a Linux terminal and reporting back the execution results.",
+        "TERMINATE": "A user that can run Python code or input command line commands at a Linux terminal and report back the execution results.",
+        "NEVER": "A user that can run Python code or input command line commands at a Linux terminal and report back the execution results.",
+    }
+
     def __init__(
         self,
         name: str,
@@ -27,6 +34,7 @@ def __init__(
         default_auto_reply: Optional[Union[str, Dict, None]] = "",
         llm_config: Optional[Union[Dict, Literal[False]]] = False,
         system_message: Optional[Union[str, List]] = "",
+        description: Optional[str] = None,
     ):
         """
         Args:
@@ -69,15 +77,20 @@ def __init__(
                 Default to false, which disables llm-based auto reply.
             system_message (str or List): system message for ChatCompletion inference.
                 Only used when llm_config is not False. Use it to reprogram the agent.
+            description (str): a short description of the agent. This description is used by other agents
+                (e.g. the GroupChatManager) to decide when to call upon this agent. (Default: system_message)
         """
         super().__init__(
-            name,
-            system_message,
-            is_termination_msg,
-            max_consecutive_auto_reply,
-            human_input_mode,
-            function_map,
-            code_execution_config,
-            llm_config,
-            default_auto_reply,
+            name=name,
+            system_message=system_message,
+            is_termination_msg=is_termination_msg,
+            max_consecutive_auto_reply=max_consecutive_auto_reply,
+            human_input_mode=human_input_mode,
+            function_map=function_map,
+            code_execution_config=code_execution_config,
+            llm_config=llm_config,
+            default_auto_reply=default_auto_reply,
+            description=description
+            if description is not None
+            else self.DEFAULT_USER_PROXY_AGENT_DESCRIPTIONS[human_input_mode],
         )

test/agentchat/test_conversable_agent.py

@@ -275,6 +275,27 @@ def test_conversable_agent():
     with pytest.raises(KeyError):
         dummy_agent_1.last_message(dummy_agent_3)
 
+    # Check the description field
+    assert dummy_agent_1.description != dummy_agent_1.system_message
+    assert dummy_agent_2.description == dummy_agent_2.system_message
+
+    dummy_agent_4 = ConversableAgent(
+        name="dummy_agent_4",
+        system_message="The fourth dummy agent used for testing.",
+        llm_config=False,
+        human_input_mode="TERMINATE",
+    )
+    assert dummy_agent_4.description == "The fourth dummy agent used for testing."  # Same as system message
+
+    dummy_agent_5 = ConversableAgent(
+        name="dummy_agent_5",
+        system_message="",
+        description="The fifth dummy agent used for testing.",
+        llm_config=False,
+        human_input_mode="TERMINATE",
+    )
+    assert dummy_agent_5.description == "The fifth dummy agent used for testing."  # Same as system message
+
 
 def test_generate_reply():
     def add_num(num_to_be_added):