letta-nightly 0.6.43.dev20250319104146__py3-none-any.whl → 0.6.43.dev20250321104124__py3-none-any.whl

letta/agent.py

@@ -522,7 +522,7 @@ class Agent(BaseAgent):
                     openai_message_dict=response_message.model_dump(),
                 )
             )  # extend conversation with assistant's reply
-            self.logger.info(f"Function call message: {messages[-1]}")
+            self.logger.debug(f"Function call message: {messages[-1]}")
 
             nonnull_content = False
             if response_message.content:
@@ -537,7 +537,7 @@ class Agent(BaseAgent):
                 response_message.function_call if response_message.function_call is not None else response_message.tool_calls[0].function
             )
             function_name = function_call.name
-            self.logger.info(f"Request to call function {function_name} with tool_call_id: {tool_call_id}")
+            self.logger.debug(f"Request to call function {function_name} with tool_call_id: {tool_call_id}")
 
             # Failure case 1: function name is wrong (not in agent_state.tools)
             target_letta_tool = None

letta/agents/ephemeral_memory_agent.py

@@ -0,0 +1,114 @@
+from typing import AsyncGenerator, Dict, List
+
+import openai
+
+from letta.agents.base_agent import BaseAgent
+from letta.helpers.tool_execution_helper import enable_strict_mode
+from letta.orm.enums import ToolType
+from letta.schemas.agent import AgentState
+from letta.schemas.enums import MessageRole
+from letta.schemas.letta_message import UserMessage
+from letta.schemas.letta_message_content import TextContent
+from letta.schemas.message import Message
+from letta.schemas.openai.chat_completion_request import ChatCompletionRequest, Tool
+from letta.schemas.user import User
+from letta.services.agent_manager import AgentManager
+from letta.services.message_manager import MessageManager
+
+
+class EphemeralMemoryAgent(BaseAgent):
+    """
+    A stateless agent that helps with offline memory computations.
+
+    """
+
+    def __init__(
+        self,
+        agent_id: str,
+        openai_client: openai.AsyncClient,
+        message_manager: MessageManager,
+        agent_manager: AgentManager,
+        actor: User,
+    ):
+        super().__init__(
+            agent_id=agent_id,
+            openai_client=openai_client,
+            message_manager=message_manager,
+            agent_manager=agent_manager,
+            actor=actor,
+        )
+
+    async def step(self, input_message: UserMessage) -> List[Message]:
+        """
+        Synchronous method that takes a user's input text and returns a summary from OpenAI.
+        Returns a list of ephemeral Message objects containing both the user text and the assistant summary.
+        """
+        agent_state = self.agent_manager.get_agent_by_id(agent_id=self.agent_id, actor=self.actor)
+
+        input_message = self.pre_process_input_message(input_message=input_message)
+        request = self._build_openai_request([input_message], agent_state)
+
+        chat_completion = await self.openai_client.chat.completions.create(**request.model_dump(exclude_unset=True))
+
+        return [
+            Message(
+                role=MessageRole.assistant,
+                content=[TextContent(text=chat_completion.choices[0].message.content.strip())],
+            )
+        ]
+
+    def pre_process_input_message(self, input_message: UserMessage) -> Dict:
+        input_prompt_augmented = f"""
+        You are a memory recall agent whose job is to comb through a large set of messages and write relevant memories in relation to a user query.
+        Your response will directly populate a "memory block" called "human" that describes the user, that will be used to answer more questions in the future.
+        You should err on the side of being more verbose, and also try to *predict* the trajectory of the conversation, and pull memories or messages you think will be relevant to where the conversation is going.
+
+        Your response should include:
+        - A high level summary of the relevant events/timeline of the conversation relevant to the query
+        - Direct citations of quotes from the messages you used while creating the summary
+
+        Here is a history of the messages so far:
+
+        {self._format_messages_llm_friendly()}
+
+        This is the query:
+
+        "{input_message.content}"
+
+        Your response:
+        """
+
+        input_message.content = input_prompt_augmented
+        # print(input_prompt_augmented)
+        return input_message.model_dump()
+
+    def _format_messages_llm_friendly(self):
+        messages = self.message_manager.list_messages_for_agent(agent_id=self.agent_id, actor=self.actor)
+
+        llm_friendly_messages = [f"{m.role}: {m.content[0].text}" for m in messages if m.content and isinstance(m.content[0], TextContent)]
+        return "\n".join(llm_friendly_messages)
+
+    def _build_openai_request(self, openai_messages: List[Dict], agent_state: AgentState) -> ChatCompletionRequest:
+        openai_request = ChatCompletionRequest(
+            model=agent_state.llm_config.model,
+            messages=openai_messages,
+            # tools=self._build_tool_schemas(agent_state),
+            # tool_choice="auto",
+            user=self.actor.id,
+            max_completion_tokens=agent_state.llm_config.max_tokens,
+            temperature=agent_state.llm_config.temperature,
+            stream=False,
+        )
+        return openai_request
+
+    def _build_tool_schemas(self, agent_state: AgentState) -> List[Tool]:
+        # Only include memory tools
+        tools = [t for t in agent_state.tools if t.tool_type in {ToolType.LETTA_CORE, ToolType.LETTA_MEMORY_CORE}]
+
+        return [Tool(type="function", function=enable_strict_mode(t.json_schema)) for t in tools]
+
+    async def step_stream(self, input_message: UserMessage) -> AsyncGenerator[str, None]:
+        """
+        This agent is synchronous-only. If called in an async context, raise an error.
+        """
+        raise NotImplementedError("EphemeralMemoryAgent does not support async step.")

letta/agents/{low_latency_agent.py → voice_agent.py}

@@ -5,7 +5,7 @@ from typing import Any, AsyncGenerator, Dict, List, Tuple
 import openai
 
 from letta.agents.base_agent import BaseAgent
-from letta.agents.ephemeral_agent import EphemeralAgent
+from letta.agents.ephemeral_memory_agent import EphemeralMemoryAgent
 from letta.constants import NON_USER_MSG_PREFIX
 from letta.helpers.datetime_helpers import get_utc_time
 from letta.helpers.tool_execution_helper import (
@@ -42,13 +42,12 @@ from letta.services.helpers.agent_manager_helper import compile_system_message
 from letta.services.message_manager import MessageManager
 from letta.services.passage_manager import PassageManager
 from letta.services.summarizer.enums import SummarizationMode
-from letta.services.summarizer.summarizer import Summarizer
 from letta.utils import united_diff
 
 logger = get_logger(__name__)
 
 
-class LowLatencyAgent(BaseAgent):
+class VoiceAgent(BaseAgent):
     """
     A function-calling loop for streaming OpenAI responses with tool execution.
     This agent:
@@ -65,9 +64,9 @@ class LowLatencyAgent(BaseAgent):
         agent_manager: AgentManager,
         block_manager: BlockManager,
         actor: User,
+        message_buffer_limit: int,
+        message_buffer_min: int,
         summarization_mode: SummarizationMode = SummarizationMode.STATIC_MESSAGE_BUFFER,
-        message_buffer_limit: int = 10,
-        message_buffer_min: int = 4,
     ):
         super().__init__(
             agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
@@ -79,75 +78,78 @@ class LowLatencyAgent(BaseAgent):
         self.passage_manager = PassageManager()  # TODO: pass this in
         # TODO: This is not guaranteed to exist!
         self.summary_block_label = "human"
-        self.summarizer = Summarizer(
-            mode=summarization_mode,
-            summarizer_agent=EphemeralAgent(
-                agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
-            ),
-            message_buffer_limit=message_buffer_limit,
-            message_buffer_min=message_buffer_min,
-        )
+        # self.summarizer = Summarizer(
+        #     mode=summarization_mode,
+        #     summarizer_agent=EphemeralAgent(
+        #         agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
+        #     ),
+        #     message_buffer_limit=message_buffer_limit,
+        #     message_buffer_min=message_buffer_min,
+        # )
         self.message_buffer_limit = message_buffer_limit
-        self.message_buffer_min = message_buffer_min
+        # self.message_buffer_min = message_buffer_min
+        self.offline_memory_agent = EphemeralMemoryAgent(
+            agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
+        )
 
     async def step(self, input_message: UserMessage) -> List[Message]:
         raise NotImplementedError("LowLatencyAgent does not have a synchronous step implemented currently.")
 
     async def step_stream(self, input_message: UserMessage) -> AsyncGenerator[str, None]:
         """
-        Async generator that yields partial tokens as SSE events, handles tool calls,
-        and streams error messages if OpenAI API failures occur.
+        Main streaming loop that yields partial tokens.
+        Whenever we detect a tool call, we yield from _handle_ai_response as well.
         """
-        input_message = self.pre_process_input_message(input_message=input_message)
-        agent_state = self.agent_manager.get_agent_by_id(agent_id=self.agent_id, actor=self.actor)
+        input_message = self.pre_process_input_message(input_message)
+        agent_state = self.agent_manager.get_agent_by_id(self.agent_id, actor=self.actor)
         in_context_messages = self.message_manager.get_messages_by_ids(message_ids=agent_state.message_ids, actor=self.actor)
         letta_message_db_queue = [create_user_message(input_message=input_message, agent_id=agent_state.id, actor=self.actor)]
         in_memory_message_history = [input_message]
 
+        # TODO: Define max steps here
         while True:
-            # Constantly pull down and integrate memory blocks
-            in_context_messages = self._rebuild_memory(in_context_messages=in_context_messages, agent_state=agent_state)
-
-            # Convert Letta messages to OpenAI messages
+            # Rebuild memory each loop
+            in_context_messages = self._rebuild_memory(in_context_messages, agent_state)
             openai_messages = convert_letta_messages_to_openai(in_context_messages)
             openai_messages.extend(in_memory_message_history)
+
             request = self._build_openai_request(openai_messages, agent_state)
 
-            # Execute the request
             stream = await self.openai_client.chat.completions.create(**request.model_dump(exclude_unset=True))
             streaming_interface = OpenAIChatCompletionsStreamingInterface(stream_pre_execution_message=True)
 
-            async for sse in streaming_interface.process(stream):
-                yield sse
+            # 1) Yield partial tokens from OpenAI
+            async for sse_chunk in streaming_interface.process(stream):
+                yield sse_chunk
 
-            # Process the AI response (buffered messages, tool execution, etc.)
-            continue_execution = await self._handle_ai_response(
-                streaming_interface, agent_state, in_memory_message_history, letta_message_db_queue
+            # 2) Now handle the final AI response. This might yield more text (stalling, etc.)
+            should_continue = await self._handle_ai_response(
+                streaming_interface,
+                agent_state,
+                in_memory_message_history,
+                letta_message_db_queue,
             )
 
-            if not continue_execution:
+            if not should_continue:
                 break
 
-        # Rebuild context window
+        # Rebuild context window if desired
         await self._rebuild_context_window(in_context_messages, letta_message_db_queue, agent_state)
-
         yield "data: [DONE]\n\n"
 
     async def _handle_ai_response(
         self,
-        streaming_interface: OpenAIChatCompletionsStreamingInterface,
+        streaming_interface: "OpenAIChatCompletionsStreamingInterface",
         agent_state: AgentState,
         in_memory_message_history: List[Dict[str, Any]],
         letta_message_db_queue: List[Any],
     ) -> bool:
         """
-        Handles AI response processing, including buffering messages, detecting tool calls,
-        executing tools, and deciding whether to continue execution.
-
-        Returns:
-            bool: True if execution should continue, False if the step loop should terminate.
+        Now that streaming is done, handle the final AI response.
+        This might yield additional SSE tokens if we do stalling.
+        At the end, set self._continue_execution accordingly.
         """
-        # Handle assistant message buffering
+        # 1. If we have any leftover content from partial stream, store it as an assistant message
         if streaming_interface.content_buffer:
             content = "".join(streaming_interface.content_buffer)
             in_memory_message_history.append({"role": "assistant", "content": content})
@@ -160,82 +162,92 @@ class LowLatencyAgent(BaseAgent):
             )
             letta_message_db_queue.extend(assistant_msgs)
 
-        # Handle tool execution if a tool call occurred
+        # 2. If a tool call was requested, handle it
         if streaming_interface.tool_call_happened:
+            tool_call_name = streaming_interface.tool_call_name
+            tool_call_args_str = streaming_interface.tool_call_args_str or "{}"
             try:
-                tool_args = json.loads(streaming_interface.tool_call_args_str)
+                tool_args = json.loads(tool_call_args_str)
             except json.JSONDecodeError:
                 tool_args = {}
 
             tool_call_id = streaming_interface.tool_call_id or f"call_{uuid.uuid4().hex[:8]}"
-
             assistant_tool_call_msg = AssistantMessage(
                 content=None,
                 tool_calls=[
                     ToolCall(
                         id=tool_call_id,
                         function=ToolCallFunction(
-                            name=streaming_interface.tool_call_name,
-                            arguments=streaming_interface.tool_call_args_str,
+                            name=tool_call_name,
+                            arguments=tool_call_args_str,
                         ),
                     )
                 ],
             )
             in_memory_message_history.append(assistant_tool_call_msg.model_dump())
 
-            tool_result, function_call_success = await self._execute_tool(
-                tool_name=streaming_interface.tool_call_name,
+            tool_result, success_flag = await self._execute_tool(
+                tool_name=tool_call_name,
                 tool_args=tool_args,
                 agent_state=agent_state,
             )
 
-            tool_message = ToolMessage(content=json.dumps({"result": tool_result}), tool_call_id=tool_call_id)
+            # 3. Provide function_call response back into the conversation
+            tool_message = ToolMessage(
+                content=json.dumps({"result": tool_result}),
+                tool_call_id=tool_call_id,
+            )
             in_memory_message_history.append(tool_message.model_dump())
 
+            # 4. Insert heartbeat message for follow-up
             heartbeat_user_message = UserMessage(
                 content=f"{NON_USER_MSG_PREFIX} Tool finished executing. Summarize the result for the user."
             )
             in_memory_message_history.append(heartbeat_user_message.model_dump())
 
+            # 5. Also store in DB
             tool_call_messages = create_tool_call_messages_from_openai_response(
                 agent_id=agent_state.id,
                 model=agent_state.llm_config.model,
-                function_name=streaming_interface.tool_call_name,
+                function_name=tool_call_name,
                 function_arguments=tool_args,
                 tool_call_id=tool_call_id,
-                function_call_success=function_call_success,
+                function_call_success=success_flag,
                 function_response=tool_result,
                 actor=self.actor,
                 add_heartbeat_request_system_message=True,
             )
             letta_message_db_queue.extend(tool_call_messages)
 
-            # Continue execution by restarting the loop with updated context
+            # Because we have new data, we want to continue the while-loop in `step_stream`
             return True
-
-        # Exit the loop if finish_reason_stop or no tool call occurred
-        return not streaming_interface.finish_reason_stop
+        else:
+            # If we got here, there's no tool call. If finish_reason_stop => done
+            return not streaming_interface.finish_reason_stop
 
     async def _rebuild_context_window(
         self, in_context_messages: List[Message], letta_message_db_queue: List[Message], agent_state: AgentState
     ) -> None:
         new_letta_messages = self.message_manager.create_many_messages(letta_message_db_queue, actor=self.actor)
+        new_in_context_messages = in_context_messages + new_letta_messages
 
-        # TODO: Make this more general and configurable, less brittle
-        target_block = next(b for b in agent_state.memory.blocks if b.label == self.summary_block_label)
-        previous_summary = self.block_manager.get_block_by_id(block_id=target_block.id, actor=self.actor).value
-        new_in_context_messages, summary_str, updated = await self.summarizer.summarize(
-            in_context_messages=in_context_messages, new_letta_messages=new_letta_messages, previous_summary=previous_summary
-        )
-
-        if updated:
-            self.block_manager.update_block(block_id=target_block.id, block_update=BlockUpdate(value=summary_str), actor=self.actor)
+        if len(new_in_context_messages) > self.message_buffer_limit:
+            cutoff = len(new_in_context_messages) - self.message_buffer_limit
+            new_in_context_messages = [new_in_context_messages[0]] + new_in_context_messages[cutoff:]
 
         self.agent_manager.set_in_context_messages(
             agent_id=self.agent_id, message_ids=[m.id for m in new_in_context_messages], actor=self.actor
         )
 
     def _rebuild_memory(self, in_context_messages: List[Message], agent_state: AgentState) -> List[Message]:
+        # Refresh memory
+        # TODO: This only happens for the summary block
+        # TODO: We want to extend this refresh to be general, and stick it in agent_manager
+        for i, b in enumerate(agent_state.memory.blocks):
+            if b.label == self.summary_block_label:
+                agent_state.memory.blocks[i] = self.block_manager.get_block_by_id(block_id=b.id, actor=self.actor)
+                break
+
         # TODO: This is a pretty brittle pattern established all over our code, need to get rid of this
         curr_system_message = in_context_messages[0]
         curr_memory_str = agent_state.memory.compile()
@@ -249,8 +261,8 @@ class LowLatencyAgent(BaseAgent):
 
         memory_edit_timestamp = get_utc_time()
 
-        num_messages = self.message_manager.size(actor=actor, agent_id=agent_id)
-        num_archival_memories = self.passage_manager.size(actor=actor, agent_id=agent_id)
+        num_messages = self.message_manager.size(actor=self.actor, agent_id=agent_state.id)
+        num_archival_memories = self.passage_manager.size(actor=self.actor, agent_id=agent_state.id)
 
         new_system_message_str = compile_system_message(
             system_prompt=agent_state.system,
@@ -296,8 +308,37 @@ class LowLatencyAgent(BaseAgent):
         else:
             tools = agent_state.tools
 
+        # Special tool state
+        recall_memory_utterance_description = (
+            "A lengthier message to be uttered while your memories of the current conversation are being re-contextualized."
+            "You should stall naturally and show the user you're thinking hard. The main thing is to not leave the user in silence."
+            "You MUST also include punctuation at the end of this message."
+        )
+        recall_memory_json = Tool(
+            type="function",
+            function=enable_strict_mode(
+                add_pre_execution_message(
+                    {
+                        "name": "recall_memory",
+                        "description": "Retrieve relevant information from memory based on a given query. Use when you don't remember the answer to a question.",
+                        "parameters": {
+                            "type": "object",
+                            "properties": {
+                                "query": {
+                                    "type": "string",
+                                    "description": "A description of what the model is trying to recall from memory.",
+                                }
+                            },
+                            "required": ["query"],
+                        },
+                    },
+                    description=recall_memory_utterance_description,
+                )
+            ),
+        )
+
         # TODO: Customize whether or not to have heartbeats, pre_exec_message, etc.
-        return [
+        return [recall_memory_json] + [
             Tool(type="function", function=enable_strict_mode(add_pre_execution_message(remove_request_heartbeat(t.json_schema))))
             for t in tools
         ]
@@ -306,19 +347,32 @@ class LowLatencyAgent(BaseAgent):
         """
         Executes a tool and returns (result, success_flag).
         """
-        target_tool = next((x for x in agent_state.tools if x.name == tool_name), None)
-        if not target_tool:
-            return f"Tool not found: {tool_name}", False
+        # Special memory case
+        if tool_name == "recall_memory":
+            # TODO: Make this safe
+            await self._recall_memory(tool_args["query"], agent_state)
+            return f"Successfully recalled memory and populated {self.summary_block_label} block.", True
+        else:
+            target_tool = next((x for x in agent_state.tools if x.name == tool_name), None)
+            if not target_tool:
+                return f"Tool not found: {tool_name}", False
 
-        try:
-            tool_result, _ = execute_external_tool(
-                agent_state=agent_state,
-                function_name=tool_name,
-                function_args=tool_args,
-                target_letta_tool=target_tool,
-                actor=self.actor,
-                allow_agent_state_modifications=False,
-            )
-            return tool_result, True
-        except Exception as e:
-            return f"Failed to call tool. Error: {e}", False
+            try:
+                tool_result, _ = execute_external_tool(
+                    agent_state=agent_state,
+                    function_name=tool_name,
+                    function_args=tool_args,
+                    target_letta_tool=target_tool,
+                    actor=self.actor,
+                    allow_agent_state_modifications=False,
+                )
+                return tool_result, True
+            except Exception as e:
+                return f"Failed to call tool. Error: {e}", False
+
+    async def _recall_memory(self, query, agent_state: AgentState) -> None:
+        results = await self.offline_memory_agent.step(UserMessage(content=query))
+        target_block = next(b for b in agent_state.memory.blocks if b.label == self.summary_block_label)
+        self.block_manager.update_block(
+            block_id=target_block.id, block_update=BlockUpdate(value=results[0].content[0].text), actor=self.actor
+        )

letta/client/client.py

@@ -2937,7 +2937,6 @@ class LocalClient(AbstractClient):
 
         Args:
             func (callable): The function to create a tool for.
-            name: (str): Name of the tool (must be unique per-user.)
             tags (Optional[List[str]], optional): Tags for the tool. Defaults to None.
             description (str, optional): The description.
             return_char_limit (int): The character limit for the tool's return value. Defaults to FUNCTION_RETURN_CHAR_LIMIT.
@@ -2950,6 +2949,7 @@ class LocalClient(AbstractClient):
         # parse source code/schema
         source_code = parse_source_code(func)
         source_type = "python"
+        name = func.__name__  # Initialize name using function's __name__
         if not tags:
             tags = []
 

letta/embeddings.py

@@ -235,7 +235,9 @@ def embedding_model(config: EmbeddingConfig, user_id: Optional[uuid.UUID] = None
 
     if endpoint_type == "openai":
         return OpenAIEmbeddings(
-            api_key=model_settings.openai_api_key, model=config.embedding_model, base_url=model_settings.openai_api_base
+            api_key=model_settings.openai_api_key,
+            model=config.embedding_model,
+            base_url=model_settings.openai_api_base,
         )
 
     elif endpoint_type == "azure":
@@ -246,19 +248,6 @@ def embedding_model(config: EmbeddingConfig, user_id: Optional[uuid.UUID] = None
                 model_settings.azure_api_version is not None,
             ]
         )
-        # from llama_index.embeddings.azure_openai import AzureOpenAIEmbedding
-
-        ## https://learn.microsoft.com/en-us/azure/ai-services/openai/reference#embeddings
-        # model = "text-embedding-ada-002"
-        # deployment = credentials.azure_embedding_deployment if credentials.azure_embedding_deployment is not None else model
-        # return AzureOpenAIEmbedding(
-        #    model=model,
-        #    deployment_name=deployment,
-        #    api_key=credentials.azure_key,
-        #    azure_endpoint=credentials.azure_endpoint,
-        #    api_version=credentials.azure_version,
-        # )
-
         return AzureOpenAIEmbedding(
             api_endpoint=model_settings.azure_base_url,
             api_key=model_settings.azure_api_key,

letta/functions/function_sets/multi_agent.py

@@ -9,6 +9,8 @@ from letta.functions.helpers import (
 )
 from letta.schemas.enums import MessageRole
 from letta.schemas.message import MessageCreate
+from letta.server.rest_api.utils import get_letta_server
+from letta.utils import log_telemetry
 
 if TYPE_CHECKING:
     from letta.agent import Agent
@@ -85,8 +87,51 @@ def send_message_to_agents_matching_tags(self: "Agent", message: str, match_all:
         response corresponds to a single agent. Agents that do not respond will not have an entry
         in the returned list.
     """
+    log_telemetry(
+        self.logger,
+        "_send_message_to_agents_matching_tags_async start",
+        message=message,
+        match_all=match_all,
+        match_some=match_some,
+    )
+    server = get_letta_server()
+
+    augmented_message = (
+        f"[Incoming message from agent with ID '{self.agent_state.id}' - to reply to this message, "
+        f"make sure to use the 'send_message' at the end, and the system will notify the sender of your response] "
+        f"{message}"
+    )
 
-    return asyncio.run(_send_message_to_agents_matching_tags_async(self, message, match_all, match_some))
+    # Retrieve up to 100 matching agents
+    log_telemetry(
+        self.logger,
+        "_send_message_to_agents_matching_tags_async listing agents start",
+        message=message,
+        match_all=match_all,
+        match_some=match_some,
+    )
+    matching_agents = server.agent_manager.list_agents_matching_tags(actor=self.user, match_all=match_all, match_some=match_some)
+
+    log_telemetry(
+        self.logger,
+        "_send_message_to_agents_matching_tags_async  listing agents finish",
+        message=message,
+        match_all=match_all,
+        match_some=match_some,
+    )
+
+    # Create a system message
+    messages = [MessageCreate(role=MessageRole.system, content=augmented_message, name=self.agent_state.name)]
+
+    result = asyncio.run(_send_message_to_agents_matching_tags_async(self, server, messages, matching_agents))
+    log_telemetry(
+        self.logger,
+        "_send_message_to_agents_matching_tags_async finish",
+        messages=message,
+        match_all=match_all,
+        match_some=match_some,
+    )
+    return result
 
 
 def send_message_to_all_agents_in_group(self: "Agent", message: str) -> List[str]: