letta-nightly 0.6.48.dev20250407104216__py3-none-any.whl → 0.6.49.dev20250408030511__py3-none-any.whl

letta/schemas/message.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import copy
 import json
+import uuid
 import warnings
 from collections import OrderedDict
 from datetime import datetime, timezone
@@ -78,6 +79,7 @@ class MessageCreate(BaseModel):
         json_schema_extra=get_letta_message_content_union_str_json_schema(),
     )
     name: Optional[str] = Field(None, description="The name of the participant.")
+    otid: Optional[str] = Field(None, description="The offline threading id associated with this message")
 
     def model_dump(self, to_orm: bool = False, **kwargs) -> Dict[str, Any]:
         data = super().model_dump(**kwargs)
@@ -168,12 +170,17 @@ class Message(BaseMessage):
         json_message["created_at"] = self.created_at.isoformat()
         return json_message
 
+    @staticmethod
+    def generate_otid():
+        return str(uuid.uuid4())
+
     @staticmethod
     def to_letta_messages_from_list(
         messages: List[Message],
         use_assistant_message: bool = True,
         assistant_message_tool_name: str = DEFAULT_MESSAGE_TOOL,
         assistant_message_tool_kwarg: str = DEFAULT_MESSAGE_TOOL_KWARG,
+        reverse: bool = True,
     ) -> List[LettaMessage]:
         if use_assistant_message:
             message_ids_to_remove = []
@@ -203,6 +210,7 @@ class Message(BaseMessage):
                 use_assistant_message=use_assistant_message,
                 assistant_message_tool_name=assistant_message_tool_name,
                 assistant_message_tool_kwarg=assistant_message_tool_kwarg,
+                reverse=reverse,
             )
         ]
 
@@ -211,6 +219,7 @@ class Message(BaseMessage):
         use_assistant_message: bool = False,
         assistant_message_tool_name: str = DEFAULT_MESSAGE_TOOL,
         assistant_message_tool_kwarg: str = DEFAULT_MESSAGE_TOOL_KWARG,
+        reverse: bool = True,
     ) -> List[LettaMessage]:
         """Convert message object (in DB format) to the style used by the original Letta API"""
         messages = []
@@ -221,18 +230,21 @@ class Message(BaseMessage):
             if self.content:
                 # Check for ReACT-style COT inside of TextContent
                 if len(self.content) == 1 and isinstance(self.content[0], TextContent):
+                    otid = Message.generate_otid_from_id(self.id, len(messages))
                     messages.append(
                         ReasoningMessage(
                             id=self.id,
                             date=self.created_at,
                             reasoning=self.content[0].text,
                             name=self.name,
+                            otid=otid,
                         )
                     )
                 # Otherwise, we may have a list of multiple types
                 else:
                     # TODO we can probably collapse these two cases into a single loop
                     for content_part in self.content:
+                        otid = Message.generate_otid_from_id(self.id, len(messages))
                         if isinstance(content_part, TextContent):
                             # COT
                             messages.append(
@@ -241,6 +253,7 @@ class Message(BaseMessage):
                                     date=self.created_at,
                                     reasoning=content_part.text,
                                     name=self.name,
+                                    otid=otid,
                                 )
                             )
                         elif isinstance(content_part, ReasoningContent):
@@ -253,6 +266,7 @@ class Message(BaseMessage):
                                     source="reasoner_model",  # TODO do we want to tag like this?
                                     signature=content_part.signature,
                                     name=self.name,
+                                    otid=otid,
                                 )
                             )
                         elif isinstance(content_part, RedactedReasoningContent):
@@ -264,6 +278,7 @@ class Message(BaseMessage):
                                     state="redacted",
                                     hidden_reasoning=content_part.data,
                                     name=self.name,
+                                    otid=otid,
                                 )
                             )
                         else:
@@ -272,6 +287,7 @@ class Message(BaseMessage):
             if self.tool_calls is not None:
                 # This is type FunctionCall
                 for tool_call in self.tool_calls:
+                    otid = Message.generate_otid_from_id(self.id, len(messages))
                     # If we're supporting using assistant message,
                     # then we want to treat certain function calls as a special case
                     if use_assistant_message and tool_call.function.name == assistant_message_tool_name:
@@ -287,6 +303,7 @@ class Message(BaseMessage):
                                 date=self.created_at,
                                 content=message_string,
                                 name=self.name,
+                                otid=otid,
                             )
                         )
                     else:
@@ -300,6 +317,7 @@ class Message(BaseMessage):
                                     tool_call_id=tool_call.id,
                                 ),
                                 name=self.name,
+                                otid=otid,
                             )
                         )
         elif self.role == MessageRole.tool:
@@ -341,6 +359,7 @@ class Message(BaseMessage):
                     stdout=self.tool_returns[0].stdout if self.tool_returns else None,
                     stderr=self.tool_returns[0].stderr if self.tool_returns else None,
                     name=self.name,
+                    otid=self.id.replace("message-", ""),
                 )
             )
         elif self.role == MessageRole.user:
@@ -357,6 +376,7 @@ class Message(BaseMessage):
                     date=self.created_at,
                     content=message_str or text_content,
                     name=self.name,
+                    otid=self.otid,
                 )
             )
         elif self.role == MessageRole.system:
@@ -372,11 +392,15 @@ class Message(BaseMessage):
                     date=self.created_at,
                     content=text_content,
                     name=self.name,
+                    otid=self.otid,
                 )
             )
         else:
             raise ValueError(self.role)
 
+        if reverse:
+            messages.reverse()
+
         return messages
 
     @staticmethod
@@ -670,6 +694,9 @@ class Message(BaseMessage):
 
         def add_xml_tag(string: str, xml_tag: Optional[str]):
             # NOTE: Anthropic docs recommends using <thinking> tag when using CoT + tool use
+            if f"<{xml_tag}>" in string and f"</{xml_tag}>" in string:
+                # don't nest if tags already exist
+                return string
             return f"<{xml_tag}>{string}</{xml_tag}" if xml_tag else string
 
         if self.role == "system":
@@ -988,6 +1015,23 @@ class Message(BaseMessage):
 
         return cohere_message
 
+    @staticmethod
+    def generate_otid_from_id(message_id: str, index: int) -> str:
+        """
+        Convert message id to bits and change the list bit to the index
+        """
+        if not 0 <= index < 128:
+            raise ValueError("Index must be between 0 and 127")
+
+        message_uuid = message_id.replace("message-", "")
+        uuid_int = int(message_uuid.replace("-", ""), 16)
+
+        # Clear last 7 bits and set them to index; supports up to 128 unique indices
+        uuid_int = (uuid_int & ~0x7F) | (index & 0x7F)
+
+        hex_str = f"{uuid_int:032x}"
+        return f"{hex_str[:8]}-{hex_str[8:12]}-{hex_str[12:16]}-{hex_str[16:20]}-{hex_str[20:]}"
+
 
 class ToolReturn(BaseModel):
     status: Literal["success", "error"] = Field(..., description="The status of the tool call")

letta/schemas/tool.py

@@ -104,6 +104,9 @@ class Tool(BaseTool):
         elif self.tool_type in {ToolType.LETTA_MULTI_AGENT_CORE}:
             # If it's letta multi-agent tool, we also generate the json_schema on the fly here
             self.json_schema = get_json_schema_from_module(module_name=LETTA_MULTI_AGENT_TOOL_MODULE_NAME, function_name=self.name)
+        elif self.tool_type in {ToolType.LETTA_SLEEPTIME_CORE}:
+            # If it's letta sleeptime core tool, we generate the json_schema on the fly here
+            self.json_schema = get_json_schema_from_module(module_name=LETTA_CORE_TOOL_MODULE_NAME, function_name=self.name)
 
         # At this point, we need to validate that at least json_schema is populated
         if not self.json_schema:

letta/schemas/usage.py

@@ -23,3 +23,4 @@ class LettaUsageStatistics(BaseModel):
     step_count: int = Field(0, description="The number of steps taken by the agent.")
     # TODO: Optional for now. This field makes everyone's lives easier
     steps_messages: Optional[List[List[Message]]] = Field(None, description="The messages generated per step")
+    run_ids: Optional[List[str]] = Field(None, description="The background task run IDs associated with the agent interaction")

letta/server/db.py

@@ -59,11 +59,13 @@ if settings.letta_pg_uri_no_default:
     # create engine
     engine = create_engine(
         settings.letta_pg_uri,
+        # f"{settings.letta_pg_uri}?options=-c%20client_encoding=UTF8",
         pool_size=settings.pg_pool_size,
         max_overflow=settings.pg_max_overflow,
         pool_timeout=settings.pg_pool_timeout,
         pool_recycle=settings.pg_pool_recycle,
         echo=settings.pg_echo,
+        # connect_args={"client_encoding": "utf8"},
     )
 else:
     # TODO: don't rely on config storage

letta/server/rest_api/app.py

@@ -139,7 +139,7 @@ def create_application() -> "FastAPI":
 
     @app.on_event("startup")
     async def configure_executor():
-        print(f"Configured event loop executor with {settings.event_loop_threadpool_max_workers} workers.")
+        print(f"INFO:     Configured event loop executor with {settings.event_loop_threadpool_max_workers} workers.")
         loop = asyncio.get_running_loop()
         executor = concurrent.futures.ThreadPoolExecutor(max_workers=settings.event_loop_threadpool_max_workers)
         loop.set_default_executor(executor)

letta/server/rest_api/chat_completions_interface.py

@@ -155,7 +155,12 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
         return
 
     def process_chunk(
-        self, chunk: ChatCompletionChunkResponse, message_id: str, message_date: datetime, expect_reasoning_content: bool = False
+        self,
+        chunk: ChatCompletionChunkResponse,
+        message_id: str,
+        message_date: datetime,
+        expect_reasoning_content: bool = False,
+        message_index: int = 0,
     ) -> None:
         """
         Called externally with a ChatCompletionChunkResponse. Transforms
@@ -172,7 +177,7 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
         """
         return
 
-    def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None) -> None:
+    def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None, chunk_index: Optional[int] = None) -> None:
         """
         Handle LLM reasoning or internal monologue. Example usage: if you want
         to capture chain-of-thought for debugging in a non-streaming scenario.
@@ -186,7 +191,7 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
         """
         return
 
-    def function_message(self, msg: str, msg_obj: Optional[Message] = None) -> None:
+    def function_message(self, msg: str, msg_obj: Optional[Message] = None, chunk_index: Optional[int] = None) -> None:
         """
         Handle function-related log messages, typically of the form:
         It's a no-op by default.

letta/server/rest_api/interface.py

@@ -165,7 +165,7 @@ class QueuingInterface(AgentInterface):
             print(vars(msg_obj))
             print(msg_obj.created_at.isoformat())
 
-    def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None) -> None:
+    def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None, chunk_index: Optional[int] = None) -> None:
         """Handle the agent's internal monologue"""
         assert msg_obj is not None, "QueuingInterface requires msg_obj references for metadata"
         if self.debug:
@@ -209,7 +209,9 @@ class QueuingInterface(AgentInterface):
 
         self._queue_push(message_api=new_message, message_obj=msg_obj)
 
-    def function_message(self, msg: str, msg_obj: Optional[Message] = None, include_ran_messages: bool = False) -> None:
+    def function_message(
+        self, msg: str, msg_obj: Optional[Message] = None, include_ran_messages: bool = False, chunk_index: Optional[int] = None
+    ) -> None:
         """Handle the agent calling a function"""
         # TODO handle 'function' messages that indicate the start of a function call
         assert msg_obj is not None, "QueuingInterface requires msg_obj references for metadata"
@@ -466,6 +468,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
         # and `content` needs to be handled outside the interface
         expect_reasoning_content: bool = False,
         name: Optional[str] = None,
+        message_index: int = 0,
     ) -> Optional[Union[ReasoningMessage, ToolCallMessage, AssistantMessage]]:
         """
         Example data from non-streaming response looks like:
@@ -478,6 +481,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
         """
         choice = chunk.choices[0]
         message_delta = choice.delta
+        otid = Message.generate_otid_from_id(message_id, message_index)
 
         if (
             message_delta.content is None
@@ -499,6 +503,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                 signature=message_delta.reasoning_content_signature,
                 source="reasoner_model" if message_delta.reasoning_content_signature else "non_reasoner_model",
                 name=name,
+                otid=otid,
             )
         elif expect_reasoning_content and message_delta.redacted_reasoning_content is not None:
             processed_chunk = HiddenReasoningMessage(
@@ -507,6 +512,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                 hidden_reasoning=message_delta.redacted_reasoning_content,
                 state="redacted",
                 name=name,
+                otid=otid,
             )
         elif expect_reasoning_content and message_delta.content is not None:
             # "ignore" content if we expect reasoning content
@@ -534,6 +540,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                         tool_call_id=None,
                     ),
                     name=name,
+                    otid=otid,
                 )
 
             except json.JSONDecodeError as e:
@@ -564,6 +571,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                 date=message_date,
                 reasoning=message_delta.content,
                 name=name,
+                otid=otid,
             )
 
         # tool calls
@@ -612,7 +620,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                         # TODO: Assumes consistent state and that prev_content is subset of new_content
                         diff = new_content.replace(prev_content, "", 1)
                         self.current_json_parse_result = parsed_args
-                        processed_chunk = AssistantMessage(id=message_id, date=message_date, content=diff, name=name)
+                        processed_chunk = AssistantMessage(id=message_id, date=message_date, content=diff, name=name, otid=otid)
                     else:
                         return None
 
@@ -645,6 +653,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                 tool_call_id=tool_call_delta.get("id"),
                             ),
                             name=name,
+                            otid=otid,
                         )
 
             elif self.inner_thoughts_in_kwargs and tool_call.function:
@@ -681,6 +690,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             date=message_date,
                             reasoning=updates_inner_thoughts,
                             name=name,
+                            otid=otid,
                         )
                         # Additionally inner thoughts may stream back with a chunk of main JSON
                         # In that case, since we can only return a chunk at a time, we should buffer it
@@ -717,6 +727,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                         tool_call_id=self.function_id_buffer,
                                     ),
                                     name=name,
+                                    otid=otid,
                                 )
 
                             # Record what the last function name we flushed was
@@ -774,6 +785,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                         date=message_date,
                                         content=combined_chunk,
                                         name=name,
+                                        otid=otid,
                                     )
                                     # Store the ID of the tool call so allow skipping the corresponding response
                                     if self.function_id_buffer:
@@ -798,7 +810,9 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                         # TODO: Assumes consistent state and that prev_content is subset of new_content
                                         diff = new_content.replace(prev_content, "", 1)
                                         self.current_json_parse_result = parsed_args
-                                        processed_chunk = AssistantMessage(id=message_id, date=message_date, content=diff, name=name)
+                                        processed_chunk = AssistantMessage(
+                                            id=message_id, date=message_date, content=diff, name=name, otid=otid
+                                        )
                                     else:
                                         return None
 
@@ -823,6 +837,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                             tool_call_id=self.function_id_buffer,
                                         ),
                                         name=name,
+                                        otid=otid,
                                     )
                                     # clear buffer
                                     self.function_args_buffer = None
@@ -838,6 +853,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                             tool_call_id=self.function_id_buffer,
                                         ),
                                         name=name,
+                                        otid=otid,
                                     )
                                     self.function_id_buffer = None
 
@@ -967,6 +983,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             tool_call_id=tool_call_delta.get("id"),
                         ),
                         name=name,
+                        otid=otid,
                     )
 
         elif choice.finish_reason is not None:
@@ -1048,6 +1065,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
         message_date: datetime,
         expect_reasoning_content: bool = False,
         name: Optional[str] = None,
+        message_index: int = 0,
     ):
         """Process a streaming chunk from an OpenAI-compatible server.
 
@@ -1074,18 +1092,20 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                 message_date=message_date,
                 expect_reasoning_content=expect_reasoning_content,
                 name=name,
+                message_index=message_index,
             )
-
         if processed_chunk is None:
             return
 
         self._push_to_buffer(processed_chunk)
 
+        return processed_chunk.message_type
+
     def user_message(self, msg: str, msg_obj: Optional[Message] = None):
         """Letta receives a user message"""
         return
 
-    def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None):
+    def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None, chunk_index: Optional[int] = None):
         """Letta generates some internal monologue"""
         if not self.streaming_mode:
 
@@ -1102,6 +1122,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                     date=msg_obj.created_at,
                     reasoning=msg,
                     name=msg_obj.name,
+                    otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                 )
 
                 self._push_to_buffer(processed_chunk)
@@ -1113,6 +1134,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             date=msg_obj.created_at,
                             reasoning=content.text,
                             name=msg_obj.name,
+                            otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                         )
                     elif isinstance(content, ReasoningContent):
                         processed_chunk = ReasoningMessage(
@@ -1122,6 +1144,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             reasoning=content.reasoning,
                             signature=content.signature,
                             name=msg_obj.name,
+                            otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                         )
                     elif isinstance(content, RedactedReasoningContent):
                         processed_chunk = HiddenReasoningMessage(
@@ -1130,6 +1153,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             state="redacted",
                             hidden_reasoning=content.data,
                             name=msg_obj.name,
+                            otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                         )
 
                     self._push_to_buffer(processed_chunk)
@@ -1142,7 +1166,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
         # NOTE: this is a no-op, we handle this special case in function_message instead
         return
 
-    def function_message(self, msg: str, msg_obj: Optional[Message] = None):
+    def function_message(self, msg: str, msg_obj: Optional[Message] = None, chunk_index: Optional[int] = None):
         """Letta calls a function"""
 
         # TODO handle 'function' messages that indicate the start of a function call
@@ -1191,6 +1215,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                 date=msg_obj.created_at,
                                 content=func_args["message"],
                                 name=msg_obj.name,
+                                otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                             )
                             self._push_to_buffer(processed_chunk)
                         except Exception as e:
@@ -1214,6 +1239,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             date=msg_obj.created_at,
                             content=func_args[self.assistant_message_tool_kwarg],
                             name=msg_obj.name,
+                            otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                         )
                         # Store the ID of the tool call so allow skipping the corresponding response
                         self.prev_assistant_message_id = function_call.id
@@ -1227,6 +1253,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                 tool_call_id=function_call.id,
                             ),
                             name=msg_obj.name,
+                            otid=Message.generate_otid_from_id(msg_obj.id, chunk_index) if chunk_index is not None else None,
                         )
 
                     # processed_chunk = {
@@ -1267,6 +1294,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                     stdout=msg_obj.tool_returns[0].stdout if msg_obj.tool_returns else None,
                     stderr=msg_obj.tool_returns[0].stderr if msg_obj.tool_returns else None,
                     name=msg_obj.name,
+                    otid=Message.generate_otid_from_id(msg_obj.id, chunk_index),
                 )
 
         elif msg.startswith("Error: "):
@@ -1282,6 +1310,7 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                 stdout=msg_obj.tool_returns[0].stdout if msg_obj.tool_returns else None,
                 stderr=msg_obj.tool_returns[0].stderr if msg_obj.tool_returns else None,
                 name=msg_obj.name,
+                otid=Message.generate_otid_from_id(msg_obj.id, chunk_index),
             )
 
         else:

letta/server/rest_api/routers/v1/agents.py

@@ -8,6 +8,7 @@ from fastapi.responses import JSONResponse
 from marshmallow import ValidationError
 from pydantic import Field
 from sqlalchemy.exc import IntegrityError, OperationalError
+from starlette.responses import StreamingResponse
 
 from letta.agents.letta_agent import LettaAgent
 from letta.constants import DEFAULT_MESSAGE_TOOL, DEFAULT_MESSAGE_TOOL_KWARG
@@ -30,7 +31,6 @@ from letta.schemas.user import User
 from letta.serialize_schemas.pydantic_agent_schema import AgentSchema
 from letta.server.rest_api.utils import get_letta_server
 from letta.server.server import SyncServer
-from letta.settings import settings
 
 # These can be forward refs, but because Fastapi needs them at runtime the must be imported normally
 
@@ -130,6 +130,10 @@ async def import_agent_serialized(
         description="If set to True, existing tools can get their source code overwritten by the uploaded tool definitions. Note that Letta core tools can never be updated externally.",
     ),
     project_id: Optional[str] = Query(None, description="The project ID to associate the uploaded agent with."),
+    strip_messages: bool = Query(
+        False,
+        description="If set to True, strips all messages from the agent before importing.",
+    ),
 ):
     """
     Import a serialized agent file and recreate the agent in the system.
@@ -149,6 +153,7 @@ async def import_agent_serialized(
             append_copy_suffix=append_copy_suffix,
             override_existing_tools=override_existing_tools,
             project_id=project_id,
+            strip_messages=strip_messages,
         )
         return new_agent
 
@@ -585,8 +590,10 @@ async def send_message(
     This endpoint accepts a message from a user and processes it through the agent.
     """
     actor = server.user_manager.get_user_or_default(user_id=actor_id)
-    if settings.use_experimental:
-        logger.warning("USING EXPERIMENTAL!")
+    # TODO: This is redundant, remove soon
+    agent = server.agent_manager.get_agent_by_id(agent_id, actor)
+
+    if agent.llm_config.model_endpoint_type == "anthropic" and not agent.enable_sleeptime and not agent.multi_agent_group:
         experimental_agent = LettaAgent(
             agent_id=agent_id,
             message_manager=server.message_manager,
@@ -639,17 +646,38 @@ async def send_message_streaming(
     It will stream the steps of the response always, and stream the tokens if 'stream_tokens' is set to True.
     """
     actor = server.user_manager.get_user_or_default(user_id=actor_id)
-    result = await server.send_message_to_agent(
-        agent_id=agent_id,
-        actor=actor,
-        messages=request.messages,
-        stream_steps=True,
-        stream_tokens=request.stream_tokens,
-        # Support for AssistantMessage
-        use_assistant_message=request.use_assistant_message,
-        assistant_message_tool_name=request.assistant_message_tool_name,
-        assistant_message_tool_kwarg=request.assistant_message_tool_kwarg,
-    )
+    # TODO: This is redundant, remove soon
+    agent = server.agent_manager.get_agent_by_id(agent_id, actor)
+
+    if agent.llm_config.model_endpoint_type == "anthropic" and not agent.enable_sleeptime and not agent.multi_agent_group:
+        experimental_agent = LettaAgent(
+            agent_id=agent_id,
+            message_manager=server.message_manager,
+            agent_manager=server.agent_manager,
+            block_manager=server.block_manager,
+            passage_manager=server.passage_manager,
+            actor=actor,
+        )
+
+        messages = request.messages
+        content = messages[0].content[0].text if messages and not isinstance(messages[0].content, str) else messages[0].content
+        result = StreamingResponse(
+            experimental_agent.step_stream(UserMessage(content=content), max_steps=10, use_assistant_message=request.use_assistant_message),
+            media_type="text/event-stream",
+        )
+    else:
+        result = await server.send_message_to_agent(
+            agent_id=agent_id,
+            actor=actor,
+            messages=request.messages,
+            stream_steps=True,
+            stream_tokens=request.stream_tokens,
+            # Support for AssistantMessage
+            use_assistant_message=request.use_assistant_message,
+            assistant_message_tool_name=request.assistant_message_tool_name,
+            assistant_message_tool_kwarg=request.assistant_message_tool_kwarg,
+        )
+
     return result
 
 
@@ -665,31 +693,17 @@ async def process_message_background(
 ) -> None:
     """Background task to process the message and update job status."""
     try:
-        # TODO(matt) we should probably make this stream_steps and log each step as it progresses, so the job update GET can see the total steps so far + partial usage?
-        if settings.use_experimental:
-            logger.warning("USING EXPERIMENTAL!")
-            experimental_agent = LettaAgent(
-                agent_id=agent_id,
-                message_manager=server.message_manager,
-                agent_manager=server.agent_manager,
-                block_manager=server.block_manager,
-                passage_manager=server.passage_manager,
-                actor=actor,
-            )
-            content = messages[0].content[0].text if messages and not isinstance(messages[0].content, str) else messages[0].content
-            result = await experimental_agent.step(UserMessage(content=content), max_steps=10)
-        else:
-            result = await server.send_message_to_agent(
-                agent_id=agent_id,
-                actor=actor,
-                messages=messages,
-                stream_steps=False,  # NOTE(matt)
-                stream_tokens=False,
-                use_assistant_message=use_assistant_message,
-                assistant_message_tool_name=assistant_message_tool_name,
-                assistant_message_tool_kwarg=assistant_message_tool_kwarg,
-                metadata={"job_id": job_id},  # Pass job_id through metadata
-            )
+        result = await server.send_message_to_agent(
+            agent_id=agent_id,
+            actor=actor,
+            messages=messages,
+            stream_steps=False,  # NOTE(matt)
+            stream_tokens=False,
+            use_assistant_message=use_assistant_message,
+            assistant_message_tool_name=assistant_message_tool_name,
+            assistant_message_tool_kwarg=assistant_message_tool_kwarg,
+            metadata={"job_id": job_id},  # Pass job_id through metadata
+        )
 
         # Update job status to completed
         job_update = JobUpdate(