feat: add reasoning toggle in agents routes (#3705)

letta/schemas/agent.py

@@ -212,6 +212,7 @@ class CreateAgent(BaseModel, validate_assignment=True):  #
         None, description="The maximum number of tokens to generate for reasoning step. If not set, the model will use its default value."
     )
     enable_reasoner: Optional[bool] = Field(False, description="Whether to enable internal extended thinking step for a reasoner model.")
+    reasoning: Optional[bool] = Field(None, description="Whether to enable reasoning for this agent.")
     from_template: Optional[str] = Field(None, description="The template id used to configure the agent")
     template: bool = Field(False, description="Whether the agent is a template")
     project: Optional[str] = Field(
@@ -335,6 +336,7 @@ class UpdateAgent(BaseModel):
     embedding: Optional[str] = Field(
         None, description="The embedding configuration handle used by the agent, specified in the format provider/model-name."
     )
+    reasoning: Optional[bool] = Field(None, description="Whether to enable reasoning for this agent.")
     enable_sleeptime: Optional[bool] = Field(None, description="If set to True, memory management will move to a background agent thread.")
     response_format: Optional[ResponseFormatUnion] = Field(None, description="The response format for the agent.")
     last_run_completion: Optional[datetime] = Field(None, description="The timestamp when the agent last completed a run.")

letta/schemas/llm_config.py

@@ -183,3 +183,30 @@ class LLMConfig(BaseModel):
             + (f" [type={self.model_endpoint_type}]" if self.model_endpoint_type else "")
             + (f" [ip={self.model_endpoint}]" if self.model_endpoint else "")
         )
+
+    @classmethod
+    def apply_reasoning_setting_to_config(cls, config: "LLMConfig", reasoning: bool):
+        if reasoning:
+            if (
+                config.model_endpoint_type == "anthropic"
+                and ("claude-opus-4" in config.model or "claude-sonnet-4" in config.model or "claude-3-7-sonnet" in config.model)
+            ) or (
+                config.model_endpoint_type == "google_vertex" and ("gemini-2.5-flash" in config.model or "gemini-2.0-pro" in config.model)
+            ):
+                config.put_inner_thoughts_in_kwargs = False
+                config.enable_reasoner = True
+                if config.max_reasoning_tokens == 0:
+                    config.max_reasoning_tokens = 1024
+            elif config.model_endpoint_type == "openai" and (
+                config.model.startswith("o1") or config.model.startswith("o3") or config.model.startswith("o4")
+            ):
+                config.put_inner_thoughts_in_kwargs = True
+                config.enable_reasoner = True
+                if config.reasoning_effort is None:
+                    config.reasoning_effort = "medium"
+            else:
+                config.put_inner_thoughts_in_kwargs = True
+                config.enable_reasoner = False
+        else:
+            config.put_inner_thoughts_in_kwargs = False
+            config.enable_reasoner = False

letta/services/agent_manager.py

@@ -51,6 +51,7 @@ from letta.schemas.enums import ProviderType, ToolType
 from letta.schemas.file import FileMetadata as PydanticFileMetadata
 from letta.schemas.group import Group as PydanticGroup
 from letta.schemas.group import ManagerType
+from letta.schemas.llm_config import LLMConfig
 from letta.schemas.memory import ContextWindowOverview, Memory
 from letta.schemas.message import Message
 from letta.schemas.message import Message as PydanticMessage
@@ -452,6 +453,9 @@ class AgentManager:
         if not agent_create.llm_config or not agent_create.embedding_config:
             raise ValueError("llm_config and embedding_config are required")
 
+        if agent_create.reasoning is not None:
+            agent_create.llm_config = LLMConfig.apply_reasoning_setting_to_config(agent_create.llm_config, agent_create.reasoning)
+
         # blocks
         block_ids = list(agent_create.block_ids or [])
         if agent_create.memory_blocks:
@@ -877,6 +881,10 @@ class AgentManager:
             agent.updated_at = datetime.now(timezone.utc)
             agent.last_updated_by_id = actor.id
 
+            if agent_update.reasoning is not None:
+                llm_config = agent_update.llm_config or agent.llm_config
+                agent_update.llm_config = LLMConfig.apply_reasoning_setting_to_config(llm_config, agent_update.reasoning)
+
             scalar_updates = {
                 "name": agent_update.name,
                 "system": agent_update.system,