testing POC w/ state extension

libs/langgraph/langgraph/agent/__init__.py

@@ -81,17 +81,16 @@ def create_agent(
     )
 
     def model_request(state: AgentState) -> AgentState:
-        request = state.model_request or ModelRequest(
+        request = state.get("model_request") or ModelRequest(
             model=model,
             tools=default_tools,
             system_prompt=system_prompt,
             response_format=response_format,
-            messages=state.messages,
+            messages=state["messages"],
             tool_choice=None,
         )
 
         # prepare messages
-        print(request.system_prompt)
         if request.system_prompt:
             messages = [SystemMessage(request.system_prompt)] + request.messages
         else:
@@ -112,7 +111,7 @@ def create_agent(
                 parallel_tool_calls=False,
             )
             output = model_.invoke(messages)
-            if state.response is not None:
+            if state.get("response") is not None:
                 return {"messages": output, "response": None}
             else:
                 return {"messages": output}
@@ -125,7 +124,7 @@ def create_agent(
             graph.add_node(
                 f"{m.__class__.__name__}.before_model",
                 m.before_model,
-                input_schema=m.State,
+                input_schema=m.state_schema,
             )
         if m.__class__.modify_model_request is not AgentMiddleware.modify_model_request:
 
@@ -137,27 +136,27 @@ def create_agent(
                     tools=default_tools,
                     system_prompt=system_prompt,
                     response_format=response_format,
-                    messages=state.messages,
+                    messages=state["messages"],
                     tool_choice=None,
                 )
 
                 return {
                     "model_request": m.modify_model_request(
-                        state.model_request or default_model_request, state
+                        state.get("model_request") or default_model_request, state
                     )
                 }
 
             graph.add_node(
                 f"{m.__class__.__name__}.modify_model_request",
                 modify_model_request_node,
-                input_schema=m.State,
+                input_schema=m.state_schema,
             )
 
         if m.__class__.after_model is not AgentMiddleware.after_model:
             graph.add_node(
                 f"{m.__class__.__name__}.after_model",
                 m.after_model,
-                input_schema=m.State,
+                input_schema=m.state_schema,
             )
 
     # add start edge
@@ -258,9 +257,9 @@ def _resolve_jump(jump_to: JumpTo | None, first_node: str) -> str | None:
 
 def _make_model_to_tools_edge(first_node: str) -> Callable[[AgentState], str | None]:
     def model_to_tools(state: AgentState) -> str | None:
-        if state.jump_to:
-            return _resolve_jump(state.jump_to, first_node)
-        message = state.messages[-1]
+        if jump_to := state.get("jump_to"):
+            return _resolve_jump(jump_to, first_node)
+        message = state["messages"][-1]
         if isinstance(message, AIMessage) and message.tool_calls:
             return "tools"
 
@@ -273,7 +272,7 @@ def _make_tools_to_model_edge(
     tool_node: ToolNode, next_node: str
 ) -> Callable[[AgentState], str | None]:
     def tools_to_model(state: AgentState) -> str | None:
-        ai_message = [m for m in state.messages if isinstance(m, AIMessage)][-1]
+        ai_message = [m for m in state["messages"] if isinstance(m, AIMessage)][-1]
         if all(
             tool_node.tools_by_name[c["name"]].return_direct
             for c in ai_message.tool_calls
@@ -302,7 +301,7 @@ def _add_middleware_edge(
 
         def jump_edge(state: AgentState) -> str:
             return (
-                _resolve_jump(state.jump_to, model_destination) or default_destination
+                _resolve_jump(state.get("jump_to"), model_destination) or default_destination
             )
 
         destinations = [default_destination, END, "tools"]

libs/langgraph/langgraph/agent/middleware/poc_middleware.py

@@ -0,0 +1,52 @@
+from dataclasses import dataclass, field
+from typing import Annotated, Any, Dict, List, cast
+
+from langchain_core.messages import AIMessage
+from typing_extensions import Annotated
+from typing import ClassVar
+from operator import add
+
+from langgraph.agent import create_agent
+from langgraph.agent.types import AgentJump, AgentMiddleware, AgentState, AgentUpdate, ModelRequest
+
+class AcceptInput:
+    ...
+
+class ExposeOutput:
+    ...
+
+# other ideas
+# state_extensions: ClassVar[dict[str, type | type[Annotated]]] = {
+#         "int1": Annotated[int, add],
+#         "int2": Annotated[int, add, AcceptInput],
+#         "int3": Annotated[int, add, ExposeOutput],
+#         "int4": Annotated[int, add, AcceptInput, ExposeOutput],
+# }
+
+class State(AgentState):
+    int1: Annotated[int, add]
+    int2: Annotated[int, add, AcceptInput]
+    int3: Annotated[int, add, ExposeOutput]
+    int4: Annotated[int, add, AcceptInput, ExposeOutput]
+
+class StateModMidleware(AgentMiddleware[State]):
+    """Terminates after a specific tool is called N times."""
+
+    state_schema: type[State] = State
+
+    def __init__(self):
+        pass
+
+    def before_model(self, state: State) -> AgentUpdate | AgentJump | None:
+        return {"int1": 1, "int2": 1, "int3": 1, "int4": 1}
+
+    def modify_model_request(self, request: ModelRequest, state: State) -> ModelRequest:
+        return request
+
+agent = create_agent(
+    model="gpt-4o",
+    tools=[],
+    system_prompt="You are a helpful assistant.",
+    middleware=[StateModMidleware()],
+)
+

libs/langgraph/langgraph/agent/middleware/tool_calls.py

@@ -14,14 +14,6 @@ class ToolCallLimitMiddleware(AgentMiddleware):
     class State(AgentMiddleware.State):
         important: Annotated[dict[str, int], Input, Output] = field(default_factory=dict)
 
-    @dataclass
-    class InputState(AgentMiddleware.State):
-        important: dict[str, int]
-
-    @dataclass
-    class OutputState(AgentMiddleware.State):
-        important: dict[str, int]
-
     def __init__(self, tool_limits: dict[str, int]):
         self.tool_limits = tool_limits
 

libs/langgraph/langgraph/agent/types.py

@@ -2,13 +2,13 @@ from __future__ import annotations
 
 from collections.abc import Sequence
 from dataclasses import dataclass
-from typing import Annotated, Any, Literal
+from typing import Annotated, Any, Literal, TypeVar, Generic, ClassVar
 
 from langchain_core.language_models.chat_models import BaseChatModel
 from langchain_core.messages import AnyMessage
 from langchain_core.tools import BaseTool
 from pydantic import BaseModel
-from typing_extensions import TypedDict
+from typing_extensions import TypedDict, Required
 
 from langgraph.channels.ephemeral_value import EphemeralValue
 from langgraph.graph.message import Messages, add_messages
@@ -28,27 +28,29 @@ class ModelRequest:
     response_format: ResponseFormat | None
 
 
-@dataclass
-class AgentState:
+class AgentState(TypedDict, total=False):
+    # TODO: figure out Required/NotRequired wrapping annotated and still registering reducer properly
     messages: Annotated[list[AnyMessage], add_messages]
-    model_request: Annotated[ModelRequest | None, EphemeralValue] = None
-    jump_to: Annotated[JumpTo | None, EphemeralValue] = None
-    response: dict | None = None
+    model_request: Annotated[ModelRequest | None, EphemeralValue]
+    jump_to: Annotated[JumpTo | None, EphemeralValue]
+    response: dict
 
+StateT = TypeVar("StateT", bound=AgentState, default=AgentState, contravariant=True)
 
-class AgentMiddleware:
-    class State(AgentState):
+class AgentMiddleware(Generic[StateT]):
+
+    # TODO: I thought this should be a ClassVar[type[StateT]] but inherently class vars can't use type vars
+    # bc they're instance dependent
+    state_schema: type[StateT]
+    tools: list[BaseTool] = []
+
+    def before_model(self, state: StateT) -> AgentUpdate | AgentJump | None:
         pass
 
-    tools: list[BaseTool]
-
-    def before_model(self, state: State) -> AgentUpdate | AgentJump | None:
-        pass
-
-    def modify_model_request(self, request: ModelRequest, state: State) -> ModelRequest:
+    def modify_model_request(self, request: ModelRequest, state: StateT) -> ModelRequest:
         return request
 
-    def after_model(self, state: State) -> AgentUpdate | AgentJump | None:
+    def after_model(self, state: StateT) -> AgentUpdate | AgentJump | None:
         pass
 
 

test_poc.py

@@ -0,0 +1,56 @@
+from dataclasses import dataclass, field
+from typing import Annotated, Any, Dict, List, cast
+
+from langchain_core.messages import AIMessage, HumanMessage
+from typing_extensions import Annotated
+from typing import ClassVar
+from operator import add
+
+from langgraph.agent import create_agent
+from langgraph.agent.types import AgentJump, AgentMiddleware, AgentState, AgentUpdate, ModelRequest
+
+class AcceptInput:
+    ...
+
+class ExposeOutput:
+    ...
+
+# other ideas
+# state_extensions: ClassVar[dict[str, type | type[Annotated]]] = {
+#         "int1": Annotated[int, add],
+#         "int2": Annotated[int, add, AcceptInput],
+#         "int3": Annotated[int, add, ExposeOutput],
+#         "int4": Annotated[int, add, AcceptInput, ExposeOutput],
+# }
+
+class State(AgentState):
+    int1: Annotated[int, add]
+    int2: Annotated[int, AcceptInput, add]
+    int3: Annotated[int, ExposeOutput, add]
+    int4: Annotated[int, AcceptInput, ExposeOutput, add]
+
+class StateModMidleware(AgentMiddleware[State]):
+    """Terminates after a specific tool is called N times."""
+
+    state_schema: type[State] = State
+
+    def __init__(self):
+        pass
+
+    def before_model(self, state: State) -> AgentUpdate | AgentJump | None:
+        return {"int1": 1, "int2": 1, "int3": 1, "int4": 1}
+
+    def modify_model_request(self, request: ModelRequest, state: State) -> ModelRequest:
+        return request
+
+agent = create_agent(
+    model="gpt-4o",
+    tools=[],
+    system_prompt="You are a helpful assistant.",
+    # TODO: figure out invariance here
+    middleware=[StateModMidleware()],
+)
+agent = agent.compile()
+
+result = agent.invoke({"messages": [HumanMessage("What is 2+2?")]})
+print(result)