Langgraph

是什么

LangGraph 是用状态图（State Graph）来描述多步智能体流程的编排框架。 用它的效果是：复杂的决策、分支、循环不再藏在 if-else 里，而是变成一张可视化、可单步调试的图。

怎么用

1. 先把业务流程拆成节点（一个节点 = 一个明确职责），让每一步的输入输出都有契约。
2. 用条件边（Conditional Edge）描述分支与循环，让回退与重试逻辑显式而非散落。
3. 把全局状态（State）建模成 TypedDict 或 Pydantic 模型，让每一次状态更新都可追踪。
4. 在关键节点接入人工审核断点（Human-in-the-loop），让高风险动作有人类签字才放行。
5. 用图遍历日志做事后回放，让失败案例能复现、能归因、能写进回归测试。

架构图

flowchart LR
  起点 --> 决策节点
  决策节点 --> 工具调用
  决策节点 --> 人工审核
  工具调用 --> 状态更新
  人工审核 --> 状态更新
  状态更新 --> 终点

LangGraph Patterns

LangGraph builds stateful multi-step LLM workflows as directed graphs. Each node is a Python function; edges define routing between them.

When to Activate

• Building a multi-step LLM pipeline (research → draft → review → publish)
• Implementing human-in-the-loop interrupts or approval steps
• Designing conditional routing based on LLM output
• Adding persistence/memory to an agent across sessions
• Streaming intermediate results to the client
• Coordinating multiple agents as subgraphs
• Debugging InvalidUpdateError, cycle errors, or state shape issues

Core Concepts

StateGraph
├── State        — TypedDict that flows through every node
├── Nodes        — functions: State → State update (partial dict)
├── Edges        — unconditional routing A → B
├── Conditional  — function decides which node to go to next
└── Checkpointer — persists state between invocations (memory)

Minimal Example

from typing import TypedDict, Annotated
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from langchain_openai import ChatOpenAI

# 1. Define state — Annotated[list, add_messages] appends instead of replacing
class State(TypedDict):
    messages: Annotated[list, add_messages]

llm = ChatOpenAI(model="gpt-4o-mini")

# 2. Define a node — receives full state, returns partial update
def chatbot(state: State) -> dict:
    return {"messages": [llm.invoke(state["messages"])]}

# 3. Build the graph
graph = (
    StateGraph(State)
    .add_node("chatbot", chatbot)
    .add_edge(START, "chatbot")
    .add_edge("chatbot", END)
    .compile()
)

# 4. Invoke
result = graph.invoke({"messages": [{"role": "user", "content": "Hello!"}]})
print(result["messages"][-1].content)

State Design

from typing import TypedDict, Annotated
from operator import add

# Annotated reducers control how values merge on update
class ResearchState(TypedDict):
    # add_messages: appends new messages, deduplicates by ID
    messages: Annotated[list, add_messages]

    # add (operator.add): appends items from each node update
    sources: Annotated[list[str], add]

    # Last-write-wins (default — no annotation needed)
    query: str
    status: str
    final_report: str | None

    # Optional fields
    error: str | None

Key rule: Nodes return a partial dict — only include keys you want to update. LangGraph merges with the existing state using the reducer.

# Node returns partial — only updates 'status' and 'sources'
def fetch_sources(state: ResearchState) -> dict:
    sources = search_web(state["query"])
    return {
        "sources": sources,      # add reducer: appends
        "status": "sources_ready",
    }

Conditional Routing

from langgraph.graph import StateGraph, START, END

def route_after_llm(state: State) -> str:
    """Return the name of the next node (or END)."""
    last_message = state["messages"][-1]

    # If the LLM called a tool, go to tools node
    if last_message.tool_calls:
        return "tools"

    # Otherwise finish
    return END

graph = StateGraph(State)
graph.add_node("llm", call_llm)
graph.add_node("tools", run_tools)

graph.add_edge(START, "llm")
graph.add_conditional_edges(
    "llm",              # source node
    route_after_llm,    # routing function → returns node name
    {                   # optional: map return values to node names
        "tools": "tools",
        END: END,
    },
)
graph.add_edge("tools", "llm")   # loop back after tools

Multiple possible routes

def classify_query(state: State) -> str:
    query = state["query"].lower()
    if "code" in query:   return "code_agent"
    if "math" in query:   return "math_agent"
    return "general_agent"

graph.add_conditional_edges(
    "classifier",
    classify_query,
    ["code_agent", "math_agent", "general_agent"],  # all possible targets
)

Tool Calling

from langchain_core.tools import tool
from langgraph.prebuilt import ToolNode

@tool
def search_web(query: str) -> str:
    """Search the web for current information."""
    return web_search_api(query)

@tool
def calculator(expression: str) -> float:
    """Evaluate a mathematical expression."""
    return eval(expression)  # use safer eval in production

tools = [search_web, calculator]
tool_node = ToolNode(tools)              # pre-built node that runs tools

llm_with_tools = ChatOpenAI(model="gpt-4o-mini").bind_tools(tools)

def call_llm(state: State) -> dict:
    return {"messages": [llm_with_tools.invoke(state["messages"])]}

def should_continue(state: State) -> str:
    return "tools" if state["messages"][-1].tool_calls else END

graph = StateGraph(State)
graph.add_node("llm", call_llm)
graph.add_node("tools", tool_node)
graph.add_edge(START, "llm")
graph.add_conditional_edges("llm", should_continue)
graph.add_edge("tools", "llm")

Persistence (Checkpointers)

Checkpointers save state after every node so the graph can be paused, resumed, or continued in a new session.

from langgraph.checkpoint.memory import MemorySaver       # in-process (dev/test)
from langgraph.checkpoint.postgres import PostgresSaver    # production

# In-memory checkpointer
memory = MemorySaver()
graph = StateGraph(State).compile(checkpointer=memory)

# PostgreSQL checkpointer
import psycopg
conn = psycopg.connect("postgresql://user:pass@localhost/db")
checkpointer = PostgresSaver(conn)
graph = StateGraph(State).compile(checkpointer=checkpointer)

# thread_id groups messages into a "conversation" — same ID = same history
config = {"configurable": {"thread_id": "user-123-session-1"}}

# First call — creates new thread
result = graph.invoke({"messages": [HumanMessage("Hello")]}, config=config)

# Second call — continues the same thread
result = graph.invoke({"messages": [HumanMessage("Follow up")]}, config=config)

# Get current state of a thread
snapshot = graph.get_state(config)
print(snapshot.values)        # current state
print(snapshot.next)          # next node to run (empty if finished)

Human-in-the-Loop (Interrupts)

from langgraph.types import interrupt, Command

# interrupt() pauses the graph and surfaces a value to the caller
def approval_step(state: State) -> dict:
    # This raises an interrupt — graph pauses here
    human_response = interrupt({
        "question": "Should I proceed?",
        "context": state["draft"],
    })
    # Execution resumes here when resumed with a Command
    if human_response == "yes":
        return {"status": "approved"}
    return {"status": "rejected"}

graph = StateGraph(State).compile(
    checkpointer=memory,
    interrupt_before=["approval_step"],   # pause BEFORE this node
    # interrupt_after=["draft"],          # pause AFTER this node
)

# First invocation — runs until interrupt
result = graph.invoke(input, config=config)
# result contains the interrupt value

# Resume after human provides input
result = graph.invoke(
    Command(resume="yes"),   # pass human decision
    config=config,
)

Streaming

# stream_mode options:
# "values"  — full state after each node
# "updates" — partial state update from each node
# "messages"— LLM token-by-token streaming

# Stream full state values
for state in graph.stream(input, config=config, stream_mode="values"):
    print(state)

# Stream node updates only
for chunk in graph.stream(input, config=config, stream_mode="updates"):
    node_name, update = list(chunk.items())[0]
    print(f"Node '{node_name}' updated: {update}")

# Stream LLM tokens (best for chat UI)
async for chunk in graph.astream(input, config=config, stream_mode="messages"):
    if hasattr(chunk, "content"):
        print(chunk.content, end="", flush=True)

# Async streaming in FastAPI
@router.get("/chat/stream")
async def stream_chat(query: str):
    async def generate():
        async for chunk in graph.astream(
            {"messages": [HumanMessage(query)]},
            stream_mode="messages",
        ):
            if hasattr(chunk, "content") and chunk.content:
                yield f"data: {chunk.content}\n\n"
    return StreamingResponse(generate(), media_type="text/event-stream")

Subgraphs (Multi-Agent)

# Define a specialised sub-agent as its own graph
researcher = (
    StateGraph(ResearchState)
    .add_node("search", search_web)
    .add_node("summarize", summarize)
    .add_edge(START, "search")
    .add_edge("search", "summarize")
    .add_edge("summarize", END)
    .compile()
)

writer = (
    StateGraph(WriterState)
    .add_node("draft", draft_content)
    .add_node("refine", refine_draft)
    .compile()
)

# Orchestrator graph uses sub-agents as nodes
def run_researcher(state: OrchestratorState) -> dict:
    result = researcher.invoke({"query": state["topic"]})
    return {"research": result["summary"]}

orchestrator = (
    StateGraph(OrchestratorState)
    .add_node("research", run_researcher)
    .add_node("write",    run_writer)
    .add_edge(START, "research")
    .add_edge("research", "write")
    .add_edge("write", END)
    .compile(checkpointer=memory)
)

Agentex Integration

In Agentex Temporal agents, LangGraph runs inside a Temporal activity (not directly in the workflow). The ADK provides helpers:

from agentex.lib import adk

# In an activity:
async def run_langgraph_agent(params: AgentParams) -> str:
    graph = build_my_graph()

    # stream_langgraph_events sends each token/update to the Agentex UI
    async for event in adk.stream_langgraph_events(
        graph=graph,
        inputs={"messages": [HumanMessage(params.user_message)]},
        task_id=params.task_id,
    ):
        pass

    return final_result

# Checkpointer backed by Agentex state (MongoDB) for persistence
checkpointer = adk.create_checkpointer(task_id=params.task_id)
graph = build_my_graph().compile(checkpointer=checkpointer)

Debugging

# Print the graph structure
print(graph.get_graph().draw_ascii())

# Print state at each step
for step in graph.stream(input, stream_mode="values"):
    print("---")
    for k, v in step.items():
        print(f"  {k}: {v}")

# Inspect checkpointed history
history = list(graph.get_state_history(config))
for snapshot in history:
    print(snapshot.values, snapshot.next, snapshot.created_at)

# Replay from a specific checkpoint
graph.invoke(None, config={**config, "checkpoint_id": old_checkpoint_id})

Common Errors

Red Flags

• Returning full state from a node — returning the entire state dict instead of a partial update overwrites all fields and breaks reducers; nodes must return only the keys they changed
• Cycles with no exit condition — a loop between two nodes with no conditional edge to END causes GraphRecursionError; always add a conditional edge that can reach END
• MemorySaver in production — in-process memory is lost on worker restart; use PostgresSaver (or another persistent backend) for any deployed graph
• Non-deterministic code in node functions — calling time.time(), random, or direct HTTP requests inside nodes makes replay unpredictable in LangGraph Cloud and Temporal-hosted graphs; use activity patterns for side effects
• Missing thread_id or reusing it across unrelated sessions — reusing a thread_id continues an old conversation; always generate a unique ID per session and pass it in the config's configurable dict
• Human-in-the-loop without a checkpointer — interrupt() silently does nothing if the graph was compiled without a checkpointer; interrupts require checkpointer= in .compile()
• Accessing relationship fields across incompatible subgraph state types — parent and subgraph states must have compatible shapes; passing keys the subgraph doesn't declare in its TypedDict causes InvalidUpdateError
• add_messages on a field that isn't a message list — annotating a plain list of strings with add_messages deduplicates by message ID and discards entries without one; use operator.add for plain list fields

Checklist

• State is a TypedDict with explicit reducers (add_messages, add) for list fields
• Nodes return partial dicts — only updated keys
• All cycles have a conditional edge that can route to END
• Tools defined with @tool decorator and bound to LLM with .bind_tools()
• Production graphs use PostgresSaver (not MemorySaver)
• thread_id in config is unique per conversation/session
• Human-in-the-loop graphs always compiled with a checkpointer
• Streaming uses astream for async contexts