入口在 libs/langchain_v1/langchain/agents/factory.py:808（前面 :740/:762/:786 是类型重载）：

# libs/langchain_v1/langchain/agents/factory.py:808
def create_agent(
    model: str | BaseChatModel,                    # 模型：实例或 "openai:gpt-4o" 字符串
    tools: Sequence[BaseTool | Callable | dict[str, Any]] | None = None,  # Day13 的货架
    *,
    system_prompt: str | SystemMessage | None = None,   # 人设
    middleware: Sequence[AgentMiddleware[...]] = (),    # 中间件：循环各切面的钩子
    response_format: ResponseFormat | ... | None = None, # 结构化输出
    state_schema: type[AgentState[ResponseT]] | None = None,
    context_schema: type[ContextT] | None = None,
    checkpointer: Checkpointer | None = None,      # ★持久化：断点续跑/多轮会话
    store: BaseStore | None = None,
    interrupt_before: list[str] | None = None,     # ★在某节点前暂停（人工确认）
    interrupt_after: list[str] | None = None,
    debug: bool = False, name: str | None = None, ...
) -> CompiledStateGraph[AgentState[ResponseT], ...]:   # ★返回值：一张编译好的 LangGraph 图！
    """Creates an agent graph that calls tools in a loop
    until a stopping condition is met."""

往下读工厂内部——先建图（factory.py:1159），再加节点（factory.py:1502）：

# libs/langchain_v1/langchain/agents/factory.py:1159（裁剪）
graph: StateGraph[AgentState[ResponseT], ...] = StateGraph(
    state_schema=resolved_state_schema,        # 棋局的形状（messages 等）
    input_schema=input_schema,
    output_schema=output_schema,
    context_schema=context_schema,
)
# ...
# factory.py:1502
graph.add_node("model", RunnableCallable(model_node, amodel_node, trace=False))  # ① 思考节点
if tool_node is not None:
    graph.add_node("tools", tool_node)         # ② 干活节点（ToolNode，来自 LangGraph）
# ...
# factory.py:1620
graph.add_edge(START, entry_node)              # ③ 起点 → 入口（无中间件时就是 "model"）

model 节点每次执行前，先用 _get_bound_model（factory.py:1272）现场绑定工具，然后 _execute_model_sync（factory.py:1406）真正调用：

# libs/langchain_v1/langchain/agents/factory.py:1396（裁剪：无结构化输出的标准路径）
if final_tools:
    return (
        request.model.bind_tools(                     # ★Day14 的 bind_tools！
            final_tools, tool_choice=request.tool_choice, **request.model_settings
        ),
        None,
    )
return request.model.bind(**request.model_settings), None   # 没工具就裸调

# factory.py:1406
def _execute_model_sync(request: ModelRequest[ContextT]) -> ModelResponse:
    model_, effective_response_format = _get_bound_model(request)  # ① 绑工具
    messages = request.messages
    if request.system_message:
        messages = [request.system_message, *messages]             # ② system 人设插最前
    output = model_.invoke(messages)                               # ③ ★一次 Day05 的标准调用
    ...
    return ModelResponse(result=messages_list, structured_response=structured_response)

先看 model 走完之后去哪——_make_model_to_tools_edge（factory.py:1840）：

# libs/langchain_v1/langchain/agents/factory.py:1846（裁剪）
def model_to_tools(state: dict[str, Any]) -> str | list[Send] | None:
    ...
    last_ai_message, tool_messages = _fetch_last_ai_and_tool_messages(state["messages"])
    ...
    # 3. If the model hasn't called any tools, exit the loop
    # this is the classic exit condition for an agent loop      ← 源码原话！
    if len(last_ai_message.tool_calls) == 0:
        return end_destination                        # ★没工单 → 收工

    pending_tool_calls = [c for c in last_ai_message.tool_calls
                          if c["id"] not in tool_message_ids     # 还没回执的工单
                          and c["name"] not in structured_output_tools]
    if pending_tool_calls:
        return [Send("tools", [tool_call]) for tool_call in pending_tool_calls]  # ★每单一发，并行执行

再看 tools 干完之后去哪——_make_tools_to_model_edge（factory.py:1921）：

# libs/langchain_v1/langchain/agents/factory.py:1928（裁剪）
def tools_to_model(state: dict[str, Any]) -> str | None:
    last_ai_message, tool_messages = _fetch_last_ai_and_tool_messages(state["messages"])
    ...
    # 2. Exit condition: All executed tools have return_direct=True
    if client_side_tool_calls and all(
        tool_node.tools_by_name[c["name"]].return_direct for c in client_side_tool_calls):
        return end_destination                        # 工具声明"结果直出"→ 不再回模型

    # 3. Exit condition: A structured output tool was executed
    if any(t.name in structured_output_tools for t in tool_messages):
        return end_destination                        # 已经"填表交卷"→ 收工

    # 4. Default: Continue the loop
    return model_destination                          # ★默认：带着回执回去继续想

今天 import 痕迹已经说明一切（factory.py:26：from langgraph.graph.state import StateGraph）。三兄弟分工：

cd /Users/bitmart/work/codes/github/AI_WORK/langchain

# 1. 工厂签名与"造图"证据
sed -n '808,830p'  libs/langchain_v1/langchain/agents/factory.py   # create_agent 签名
sed -n '1159,1170p' libs/langchain_v1/langchain/agents/factory.py  # StateGraph(...)
grep -n 'add_node("model"\|add_node("tools"' libs/langchain_v1/langchain/agents/factory.py

# 2. 循环的两条边（重点读注释）
sed -n '1846,1892p' libs/langchain_v1/langchain/agents/factory.py  # model_to_tools
sed -n '1928,1954p' libs/langchain_v1/langchain/agents/factory.py  # tools_to_model

# 3. 现成中间件目录：每个文件名都是一个生产级痛点
ls libs/langchain_v1/langchain/agents/middleware/