第 10 章：子 Agent 与结构化交接

子 Agent 的目的不是把任何函数都拆成 Agent，而是隔离专业工具、上下文和失败边界。父 Agent 只传任务契约，子 Agent 只返结果契约，禁止把完整父对话、认证信息或所有 Store 内容直接转发。

用户问题 -> 父：TaskContract(topic, constraints)
        -> 子图：只获得任务和必要 context
        -> 父：ResultContract(summary, sources, uncertainty)
        -> 最终回答

当前项目已经使用推荐的显式方式：deep_researcher.py 编译 supervisor_subgraph 和 researcher_subgraph，ConductResearch 定义任务，ResearcherOutputState 定义子图输出，supervisor 汇总 compressed_research。这需要并发、压缩和明确 state，所以继续用 LangGraph compiled subgraph，而不是强行嵌入 create_agent。

deepagents 的 SubAgentMiddleware 是额外包的可选扩展，不是 LangChain core，也不在本仓库依赖中。本课不安装它；需要其自动委派能力时，应单独评估隔离、成本、可观测性和版本兼容性。运行 15_structured_handoff.py 会以 Pydantic TaskContract/ResultContract 验证最小父子图契约；该示例不联网。

相关资源

查看示例代码：docs/langchain/examples/15_structured_handoff.py

"""Chapter 10: pass a narrow Pydantic task/result contract through a subgraph."""

from typing import TypedDict

from langgraph.graph import END, START, StateGraph
from pydantic import BaseModel, Field


class TaskContract(BaseModel):
    topic: str = Field(min_length=3)
    constraints: list[str] = Field(default_factory=list)


class ResultContract(BaseModel):
    summary: str
    uncertainty: str


class ChildState(TypedDict):
    task: TaskContract
    result: ResultContract


def research_task(state: ChildState) -> dict:
    task = state["task"]
    return {"result": ResultContract(summary=f"已处理：{task.topic}", uncertainty="未调用外部来源")}


child = StateGraph(ChildState)
child.add_node("research", research_task)
child.add_edge(START, "research")
child.add_edge("research", END)
compiled_child = child.compile()


def main() -> None:
    result = compiled_child.invoke({"task": TaskContract(topic="LangGraph 状态", constraints=["中文"])} )
    assert result["result"].summary == "已处理：LangGraph 状态"
    print(result["result"].model_dump_json())


if __name__ == "__main__":
    main()