14 附录：手写 `@tool` 包装子智能体 vs Deep Agents 的 `subagents=`——同一个原理，两种实现

1
class TaskToolSchema(BaseModel):
2
    """Input schema for the `task` tool."""
3

4
    description: str = Field(description=(
5
        "A detailed description of the task for the subagent to perform autonomously. "
6
        "Include all necessary context and specify the expected output format."))
7
    subagent_type: str = Field(description=(
8
        "The type of subagent to use. Must be one of the available agent types listed in the tool description."))

1
subagent_description_str = "\n".join(f"- {s['name']}: {s['description']}" for s in compiled_subagents)
2
description = TASK_TOOL_DESCRIPTION.format(available_agents=subagent_description_str)

1
subagent_state["messages"] = [HumanMessage(content=description)]

1
for msg in reversed(result["messages"]):
2
    if isinstance(msg, AIMessage):
3
        text = msg.text.rstrip() if msg.text else ""
4
        if text:
5
            content = text
6
            break
7
return Command(update={**state_update, "messages": [ToolMessage(content, tool_call_id=tool_call_id)]})

1
• ls            params=['path']
2
• read_file     params=['file_path','offset','limit']
3
• write_file    params=['file_path','content']
4
• edit_file     params=['file_path','old_string','new_string','replace_all']
5
• delete        params=['file_path']
6
• glob          params=['pattern','path']
7
• grep          params=['pattern','path','glob','output_mode','max_count']
8
• task          params=['description','subagent_type']      ← 唯一的派发入口
9
• save_note     params=['title','body']                     ← 我们自己的业务工具

1
Launch an ephemeral subagent to handle a complex, multi-step task in an isolated context window.
2

3
Available agent types and the tools they have access to:
4
- general-purpose: General-purpose agent for researching complex questions, ...
5
- research-agent: 用于需要多步调研、比较、查证的问题
6

7
Specify subagent_type to select the agent. Usage notes:
8
- Launch multiple agents concurrently when their tasks are independent, using a single message with multiple tool calls.
9
- Each invocation is stateless: the agent sees only the prompt you give it and returns a single final report. ...