deep-agents-orchestration

Deep Agents 包含三种编排能力：

1. SubAgentMiddleware：通过 task 工具将工作委派给专用 Agent
2. TodoListMiddleware：通过 write_todos 工具规划并跟踪任务
3. HumanInTheLoopMiddleware：在执行敏感操作前要求审批

这三种能力都会自动包含在 create_deep_agent() 中。

---

子 Agent（任务委派）

适合使用子 Agent 的情况	适合使用主 Agent 的情况
任务需要专用工具	通用工具已足够
希望隔离复杂工作	单步操作
需要为主 Agent 提供干净的上下文	可以接受上下文膨胀

主 Agent 拥有 `task` 工具 -> 创建全新的子 Agent -> 子 Agent 自主执行 -> 返回最终报告。

默认子 Agent："general-purpose"——自动可用，且与主 Agent 使用相同的工具和配置。

创建一个自定义的 "researcher" 子 Agent，并为其配置用于搜索学术论文的专用工具。 ```python from deepagents import create_deep_agent from langchain.tools import tool

@tool def search_papers(query: str) -> str: """Search academic papers.""" return f"Found 10 papers about {query}"

agent = create_deep_agent( subagents=[ { "name": "researcher", "description": "Conduct web research and compile findings", "system_prompt": "Search thoroughly, return concise summary", "tools": [search_papers], } ] )

Main agent delegates: task(agent="researcher", instruction="Research AI trends")

</python>
<typescript>
创建一个自定义的 "researcher" 子 Agent，并为其配置用于搜索学术论文的专用工具。
```typescript
import { createDeepAgent } from "deepagents";
import { tool } from "@langchain/core/tools";
import { z } from "zod";

const searchPapers = tool(
  async ({ query }) => `Found 10 papers about ${query}`,
  { name: "search_papers", description: "Search papers", schema: z.object({ query: z.string() }) }
);

const agent = await createDeepAgent({
  subagents: [
    {
      name: "researcher",
      description: "Conduct web research and compile findings",
      systemPrompt: "Search thoroughly, return concise summary",
      tools: [searchPapers],
    }
  ]
});

// Main agent delegates: task(agent="researcher", instruction="Research AI trends")

使用 CompiledSubAgent 将预构建的 LangGraph 图封装为子 Agent。 ```python from deepagents import create_deep_agent, CompiledSubAgent from langgraph.graph import StateGraph, MessagesState

builder = StateGraph(MessagesState)

... add nodes/edges ...

graph = builder.compile()

agent = create_deep_agent( subagents=[ CompiledSubAgent( name="my-graph", description="Run the custom LangGraph workflow", runnable=graph, # Must be a compiled graph; state must have "messages" key ) ] )

</python>
</ex-compiledsubagent>

<ex-subagent-with-hitl>
<python>
配置一个在执行敏感操作时需要 HITL 审批的子 Agent。
```python
from deepagents import create_deep_agent
from langgraph.checkpoint.memory import MemorySaver

agent = create_deep_agent(
    subagents=[
        {
            "name": "code-deployer",
            "description": "Deploy code to production",
            "system_prompt": "You deploy code after tests pass.",
            "tools": [run_tests, deploy_to_prod],
            "interrupt_on": {"deploy_to_prod": True},  # Require approval
        }
    ],
    checkpointer=MemorySaver()  # Required for interrupts
)

子 Agent 是无状态的——请在一次调用中提供完整指令。 ```python # WRONG: Subagents don't remember previous calls # task(agent='research', instruction='Find data') # task(agent='research', instruction='What did you find?') # Starts fresh!

CORRECT: Complete instructions upfront

task(agent='research', instruction='Find data on AI, save to /research/, return summary')

</python>
<typescript>
子 Agent 是无状态的——请在一次调用中提供完整指令。
```typescript
// WRONG: Subagents don't remember previous calls
// task research: Find data
// task research: What did you find?  // Starts fresh!

// CORRECT: Complete instructions upfront
// task research: Find data on AI, save to /research/, return summary

自定义子 Agent 不会继承主 Agent 的技能。 ```python # WRONG: Custom subagent won't have main agent's skills agent = create_deep_agent( skills=["/main-skills/"], subagents=[{"name": "helper", ...}] # No skills inherited )

CORRECT: Provide skills explicitly (general-purpose subagent DOES inherit)

agent = create_deep_agent( skills=["/main-skills/"], subagents=[{"name": "helper", "skills": ["/helper-skills/"], ...}] )

</python>
</fix-custom-subagents-dont-inherit-skills>

---

## TodoList（任务规划）

<when-to-use-todolist>

| 适合使用 TodoList 的情况 | 适合跳过 TodoList 的情况 |
|------------------|-------------------|
| 复杂的多步骤任务 | 简单的单次操作任务 |
| 长时间运行的操作 | 快速操作（< 3 步） |

</when-to-use-todolist>

<todolist-tool>

write_todos(todos: list[dict]) -> None

每个待办事项包含：
- `content`：任务描述
- `status`：取值为 `"pending"`、`"in_progress"` 或 `"completed"` 之一
</todolist-tool>

<ex-todolist-usage>
<python>
调用一个会自动为多步骤任务创建待办列表的 Agent。
```python
from deepagents import create_deep_agent

agent = create_deep_agent()  # TodoListMiddleware included by default

result = agent.invoke({
    "messages": [{"role": "user", "content": "Create a REST API: design models, implement CRUD, add auth, write tests"}]
}, config={"configurable": {"thread_id": "session-1"}})

# Agent's planning via write_todos:
# [
#   {"content": "Design data models", "status": "in_progress"},
#   {"content": "Implement CRUD endpoints", "status": "pending"},
#   {"content": "Add authentication", "status": "pending"},
#   {"content": "Write tests", "status": "pending"}
# ]

调用一个会自动为多步骤任务创建待办列表的 Agent。 ```typescript import { createDeepAgent } from "deepagents";

const agent = await createDeepAgent(); // TodoListMiddleware included

const result = await agent.invoke({ messages: [{ role: "user", content: "Create a REST API: design models, implement CRUD, add auth, write tests" }] }, { configurable: { thread_id: "session-1" } });

</typescript>
</ex-todolist-usage>

<ex-access-todo-state>
<python>
调用后，从 Agent 的最终状态中访问待办列表。
```python
result = agent.invoke({...}, config={"configurable": {"thread_id": "session-1"}})

# Access todo list from final state
todos = result.get("todos", [])
for todo in todos:
    print(f"[{todo['status']}] {todo['content']}")

待办列表状态需要 `thread_id`，才能在多次调用之间持久保存。 ```python # WRONG: Fresh state each time without thread_id agent.invoke({"messages": [...]})

CORRECT: Use thread_id

config = {"configurable": {"thread_id": "user-session"}} agent.invoke({"messages": [...]}, config=config) # Todos preserved

</python>
</fix-todolist-requires-thread-id>

---
## 人机协同（审批工作流）

<when-to-use-hitl>

| 适用 HITL 的情况 | 不适用 HITL 的情况 |
|--------------|---------------|
| 高风险操作（数据库写入、部署） | 只读操作 |
| 合规要求人工监督 | 全自动工作流 |

</when-to-use-hitl>

<ex-hitl-setup>
<python>
配置执行前需要人工审批的工具。
```python
from deepagents import create_deep_agent
from langgraph.checkpoint.memory import MemorySaver

agent = create_deep_agent(
    interrupt_on={
        "write_file": True,  # All decisions allowed
        "execute_sql": {"allowed_decisions": ["approve", "reject"]},
        "read_file": False,  # No interrupts
    },
    checkpointer=MemorySaver()  # REQUIRED for interrupts
)

配置执行前需要人工审批的工具。 ```typescript import { createDeepAgent } from "deepagents"; import { MemorySaver } from "@langchain/langgraph";

const agent = await createDeepAgent({ interruptOn: { write_file: true, execute_sql: { allowedDecisions: ["approve", "reject"] }, read_file: false, }, checkpointer: new MemorySaver() // REQUIRED });

</typescript>
</ex-hitl-setup>

<ex-approval-workflow>
<python>
完整工作流：触发中断、检查状态、批准操作并恢复执行。
```python
from deepagents import create_deep_agent
from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command

agent = create_deep_agent(
    interrupt_on={"write_file": True},
    checkpointer=MemorySaver()
)

config = {"configurable": {"thread_id": "session-1"}}

# Step 1: Agent proposes write_file - execution pauses
result = agent.invoke({
    "messages": [{"role": "user", "content": "Write config to /prod.yaml"}]
}, config=config)

# Step 2: Check for interrupts
state = agent.get_state(config)
if state.next:
    print(f"Pending action")

# Step 3: Approve and resume
result = agent.invoke(Command(resume={"decisions": [{"type": "approve"}]}), config=config)

完整工作流：触发中断、检查状态、批准操作并恢复执行。 ```typescript import { createDeepAgent } from "deepagents"; import { MemorySaver, Command } from "@langchain/langgraph";

const agent = await createDeepAgent({ interruptOn: { write_file: true }, checkpointer: new MemorySaver() });

const config = { configurable: { thread_id: "session-1" } };

// Step 1: Agent proposes write_file - execution pauses let result = await agent.invoke({ messages: [{ role: "user", content: "Write config to /prod.yaml" }] }, config);

// Step 2: Check for interrupts const state = await agent.getState(config); if (state.next) { console.log("Pending action"); }

// Step 3: Approve and resume result = await agent.invoke( new Command({ resume: { decisions: [{ type: "approve" }] } }), config );

</typescript>
</ex-approval-workflow>

<ex-reject-with-feedback>
<python>
拒绝待处理的操作并提供反馈，提示 Agent 尝试其他方法。
```python
result = agent.invoke(
    Command(resume={"decisions": [{"type": "reject", "message": "Run tests first"}]}),
    config=config,
)

拒绝待处理的操作并提供反馈，提示 Agent 尝试其他方法。 ```typescript const result = await agent.invoke( new Command({ resume: { decisions: [{ type: "reject", message: "Run tests first" }] } }), config, ); ``` 允许执行前，编辑拟执行操作的参数。 ```python result = agent.invoke( Command(resume={"decisions": [{ "type": "edit", "edited_action": { "name": "execute_sql", "args": {"query": "DELETE FROM users WHERE last_login < '2020-01-01' LIMIT 100"}, }, }]}), config=config, ) ``` ### Agent 可以配置的内容

• 子 Agent 的名称、工具、模型和系统提示词
• 哪些工具需要审批
• 每种工具允许的决策类型
• TodoList 的内容和结构

Agent 不能配置的内容

• 工具名称（task、write_todos）
• HITL 协议（批准／编辑／拒绝结构）
• 跳过中断所需的 checkpointer 要求
• 让子 Agent 保持有状态（它们是临时的）

使用 interrupt_on 实现 HITL 工作流时，必须配置 Checkpointer。 ```python # WRONG agent = create_deep_agent(interrupt_on={"write_file": True})

CORRECT

agent = create_deep_agent(interrupt_on={"write_file": True}, checkpointer=MemorySaver())

</python>
<typescript>
使用 interruptOn 实现 HITL 工作流时，必须配置 Checkpointer。
```typescript
// WRONG
const agent = await createDeepAgent({ interruptOn: { write_file: true } });

// CORRECT
const agent = await createDeepAgent({ interruptOn: { write_file: true }, checkpointer: new MemorySaver() });

恢复被中断的工作流需要一致的 thread_id。 ```python # WRONG: Can't resume without thread_id agent.invoke({"messages": [...]})

CORRECT

config = {"configurable": {"thread_id": "session-1"}} agent.invoke({...}, config=config)

Resume with Command using same config

agent.invoke(Command(resume={"decisions": [{"type": "approve"}]}), config=config)

</python>
<typescript>
恢复被中断的工作流需要一致的 thread_id。
```typescript
// WRONG: Can't resume without thread_id
await agent.invoke({ messages: [...] });

// CORRECT
const config = { configurable: { thread_id: "session-1" } };
await agent.invoke({ messages: [...] }, config);
// Resume with Command using same config
await agent.invoke(new Command({ resume: { decisions: [{ type: "approve" }] } }), config);

中断发生在 invoke() 调用之间，而非执行过程中。 ```python result = agent.invoke({...}, config=config) # Step 1: triggers interrupt if "__interrupt__" in result: # Step 2: check for interrupt result = agent.invoke( # Step 3: resume Command(resume={"decisions": [{"type": "approve"}]}), config=config, ) ```