Deep Agents overview - Docs by LangChain中文

这是开始构建由 LLMs 驱动的 agents 和 applications 的最简单方式，内置 task planning、用于 context management 的 file systems、subagent-spawning 和 long-term memory。你可以将 deep agents 用于任何 task，包括复杂的 multi-step tasks。 Deep Agents 是一种 “agent harness”。它与其他 agent frameworks 使用相同的 core tool calling loop，但内置了让 agents 能可靠处理真实 tasks 的 capabilities：

Take actions in an environment

通过 tools 执行动作，read 和 write files，并 execute code

Connect to your data

在合适时机加载 memories、skills 和 domain knowledge

Manage growing context

在 long runs 中 summarize history，并 offload large results

Parallelize tasks

委派给在 isolated context windows 中运行的 general 或 specialized subagents

Stay in the loop

在关键 decision points 暂停以等待 human approval

Improve over time

根据真实 usage 更新 memory、skills 和 prompts

请参阅 Harness capabilities，了解每个 component 的完整拆解。 deepagents 是一个基于 LangChain agents core building blocks 构建的 standalone library。它使用 LangGraph runtime 提供 durable execution、streaming、human-in-the-loop 和其他 features。 LangChain 是为你的 agents 提供 core building blocks 的 framework。若要进一步了解 LangChain、LangGraph 和 Deep Agents 之间的差异，请参阅 Frameworks, runtimes, and harnesses。若要与 Anthropic 的 harness 进行 side-by-side comparison，请参阅 Deep Agents vs. Claude Agent SDK。

Create a deep agent

# pip install -qU deepagents langchain-google-genai
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="google_genai:gemini-3.5-flash",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

# pip install -qU deepagents langchain-openai
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="openai:gpt-5.4",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

# pip install -qU deepagents langchain-anthropic
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

# pip install -qU deepagents langchain-openrouter
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="openrouter:anthropic/claude-sonnet-4-6",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

# pip install -qU deepagents langchain-fireworks
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="fireworks:accounts/fireworks/models/qwen3p5-397b-a17b",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

# pip install -qU deepagents langchain-baseten
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="baseten:zai-org/GLM-5",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

# pip install -qU deepagents langchain-ollama
from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"

agent = create_deep_agent(
    model="ollama:devstral-2",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

请参阅 Quickstart 和 Customization guide，开始构建你自己的 Deep Agents agents 和 applications。

使用 LangSmith trace requests、debug agent behavior 并 evaluate outputs。按照 observability quickstart 完成设置。准备进入 production 时，请参阅 Going to production 了解 LangSmith deployment options。

Core capabilities

使用 Deep Agents SDK 构建可跨 任何 model provider 处理复杂 multi-step tasks 的 agents。SDK 随附以下 built-in capabilities：

Planning and task decomposition

内置 write_todos tool 让 agents 可以把 complex tasks 拆成 discrete steps、track progress，并在出现新信息时 adapt plans。

Context management

内置 context compression 会将 large tool inputs 和 results offload 到 virtual filesystem，并 summarizes older messages，让 agents 在 extended sessions 中保持有效。

Pluggable filesystem backends

通过 pluggable backends swap virtual filesystem：in-memory state、local disk、LangGraph store、composite routing，或带 read/write permission rules 的 custom backend。

Shell execution

Shell-capable backends 会添加 execute tool，用于 tests、builds、git operations 和 system tasks。Local development 时在 host 上使用 LocalShellBackend；需要与 host system 隔离时使用 sandbox backend。

Interpreters

添加 interpreter，在 in-memory runtime 中运行 JavaScript。Interpreters 让 agents 无需 full shell environment，也能 programmatically compose tools、orchestrate subagents，并 transform structured data。

Subagent spawning

内置 task tool 会为 subtasks spawn general-purpose 或 specialized subagents，以实现 context isolation。对于 long-running 或 parallel work，async subagents 会在 background 中运行，并支持 progress checks、follow-ups 和 cancellation。

Streaming

Event streaming 会将 agent runs 暴露为 messages、tool calls、values 和 output 的 typed projections。Deep Agents 添加了 stream.subagents，让每个 delegated task 都有自己的 handle，并拥有独立 message、tool-call 和 nested subagent streams。

Long-term memory

使用 LangGraph 的 Memory Store 在 threads 和 conversations 之间持久化 memory。

Filesystem permissions

声明 permission rules，控制 agents 可以 read 或 write 哪些 files 和 directories。Subagents 可以 inherit 或 override parent 的 rules。

Human-in-the-loop

使用 LangGraph 的 interrupt capabilities 为 sensitive tool operations 配置 human approval。

Skills

使用 reusable skills 扩展 agents，提供 specialized workflows、domain knowledge 和 custom instructions。

Smart defaults

附带 opinionated system prompts，教 model 在行动前 plan、verify work，并 manage context。你可以按需 customize 或 replace defaults。

若要构建不包含这些 builtin capabilities 的 custom agents，请考虑使用 LangChain 的 create_agent，或构建 custom LangGraph workflow。

Get started

Quickstart

构建你的 first deep agent

Customization

了解 customization options

Models

配置 models 和 providers

Backends

选择并配置 pluggable filesystem backends

Sandboxes

在 isolated environments 中 execute code

Interpreters

在 QuickJS 中 compose tools 和 transform data

Permissions

使用 permission rules 控制 filesystem access

Human-in-the-loop

为 sensitive operations 配置 approval

Code

使用 Deep Agents Code

ACP

通过 ACP 在 code editors 中使用 deep agents

Reference

查看 deepagents API reference

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.

Edit this page on GitHub or file an issue.

Take actions in an environment

Connect to your data

Manage growing context

Parallelize tasks

Stay in the loop

Improve over time

​ Create a deep agent

​Core capabilities

Planning and task decomposition

Context management

Pluggable filesystem backends

Shell execution

Interpreters

Subagent spawning

Streaming

Long-term memory

Filesystem permissions

Human-in-the-loop

Skills

Smart defaults

​Get started

Quickstart

Customization

Models

Backends

Sandboxes

Interpreters

Permissions

Human-in-the-loop

Code

ACP

Reference

Create a deep agent

Core capabilities

Get started