deep-agents

LangChain Deep Agents

Package: deepagents v0.5.0 (Beta, MIT, Python >=3.11) Homepage: https://docs.langchain.com/oss/python/deepagents/overview GitHub: https://github.com/langchain-ai/deepagents

pip install deepagents
uv add deepagents

Returns a compiled CompiledStateGraph (LangGraph) — all LangGraph primitives (.stream(), .ainvoke(), Studio, Platform) work on the result.

Reference Pages

Core API

create_deep_agent()

from deepagents import create_deep_agent

agent = create_deep_agent(
    name: str | None = None,
    model: str | BaseChatModel | None = None,       # default: "claude-sonnet-4-6"
    tools: Sequence[BaseTool | Callable | dict] | None = None,
    *,
    system_prompt: str | SystemMessage | None = None,
    backend: BackendProtocol | None = None,          # default: StateBackend()
    store: BaseStore | None = None,
    checkpointer: BaseCheckpointSaver | None = None, # REQUIRED for human-in-the-loop
    interrupt_on: dict | None = None,
    subagents: list[dict | CompiledSubAgent] | None = None,
    skills: list[str] | None = None,
    memory: list[str] | None = None,
    context_schema: type | None = None,
    response_format: type[BaseModel] | None = None,  # structured output
)
# Returns: CompiledStateGraph

Model format: "provider:model" string or BaseChatModel instance. Supported providers: Anthropic, OpenAI, Google, Azure, AWS Bedrock, OpenRouter, Fireworks, Baseten, Ollama.

Invocation

# Single-turn
result = agent.invoke({"messages": [{"role": "user", "content": "..."}]})

# With thread (required for checkpointing, human-in-the-loop)
result = agent.invoke(
    {"messages": [{"role": "user", "content": "..."}]},
    config={"configurable": {"thread_id": "some-id"}},
)

# Async
result = await agent.ainvoke(...)

# Structured output result key
result["structured_response"]  # when response_format= is set

Built-in Tools

These are injected automatically based on middleware:

Minimal Example

from deepagents import create_deep_agent

def get_weather(city: str) -> str:
    """Return current weather for a city."""
    return f"Sunny, 72°F in {city}"

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

result = agent.invoke({"messages": [{"role": "user", "content": "Weather in NYC?"}]})

Tool Registration

Any Python callable with type hints and a docstring works. LangChain BaseTool and @tool-decorated functions also accepted.

from langchain.tools import tool

@tool
def search(query: str) -> str:
    """Search the web for information."""
    return search_api(query)

agent = create_deep_agent(tools=[search, my_plain_callable])

Streaming

for chunk in agent.stream(
    {"messages": [{"role": "user", "content": "..."}]},
    stream_mode=["updates", "messages", "custom"],
    subgraphs=True,
    version="v2",
):
    # chunk["type"] is "updates", "messages", or "custom"
    # chunk["ns"] is () for main agent, ("tools:<id>",) for subagent
    ...

See references/streaming.md for event structure details.

Human-in-the-Loop

from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command

agent = create_deep_agent(
    tools=[sensitive_tool],
    checkpointer=MemorySaver(),           # REQUIRED
    interrupt_on={"sensitive_tool": True},
)

thread = {"configurable": {"thread_id": "t1"}}
result = agent.invoke(input, config=thread, version="v2")

# result.interrupts[0].value contains action_requests and review_configs
# Resume:
agent.invoke(Command(resume={"decisions": ["approve"]}), config=thread, version="v2")

interrupt_on values: True, False, or {"allowed_decisions": ["approve", "edit", "reject"]}.

Standards

Always do

1. Set a thread_id whenever using checkpointer or needing human-in-the-loop.
2. Use StoreBackend with namespace factories for multi-tenant deployments — never StateBackend in production serving multiple users.
3. Pass checkpointer=MemorySaver() (or persistent saver) when using interrupt_on.
4. Provide type hints and docstrings on all tool callables — the agent uses them to understand the tool.
5. Use backend.upload_files() before invoking when seeding a sandbox with source code or data.
6. Read version="v2" when streaming — required for unified event format.

Never do

1. Don't use StateBackend for multi-user — files are not isolated between threads.
2. Don't put secrets in sandbox environments — they can be read/exfiltrated via context injection.
3. Don't skip checkpointer with interrupt_on — human-in-the-loop silently fails without it.
4. Don't expect subagents to inherit the parent's tools, middleware, or skills — they must be specified explicitly.
5. Don't use the deprecated runtime constructor arg on StateBackend/StoreBackend.

Vs. Claude Agent SDK

Deep Agents adds on top of LangGraph: virtual filesystem with pluggable backends, composable middleware, long-term memory via Store, and sandbox-as-tool pattern. Use Deep Agents when you need multi-backend filesystems, middleware composition, or model-agnostic support. Use Claude Agent SDK when tight Claude Code integration (hooks, permissions, CLAUDE.md) is the priority. See the comparison page.