Define subgraph#

class SubgraphState(TypedDict): # note that none of these keys are shared with the parent graph state bar: str baz: str

def subgraph_node_1(state: SubgraphState): return {“baz”: “baz”}

def subgraph_node_2(state: SubgraphState): return {“bar”: state[“bar”] + state[“baz”]}

subgraph_builder = StateGraph(SubgraphState) subgraph_builder.add_node(subgraph_node_1) subgraph_builder.add_node(subgraph_node_2) subgraph_builder.add_edge(START, “subgraph_node_1”) subgraph_builder.add_edge(“subgraph_node_1”, “subgraph_node_2”) subgraph = subgraph_builder.compile()

Define parent graph#

class ParentState(TypedDict): foo: str

def node_1(state: ParentState): return {“foo”: “hi! " + state[“foo”]}

def node_2(state: ParentState): # Transform the state to the subgraph state response = subgraph.invoke({“bar”: state[“foo”]}) # Transform response back to the parent state return {“foo”: response[“bar”]}

builder = StateGraph(ParentState) builder.add_node(“node_1”, node_1) builder.add_node(“node_2”, node_2) builder.add_edge(START, “node_1”) builder.add_edge(“node_1”, “node_2”) graph = builder.compile()

for chunk in graph.stream({“foo”: “foo”}, subgraphs=True): print(chunk)

((), {’node_1’: {‘foo’: ‘hi! foo’}}) ((’node_2:577b710b-64ae-31fb-9455-6a4d4cc2b0b9’,), {‘subgraph_node_1’: {‘baz’: ‘baz’}}) ((’node_2:577b710b-64ae-31fb-9455-6a4d4cc2b0b9’,), {‘subgraph_node_2’: {‘bar’: ‘hi! foobaz’}}) ((), {’node_2’: {‘foo’: ‘hi! foobaz’}})

</Accordion>

<Accordion title="Full example: different state schemas (two levels of subgraphs)">
This is an example with two levels of subgraphs: parent -> child -> grandchild.

```python
# Grandchild graph
from typing_extensions import TypedDict
from langgraph.graph.state import StateGraph, START, END

class GrandChildState(TypedDict):
    my_grandchild_key: str

def grandchild_1(state: GrandChildState) -> GrandChildState:
    # NOTE: child or parent keys will not be accessible here
    return {"my_grandchild_key": state["my_grandchild_key"] + ", how are you"}


grandchild = StateGraph(GrandChildState)
grandchild.add_node("grandchild_1", grandchild_1)

grandchild.add_edge(START, "grandchild_1")
grandchild.add_edge("grandchild_1", END)

grandchild_graph = grandchild.compile()

# Child graph
class ChildState(TypedDict):
    my_child_key: str

def call_grandchild_graph(state: ChildState) -> ChildState:
    # NOTE: parent or grandchild keys won't be accessible here
    grandchild_graph_input = {"my_grandchild_key": state["my_child_key"]}
    grandchild_graph_output = grandchild_graph.invoke(grandchild_graph_input)
    return {"my_child_key": grandchild_graph_output["my_grandchild_key"] + " today?"}

child = StateGraph(ChildState)
# We're passing a function here instead of just compiled graph (`grandchild_graph`)
child.add_node("child_1", call_grandchild_graph)
child.add_edge(START, "child_1")
child.add_edge("child_1", END)
child_graph = child.compile()

# Parent graph
class ParentState(TypedDict):
    my_key: str

def parent_1(state: ParentState) -> ParentState:
    # NOTE: child or grandchild keys won't be accessible here
    return {"my_key": "hi " + state["my_key"]}

def parent_2(state: ParentState) -> ParentState:
    return {"my_key": state["my_key"] + " bye!"}

def call_child_graph(state: ParentState) -> ParentState:
    child_graph_input = {"my_child_key": state["my_key"]}
    child_graph_output = child_graph.invoke(child_graph_input)
    return {"my_key": child_graph_output["my_child_key"]}

parent = StateGraph(ParentState)
parent.add_node("parent_1", parent_1)
# We're passing a function here instead of just a compiled graph (`child_graph`)
parent.add_node("child", call_child_graph)
parent.add_node("parent_2", parent_2)

parent.add_edge(START, "parent_1")
parent.add_edge("parent_1", "child")
parent.add_edge("child", "parent_2")
parent.add_edge("parent_2", END)

parent_graph = parent.compile()

for chunk in parent_graph.stream({"my_key": "Bob"}, subgraphs=True):
    print(chunk)

((), {'parent_1': {'my_key': 'hi Bob'}})
(('child:2e26e9ce-602f-862c-aa66-1ea5a4655e3b', 'child_1:781bb3b1-3971-84ce-810b-acf819a03f9c'), {'grandchild_1': {'my_grandchild_key': 'hi Bob, how are you'}})
(('child:2e26e9ce-602f-862c-aa66-1ea5a4655e3b',), {'child_1': {'my_child_key': 'hi Bob, how are you today?'}})
((), {'child': {'my_key': 'hi Bob, how are you today?'}})
((), {'parent_2': {'my_key': 'hi Bob, how are you today? bye!'}})

Define subgraph#

class SubgraphState(TypedDict): foo: str # shared with parent graph state bar: str # private to SubgraphState

def subgraph_node_1(state: SubgraphState): return {“bar”: “bar”}

def subgraph_node_2(state: SubgraphState): # note that this node is using a state key (‘bar’) that is only available in the subgraph # and is sending update on the shared state key (‘foo’) return {“foo”: state[“foo”] + state[“bar”]}

subgraph_builder = StateGraph(SubgraphState) subgraph_builder.add_node(subgraph_node_1) subgraph_builder.add_node(subgraph_node_2) subgraph_builder.add_edge(START, “subgraph_node_1”) subgraph_builder.add_edge(“subgraph_node_1”, “subgraph_node_2”) subgraph = subgraph_builder.compile()

Define parent graph#

class ParentState(TypedDict): foo: str

def node_1(state: ParentState): return {“foo”: “hi! " + state[“foo”]}

builder = StateGraph(ParentState) builder.add_node(“node_1”, node_1) builder.add_node(“node_2”, subgraph) builder.add_edge(START, “node_1”) builder.add_edge(“node_1”, “node_2”) graph = builder.compile()

for chunk in graph.stream({“foo”: “foo”}): print(chunk)

{’node_1’: {‘foo’: ‘hi! foo’}} {’node_2’: {‘foo’: ‘hi! foobar’}}

</Accordion>

## Subgraph persistence

When you use a subgraph, you need to decide what happens to its internal data between calls. Consider a customer support bot that delegates to specialist subagents: should the "billing expert" subagent remember the customer's earlier questions, or start fresh each time it's called?

By default, subgraphs are **stateless** (no memory): each call starts with a blank slate. This is the right choice for most applications, including [multi-agent](/oss/python/langchain/multi-agent) systems where subagents handle independent requests. If a subagent needs multi-turn conversation memory (for example, a research assistant that builds context over several exchanges) you can make it **stateful** (persistent memory) so its conversation history and data accumulate across calls on the same thread.

<span class="callout-start" data-callout-type="note"></span>
The parent graph must be compiled with a checkpointer for subgraph persistence features (interrupts, state inspection, stateful memory) to work. See [persistence](/oss/python/langgraph/persistence).
<span class="callout-end"></span>

<span class="callout-start" data-callout-type="info"></span>
The examples below use LangChain's [`create_agent`](https://reference.langchain.com/python/langchain/agents/factory/create_agent), which is a common way to build agents. `create_agent` produces a [LangGraph graph](/oss/python/langgraph/graph-api) under the hood, so all subgraph persistence concepts apply directly. If you're building with raw LangGraph `StateGraph`, the same patterns and configuration options apply — see the [Graph API](/oss/python/langgraph/graph-api) for details.
<span class="callout-end"></span>

### Stateless

Use stateless subgraphs when each call to the subgraph is independent and the subagent doesn't need to remember anything from previous calls. This is the most common pattern, especially for [multi-agent](/oss/python/langchain/multi-agent) systems where subagents handle one-off requests like "look up this customer's order" or "summarize this document."

There are two stateless options depending on whether you need [interrupts](/oss/python/langgraph/interrupts) (human-in-the-loop pausing) and [durable execution](/oss/python/langgraph/durable-execution) within the subgraph.

#### With interrupts

<span class="callout-start" data-callout-type="tip"></span>
This is the recommended mode for most applications, including [multi-agent](/oss/python/langchain/multi-agent) systems where subagents are invoked as tools. It supports interrupts, [durable execution](/oss/python/langgraph/durable-execution), and parallel calls while keeping each invocation isolated.
<span class="callout-end"></span>

Use this when you want a subagent with no memory across calls, but you still need durable execution and the ability to pause mid-run for user input (for example, asking for approval before taking an action). This is the default behavior: omit `checkpointer` or set it to `None`. Each call starts fresh, but within a single call, the subgraph can use `interrupt()` to pause and resume.

The following examples use two subagents (fruit expert, veggie expert) wrapped as tools for an outer agent:

```python
from langchain.agents import create_agent
from langchain.tools import tool
from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command, interrupt

@tool
def fruit_info(fruit_name: str) -> str:
  """Look up fruit info."""
  return f"Info about {fruit_name}"

@tool
def veggie_info(veggie_name: str) -> str:
  """Look up veggie info."""
  return f"Info about {veggie_name}"

# Subagents — no checkpointer setting (inherits parent)
fruit_agent = create_agent(
  model="gpt-4.1-mini",
  tools=[fruit_info],
  prompt="You are a fruit expert. Use the fruit_info tool. Respond in one sentence.",
)

veggie_agent = create_agent(
  model="gpt-4.1-mini",
  tools=[veggie_info],
  prompt="You are a veggie expert. Use the veggie_info tool. Respond in one sentence.",
)

# Wrap subagents as tools for the outer agent
@tool
def ask_fruit_expert(question: str) -> str:
  """Ask the fruit expert. Use for ALL fruit questions."""
  response = fruit_agent.invoke(
      {"messages": [{"role": "user", "content": question}]},
  )
  return response["messages"][-1].content

@tool
def ask_veggie_expert(question: str) -> str:
  """Ask the veggie expert. Use for ALL veggie questions."""
  response = veggie_agent.invoke(
      {"messages": [{"role": "user", "content": question}]},
  )
  return response["messages"][-1].content

# Outer agent with checkpointer
agent = create_agent(
  model="gpt-4.1-mini",
  tools=[ask_fruit_expert, ask_veggie_expert],
  prompt=(
      "You have two experts: ask_fruit_expert and ask_veggie_expert. "
      "ALWAYS delegate questions to the appropriate expert."
  ),
  checkpointer=MemorySaver(),
)

Each invocation can use interrupt() to pause and resume. Add interrupt() to a tool function to require user approval before proceeding:

    @tool
    def fruit_info(fruit_name: str) -> str:
        """Look up fruit info."""
        interrupt("continue?")  # [!code highlight]
        return f"Info about {fruit_name}"
    ```

```python
    config = {"configurable": {"thread_id": "1"}}

    # Invoke — the subagent's tool calls interrupt()
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Tell me about apples"}]},
        config=config,
    )
    # response contains __interrupt__

    # Resume — approve the interrupt
    response = agent.invoke(Command(resume=True), config=config)  # [!code highlight]
    # Subagent message count: 4
    ```
  <span class="tab-end"></span>

  <span class="tab-start" data-tab-title="Multi-turn"></span>
Each invocation starts with a fresh subagent state. The subagent does not remember previous calls:

```python
    config = {"configurable": {"thread_id": "1"}}

    # First call
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Tell me about apples"}]},
        config=config,
    )
    # Subagent message count: 4

    # Second call — subagent starts fresh, no memory of apples
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Now tell me about bananas"}]},
        config=config,
    )
    # Subagent message count: 4 (still fresh!)
    ```
  <span class="tab-end"></span>

  <span class="tab-start" data-tab-title="Multiple subgraph calls"></span>
Multiple calls to the same subgraph work without conflicts, since each invocation gets its own checkpoint namespace:

```python
    config = {"configurable": {"thread_id": "1"}}

    # LLM calls ask_fruit_expert for both apples and bananas
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Tell me about apples and bananas"}]},
        config=config,
    )
    # Subagent message count: 4 (apples — fresh)
    # Subagent message count: 4 (bananas — fresh)
    ```
  <span class="tab-end"></span>
<span class="tab-group-end"></span>

#### Without interrupts

Use this when you want to run a subagent like a normal function call with no checkpointing overhead. The subgraph cannot pause/resume and does not benefit from [durable execution](/oss/python/langgraph/persistence). Compile with `checkpointer=False`.

<span class="callout-start" data-callout-type="warning"></span>
  Without checkpointing, the subgraph has no durable execution. If the process crashes mid-run, the subgraph cannot recover and must be re-run from the beginning.
<span class="callout-end"></span>

```python
subgraph_builder = StateGraph(...)
subgraph = subgraph_builder.compile(checkpointer=False)  # [!code highlight]

Stateful#

Use stateful subgraphs when a subagent needs to remember previous interactions. For example, a research assistant that builds up context over several exchanges, or a coding assistant that tracks what files it has already edited. With stateful persistence, the subagent’s conversation history and data accumulate across calls on the same thread. Each call picks up where the last one left off.

Compile with checkpointer=True to enable this behavior.

Stateful subgraphs do not support parallel tool calls. When an LLM has access to a stateful subagent as a tool, it may try to call that tool multiple times in parallel (for example, asking the fruit expert about apples and bananas simultaneously). This causes checkpoint conflicts because both calls write to the same namespace.

The examples below use LangChain’s ToolCallLimitMiddleware to prevent this. If you’re building with pure LangGraph StateGraph, you need to prevent parallel tool calls yourself — for example, by configuring your model to disable parallel tool calling or by adding logic to ensure the same subgraph is not invoked multiple times in parallel.

The following examples use a fruit expert subagent compiled with checkpointer=True:

from langchain.agents import create_agent
from langchain.agents.middleware import ToolCallLimitMiddleware
from langchain.tools import tool
from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command, interrupt

@tool
def fruit_info(fruit_name: str) -> str:
    """Look up fruit info."""
    return f"Info about {fruit_name}"

# Subagent with checkpointer=True for persistent state
fruit_agent = create_agent(
    model="gpt-4.1-mini",
    tools=[fruit_info],
    prompt="You are a fruit expert. Use the fruit_info tool. Respond in one sentence.",
    checkpointer=True,  # [!code highlight]
)

# Wrap subagent as a tool for the outer agent
@tool
def ask_fruit_expert(question: str) -> str:
    """Ask the fruit expert. Use for ALL fruit questions."""
    response = fruit_agent.invoke(
        {"messages": [{"role": "user", "content": question}]},
    )
    return response["messages"][-1].content

# Outer agent with checkpointer
# Use ToolCallLimitMiddleware to prevent parallel calls to stateful subagents,
# which would cause checkpoint conflicts.
agent = create_agent(
    model="gpt-4.1-mini",
    tools=[ask_fruit_expert],
    prompt="You have a fruit expert. ALWAYS delegate fruit questions to ask_fruit_expert.",
    middleware=[  # [!code highlight]
        ToolCallLimitMiddleware(tool_name="ask_fruit_expert", run_limit=1),  # [!code highlight]
    ],  # [!code highlight]
    checkpointer=MemorySaver(),
)

Stateful subagents support interrupt() just like per-invocation. Add interrupt() to a tool function to require user approval:

    @tool
    def fruit_info(fruit_name: str) -> str:
        """Look up fruit info."""
        interrupt("continue?")  # [!code highlight]
        return f"Info about {fruit_name}"
    ```

```python
    config = {"configurable": {"thread_id": "1"}}

    # Invoke — the subagent's tool calls interrupt()
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Tell me about apples"}]},
        config=config,
    )
    # response contains __interrupt__

    # Resume — approve the interrupt
    response = agent.invoke(Command(resume=True), config=config)  # [!code highlight]
    # Subagent message count: 4
    ```
  <span class="tab-end"></span>

  <span class="tab-start" data-tab-title="Multi-turn"></span>
State accumulates across invocations — the subagent remembers past conversations:

```python
    config = {"configurable": {"thread_id": "1"}}

    # First call
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Tell me about apples"}]},
        config=config,
    )
    # Subagent message count: 4

    # Second call — subagent REMEMBERS apples conversation
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Now tell me about bananas"}]},
        config=config,
    )
    # Subagent message count: 8 (accumulated!)
    ```
  <span class="tab-end"></span>

  <span class="tab-start" data-tab-title="Multiple subgraph calls"></span>
When you have multiple **different** stateful subgraphs (for example, a fruit expert and a veggie expert), each one needs its own storage space so their checkpoints don't overwrite each other. This is called **namespace isolation**.

If you [call subgraphs inside a node](#call-a-subgraph-inside-a-node), LangGraph assigns namespaces based on call order (first call, second call, etc.). This means reordering your calls can mix up which subgraph loads which state. To avoid this, wrap each subagent in its own `StateGraph` with a unique node name — this gives each subgraph a stable, unique namespace:

```python
    from langgraph.graph import MessagesState, StateGraph

    def create_sub_agent(model, *, name, **kwargs):
        """Wrap an agent with a unique node name for namespace isolation."""
        agent = create_agent(model=model, name=name, **kwargs)
        return (
            StateGraph(MessagesState)
            .add_node(name, agent)  # unique name → stable namespace  # [!code highlight]
            .add_edge("__start__", name)
            .compile()
        )

    fruit_agent = create_sub_agent(
        "gpt-4.1-mini", name="fruit_agent",
        tools=[fruit_info], prompt="...", checkpointer=True,
    )
    veggie_agent = create_sub_agent(
        "gpt-4.1-mini", name="veggie_agent",
        tools=[veggie_info], prompt="...", checkpointer=True,
    )

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

    # First call — LLM calls both fruit and veggie experts
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Tell me about cherries and broccoli"}]},
        config=config,
    )
    # Fruit subagent message count: 4
    # Veggie subagent message count: 4

    # Second call — both agents accumulate independently
    response = agent.invoke(
        {"messages": [{"role": "user", "content": "Now tell me about oranges and carrots"}]},
        config=config,
    )
    # Fruit subagent message count: 8 (remembers cherries!)
    # Veggie subagent message count: 8 (remembers broccoli!)
    ```

Subgraphs [added as nodes](#add-a-subgraph-as-a-node) already get name-based namespaces automatically, so they don't need this wrapper.
  <span class="tab-end"></span>
<span class="tab-group-end"></span>

### Checkpointer reference

Control subgraph persistence with the `checkpointer` parameter on `.compile()`:

```python
subgraph = builder.compile(checkpointer=False)  # or True / None

• Interrupts (HITL): The subgraph can use interrupt() to pause execution and wait for user input, then resume where it left off.
• Multi-turn memory: The subgraph retains its state across multiple invocations within the same thread. Each call picks up where the last one left off rather than starting fresh.
• Multiple calls (different subgraphs): Multiple different subgraph instances can be invoked within a single node without checkpoint namespace conflicts.
• Multiple calls (same subgraph): The same subgraph instance can be invoked multiple times within a single node. With stateful persistence, these calls write to the same checkpoint namespace and conflict — use per-invocation persistence instead.
• State inspection: The subgraph’s state is available via get_state(config, subgraphs=True) for debugging and monitoring.

View subgraph state#

When you enable persistence, you can inspect the subgraph state using the subgraphs option. With checkpointer=False, no subgraph checkpoints are saved, so subgraph state is not available.

Viewing subgraph state requires that LangGraph can statically discover the subgraph — i.e., it is added as a node or called inside a node. It does not work when a subgraph is called inside a tool function or other indirection (e.g., the subagents pattern). Interrupts still propagate to the top-level graph regardless of nesting.

Returns subgraph state for the current invocation only. Each invocation starts fresh.

    from langgraph.graph import START, StateGraph
    from langgraph.checkpoint.memory import MemorySaver
    from langgraph.types import interrupt, Command
    from typing_extensions import TypedDict

    class State(TypedDict):
        foo: str

    # Subgraph
    def subgraph_node_1(state: State):
        value = interrupt("Provide value:")
        return {"foo": state["foo"] + value}

    subgraph_builder = StateGraph(State)
    subgraph_builder.add_node(subgraph_node_1)
    subgraph_builder.add_edge(START, "subgraph_node_1")
    subgraph = subgraph_builder.compile()  # inherits parent checkpointer

    # Parent graph
    builder = StateGraph(State)
    builder.add_node("node_1", subgraph)
    builder.add_edge(START, "node_1")

    checkpointer = MemorySaver()
    graph = builder.compile(checkpointer=checkpointer)

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

    graph.invoke({"foo": ""}, config)

    # View subgraph state for the current invocation
    subgraph_state = graph.get_state(config, subgraphs=True).tasks[0].state  # [!code highlight]

    # Resume the subgraph
    graph.invoke(Command(resume="bar"), config)
    ```
  <span class="tab-end"></span>

  <span class="tab-start" data-tab-title="Stateful"></span>
Returns **accumulated** subgraph state across all invocations on this thread.

```python
    from langgraph.graph import START, StateGraph, MessagesState
    from langgraph.checkpoint.memory import MemorySaver

    # Subgraph with its own persistent state
    subgraph_builder = StateGraph(MessagesState)
    # ... add nodes and edges
    subgraph = subgraph_builder.compile(checkpointer=True)  # [!code highlight]

    # Parent graph
    builder = StateGraph(MessagesState)
    builder.add_node("agent", subgraph)
    builder.add_edge(START, "agent")

    checkpointer = MemorySaver()
    graph = builder.compile(checkpointer=checkpointer)

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

    graph.invoke({"messages": [{"role": "user", "content": "hi"}]}, config)
    graph.invoke({"messages": [{"role": "user", "content": "what did I say?"}]}, config)

    # View accumulated subgraph state (includes messages from both invocations)
    subgraph_state = graph.get_state(config, subgraphs=True).tasks[0].state  # [!code highlight]
    ```
  <span class="tab-end"></span>
<span class="tab-group-end"></span>

## Stream subgraph outputs

To include outputs from subgraphs in the streamed outputs, you can set the subgraphs option in the stream method of the parent graph. This will stream outputs from both the parent graph and any subgraphs.

```python
for chunk in graph.stream(
    {"foo": "foo"},
    subgraphs=True, # [!code highlight]
    stream_mode="updates",
):
    print(chunk)

</Accordion>

***

<span class="callout-start" data-callout-type="note"></span>
[Edit this page on GitHub](https://github.com/langchain-ai/docs/edit/main/src/oss/langgraph/use-subgraphs.mdx) or [file an issue](https://github.com/langchain-ai/docs/issues/new/choose).
<span class="callout-end"></span>

<span class="callout-start" data-callout-type="note"></span>
[Connect these docs](/use-these-docs) to Claude, VSCode, and more via MCP for real-time answers.
<span class="callout-end"></span>