Generative AI Google Brain

gemini-agents-genai-experience-concierge-langgraph-demo-backend-notebooks-langgraph-agent

2 min read

Constructing a LangGraphAgent

Source notebook

Repo path: gemini/agents/genai-experience-concierge/langgraph-demo/backend/notebooks/langgraph-agent.ipynb · Open on GitHub · intermediate

Wraps a simple LangGraph StateGraph in LangGraphAgent and tests stateful/stateless streaming.

Summary

The notebook teaches how to transform a LangGraph graph into a LangGraphAgent using concierge.langgraph_server. It defines a typed state with an id, builds async SET_ID and REVERSE_ID nodes, and wires them into a StateGraph. It then demonstrates a memory-backed stateful agent with debug, values, and updates stream modes, plus a stateless agent created by disabling the checkpointer.

Key code patterns

Typed graph state

class State(TypedDict, total=False):
    id: str

Defines the state schema passed into StateGraph.

Command-based node routing

async def set_id(state: State):
    new_id = state.get("id")
    assert new_id is not None, "must set ID"
    return types.Command(update=State(id=new_id), goto="REVERSE_ID")

Shows a LangGraph node updating state and routing to the next node.

StateGraph construction

state_graph = graph.StateGraph(state_schema=State)
state_graph.add_node("SET_ID", set_id)
state_graph.add_node("REVERSE_ID", reverse_id)
state_graph.set_entry_point("SET_ID")

Builds the graph that LangGraphAgent wraps.

Stateful LangGraphAgent

basic_agent = langgraph_agent.LangGraphAgent(
    state_graph=state_graph,
    checkpointer_config=schemas.MemoryBackendConfig(),
)

Uses a memory checkpointer for stateful threaded requests.

Streaming with thread ID

async for chunk in basic_agent.stream(
    input=State(id="hello"),
    config={"configurable": {"thread_id": uuid.uuid4().hex}},
    stream_mode="updates",
):
    print("updates", "->", chunk)

Demonstrates async streaming for a stateful agent.

Stateless LangGraphAgent

stateless_basic_agent = langgraph_agent.LangGraphAgent(
    state_graph=state_graph,
    checkpointer_config=None,
)

Disables the checkpointer so the agent works with stateless requests.

Models & APIs used

  • SDKs / libraries: langgraph, concierge.langgraph_server

When to use this

Use this pattern to prototype a LangGraph graph as a LangGraphAgent interactively before deploying it on a server.

Gotchas & caveats

  • MemoryBackendConfig requires a configurable thread_id for stateful requests.
  • checkpointer_config=None means the agent only works with stateless requests.
  • The node functions assert that id is present before updating or reversing it.
  • concierge.langgraph_server must be importable in the notebook environment.
  • The stream examples use async for, so execution must run in an async-capable context.

Best practices

  • Define an explicit TypedDict state schema before constructing the StateGraph.
  • Return types.Command from nodes to update state and route execution.
  • Use separate stream modes to inspect debug events, full values, and node updates.
  • Use unique uuid.uuid4().hex thread IDs for stateful stream calls.
  • Build a separate agent with checkpointer_config=None for stateless requests.

Graph View

Backlinks